Assignement

Anonymous009
1904.02697.pdf

1

Abstract—Autonomous Vehicles (AV) are expected to bring

considerable benefits to society, such as traffic optimization and

accidents reduction. They rely heavily on advances in many

Artificial Intelligence (AI) approaches and techniques. However,

while some researchers in this field believe AI is the core element

to enhance safety, others believe AI imposes new challenges to

assure the safety of these new AI-based systems and applications.

In this non-convergent context, this paper presents a systematic

literature review to paint a clear picture of the state of the art of

the literature in AI on AV safety. Based on an initial sample of

4870 retrieved papers, 59 studies were selected as the result of the

selection criteria detailed in the paper. The shortlisted studies were

then mapped into six categories to answer the proposed research

questions. An AV system model was proposed and applied to

orient the discussions about the SLR findings. As a main result, we

have reinforced our preliminary observation about the necessity

of considering a serious safety agenda for the future studies on AI-

based AV systems.

Keywords: Autonomous vehicles, safety, artificial intelligence,

machine intelligence, machine learning, SLR.

I. INTRODUCTION

Advances in Artificial Intelligence (AI) are one of the key

enablers of the Autonomous Vehicles (AVs) development. In

fact, AVs rely on AI to interpret the environment, understand

its conditions, and make driving-related decisions. Thus, it

basically replicates the human driver actions when driving a

vehicle. In this context, AI applied to AV has become an

important research topic.

AV is a safety-critical system. When operating in an

undesirable way, AV can jeopardize human lives or the

environment in which it operates. It has the potential to threaten

the lives of its own passengers, pedestrians and people in other

vehicles, and damage other transportation system elements (e.g.

other vehicles and transportation infrastructure). Therefore, it is

mandatory to assure AV is safe, mainly when operating on

public roads in which resources will be shared with other

systems (and people).

Although safety is a mandatory characteristic to AV, and

although the researchers seem to agree on the importance of AI

This work was supported in part by the Research, Development and

innovation Center, Ericsson Telecommunication S.A., Brazil. 1Safety Analysis Group from Polytechnic School of University of São Paulo,

São Paulo, Brazil (e-mail: alexandremoreiranascimento@gmail.com).

applied to autonomous vehicles, they seem to disagree on the

AIs impact on AV safety. Many researchers, in special those

related to the AI community and AV manufacturers, advocate

AI as one of the core elements to enhance AV safety. Their

hypothesis is the automation of the driving tasks will lead to a

significant reduction of the car accidents. However, other

researchers, mainly in the system safety community, argue that

AI can potentially jeopardize AVs safety.

This study is the first, as far as we are aware, to map and to

organize the related literature and to provide a complete view

of the aspects related to both visions, and to subsidize future

studies. A preliminary study on the concerns about the

differences between AI and system safety mindsets impacting

AV safety was published in [1]. In this non-convergent context,

this paper presents a systematic literature review (SLR) aiming

to present a clear picture of the state of the art of the literature

in AI on AVs safety.

This paper is structured into 5 sections. Section II presents

details about the research methodology used. Section III

presents the data analysis results from the SLR based on the

proposed methodology. Section IV proposes an AV system

model that is used to orient the discussions about SLR findings.

Finally, Section V presents the final remarks.

II. RESEARCH METHODOLOGY

This study was performed using the systematic literature

review (SLR) method. The reasons supporting the SLR use are:

(1) its established tradition as a tool to understand state-of-the-

art research in technology-related fields [2]; (2) it helps to

understand existing studies and supports readers in identifying

new directions in the research field [3]; and (3) it helps to create

a foundation for advancing knowledge [4].

The protocol used (Figure 1) was based on the tasks

suggested by [5][6] for defining the research questions,

identification of search string, source selection, study selection

criteria, and data mapping. Also, the protocol followed the

recommendations of [7], [8], [4] and [9] for extracting,

analyzing, interpreting and reporting the literature-based

findings.

2Ericsson Research, Ericsson AB, Stockholm, Sweden (e-mail:

rafia.inam@ericsson.com). 3Ericsson Telecommunication S.A., Indaiatuba, SP, Brazil (e-mail:

maria.marquezini@ericsson.com).

A Systematic Literature Review about the

impact of Artificial Intelligence on

Autonomous Vehicle Safety

A. M. Nascimento1, L. F. Vismari1, C. B. S. T. Molina1, P.S. Cugnasca1, J.B. Camargo Jr.1, J.R. de

Almeida Jr.1, R. Inam2, E. Fersman2, M. V. Marquezini3, and A. Y. Hata3

2

Figure 1. Protocol used to support systematic literature review.

A. Definition of Research Questions

The first step was to define the research questions (RQ). In

order to support the research goal of presenting a clear picture

of the state of the art in the literature about AI on AV safety, the

following research questions were posed:

 RQ1. How do AI-based systems impact system safety?

 RQ2. Which are the topics (context domain) of the studies

identified?

 RQ3. Which AI-related techniques are used on the

studies?

 RQ4. Which problems do the techniques seek to address?

 RQ5. Which findings are reported by the study’s authors?

 RQ6. Which future studies are suggested in these studies?

B. Identification of Search String and Source Selection

The search strategy was structured through the selection of

source databases and the appropriate search terms. No date

range was used, to ensure that relevant studies were covered,

regardless of their publication date. A broad selection of online

databases indexing scientific literature was considered: ACM,

Engineering Village, ScienceDirect, Scopus, SpringerLink,

Wiley and Web of Science (WoS). Please note that IEEExplore

is already covered by the selected databases for this SLR study.

The search string was designed based on the synonyms of the

3 main concepts related to the investigated topics: Safety,

Artificial Intelligence and Autonomous Vehicle. Many

synonyms are present in the literature for the terms "artificial

intelligence" and "autonomous vehicle". Therefore, an

exploratory study of their most representative synonyms was

performed. Then, a careful selection of synonyms was made to

ensure the search process would have an appropriate coverage.

As a result, the following string with Boolean operators was

selected: (“safety” AND (“artificial intelligence” OR

“machine intelligence” OR “machine learning”) AND

(“autonomous vehicle” OR “autonomous car” OR “automated

vehicle” OR “automated car” OR “self-driven vehicle” OR

“self-driving” OR “driverless”)). Note that the synonyms for

each one of the topics are already presented in the Boolean

string previously displayed.

Different instances of the search string were created to adapt

it to the distinct database search syntax rules, but the same

logical value was kept. In each database, the appropriate options

were selected to limit the search process to the Title-Abstract-

Keyword (TAK) field set. This is an important measure to

reduce the number of non-related or duplicated studies

retrieved. However, it was observed that not all databases

support a search limited on TAK field set, leading to an inflated

number of papers found (e.g. SpringerLink). Table I shows the

initial number of papers found per database.

TABLE I

NUMBER OF PAPERS PER DATABASE

Database #Entries

ACM 36

Engineering Village 191

ScienceDirect 81

Scopus 182

SpringerLink 3999

Wiley 329

WoS 52

Total 4870

C. Study Selection Criteria and Papers Review

The study selection process is shown in Figure 2. Each step

indicates the number of papers remaining as a sample after the

corresponding step was executed. The first selection criterion

applied was to ensure that only the studies with the TAK fields

returning positive to the Boolean search expression would be

selected. The information (metadata) available for each paper

found, in the first step of the selection process, was collected by

exporting the results to a spreadsheet. A spreadsheet macro was

developed to analyze the TAK fields and to properly select the

papers. After this check, only 230 papers remained as a sample.

Using the spreadsheet Remove Duplicate tools, the duplicated

entries were removed. The 97 remaining papers composed the

selected sample.

Figure 2. Study selection process.

As a reasonable number of papers (97) was found [10], book

chapters, editorials, notes or reports were excluded - level 5

exclusion [10] - and 86 papers remained. The abstracts, titles

and keywords of the remaining 86 peer reviewed papers were

scrutinized to check their fitness with the goals of this research.

After a careful examination (sometimes a full-paper skimming

was necessary), 27 papers were considered not related to this

Complete Review (59)

Level 5 Exclusion

(86)

Duplicated Removal (97)

TAK Filtering (230)

Papers Retrieved (4870)

3

research and were excluded from the sample of the literature

mapping. Finally, a sample of 59 papers was considered for this

study.

There was a considerable drop in the number of studies, from

the initial 4870 to the final 59 papers selected. It occurred for

different reasons, such as: misuse of the terminology; correct

use of the terminology in the context of an example within a

paper that did not actually focus on the topic; or lack of

restricted search in TAK fields in some databases (in our study,

the SpringerLink).

D. Data Mapping

The data mapping from the selected papers were executed

after they had been completely reviewed and scrutinized. It was

performed categorizing the 59 sample papers into 6 categories

(CT.1-6) to answer the 6 research questions (RQ.1-6),

respectively. The categories defined were based on the

corresponding research question: (CT.1) Impact, (CT.2)

Topics, (CT.3) Techniques, (CT.4) Problem, (CT.5) Findings

and (CT.6) Future Studies. The categorization process was

based on the agreement of researchers working in this study.

Different strategies were used to create the codes for each of the

categories. For (CT.1) Impact, the code increase was used when

the paper described AI as a factor of increasing the safety risk

(negative impact on safety) and the code decrease was used

when the paper presented AI as a factor of decreasing the safety

risk (positive impact on safety). For (CT.2) Topics, (CT.5)

Findings and (CT.6) Future Studies, the codifications were

derived by the context domain of the study according to what

was reported by their authors, as suggested by [11]. Lastly, for

(CT.3) Techniques and (CT.4) Problem, similarly to other

categories, the codes were based on what was reported by their

authors [11] and, due to the wide range of techniques, subfields

and misuses of terms, the terminologies were adapted and

normalized according to field references [12][13]

[14][15][16][17].

III. DATA ANALYSIS

The distribution of the studies over the years can provide an

overview of the size and evolution of the field (Figure 3). The

left chart in Figure 3 shows the distribution from 1987 until

2018 (April). The oldest study found dates back to 1987. No

work was found for over a decade – from 1991 to 2002 –

considering the adopted search criteria.

This period can be labeled as the "first winter" in this research

topic as an analogy to the Artificial Intelligence "winters"1.

Only one paper a year was found over the following 3 years –

from 2003 to 2005. A second short winter was found from 2006

to 2008. Only 1 paper was found in 2009 and another in 2011,

while no paper was found in 2010. Finally, the combination of

AI, safety and autonomous vehicles started to get more

attention from the scientific community in 2012 when 5 papers

were found, although no paper was found in 2013. In fact, 86%

(51) of the papers found were published from 2012 to 2018.

The right chart in Figure 3 shows the distribution of the

studies over the last decade. The year 2018 was excluded from

the plot to avoid misinterpretation. Considering the results

presented in Figure 3, the field is gaining momentum based on

the continuous growth in the number of published studies since

2014. The trend line built in the last decade data shows a higher

angular coefficient, indicating the momentum in recent years.

Most of the papers found are from conference proceedings.

In fact, 45 papers (76%) are from conferences. Only 14 papers

(24%) were published by journals. Therefore, it is reasonable to

expect a growth in the number of publications about this topic

in journals. Besides evaluating the time distribution of papers,

another important aspect is the consistency-check of the

selected keywords in the papers considered. This was

performed by checking the most representative keywords

among all the synonyms of each of the 3 sets (previously

presented) in the search string. All the keywords from the

search string found on each paper TAK were accounted. As a

result, the total number of hits per keyword was computed.

Table II shows the number of studies with each keyword

present (hits per keyword) and the percentage of the 59 sample

papers with the keyword. Note the sum of the number of hits

does not totalize 59. Also, the sum of the percentages for all the

keywords for each distinct concept does not totalize 100%. This

is because many papers have more than 1 synonym present,

which makes it be accounted more than once.

Thus, it is possible to note the most representative keyword

for each concept: safety, artificial intelligence and autonomous

vehicle. In fact, a search string using only those keywords

would result in 36 papers, which corresponds to 61% of the

sample size of the present study. However, many other

keywords used could not be ignored, since they have a

considerable representativeness, such as: machine learning,

automated vehicle, self-driving and autonomous car.

Conversely the keyword autonomous truck surprisingly had

only one hit.

The following sub-sections present the results for each

research question (RQ.1-6).

A. AI-based systems impact on safety (RQ.1)

The RQ.1 was answered with the categorization of the

sample studies into CT.1 (Impact). Most studies consider AI a

technology that increases the system safety (positive impacts on

safety). So, 81% (48) of the papers were actually coded as

decrease, because they argue that AI decreases the safety risks.

Only 19% (11) of the studies consider AI a potential threat to

the system safety.

B. Main topics of the studies (RQ.2)

In order to answer RQ.2, the sample papers were classified

into the category CT.2 (Topics). Studies were grouped based on

their CT.1 coding into two distinct sets: Increase Safety

1 The Artificial Intelligence field had periods of warm enthusiasms and some

periods of very low enthusiasm, with a much lower number of publications

and contributions. The literature named those low enthusiasm periods as AI

winters.

4

Figure 3. Studies distribution over the years: a) depicting all studies till 2018; b) depicting studies in the last ten years.

TABLE II

KEYWORDS HITS

Concept Keyword #Hits %Papers

Safety Safety 59 100

Artificial

Intelligence

Artificial intelligence 36 61

Machine learning 27 46

Machine intelligence 1 2

Autonomous

Vehicle

Autonomous vehicle 37 63

Automated vehicle 12 20

Self-driving 11 19

Autonomous car 6 10

Driverless 2 3

Autonomous truck 1 2

Risks and Decrease Safety Risks. Then studies were grouped

by their similarities and each group was coded with a label that

could encompass all its members. Table III shows the results of

this coding process. As observed, the papers positioning AI as

a factor that decreases safety risks (48 papers, 81%), they

studied the subjects related to five main topics: Sensors and

Perception (21 papers, 44%), Navigation and Control (13

papers, 27%), Fault Prevention (6 papers, 13%), Conceptual

Model and Framework (4 papers, 8%) and Human Factor (4

papers, 8%). In turn, the papers positioning AI as a risk to

system safety (11 papers, 19%) studied subjects related to three

main topics: Fault Forecasting (5 papers, 45%), Ethics and

Policies (4 papers, 36%) and Dependability and Trust (2 papers,

18%). The complete list of references for each code in this

category can also be found in Table III.

The main topics for each group of papers differ reasonably

from each other. While the papers in the category decrease

focus on important aspects to support or to enhance the vehicle

autonomy, the papers in the category increase (endanger safety)

focus on topics related to safety assurance.

Sensors and Perception is the topic with the largest number

of studies (21). They are mostly related to computer vision and

detection techniques necessary for adding the necessary

capabilities to detect different aspects of the navigation

environment and supporting the autonomy of the AVs, such as:

general computer vision [19], Doppler sensing [20], lane

detection [21], daylight detection and evaluation [22], obstacles

detection [23][24][25], pedestrian detection [26][27],

pedestrian trajectory prediction [28], road detection

[24][29][75], road junction detection [30], road terrain

detection [31], traffic signal detection [32][18], situation

awareness [33], speed bump detection [25][34], traffic light

detection [35], vehicle detection [36] and virtual worlds for

training detection [37].

The second largest number of papers (13) found

encompasses studies related to Navigation and Control. They

are mostly related to techniques necessary to ensure the proper

autonomous navigation and control capabilities required by

AVs, such as: remote-controlled semi-AV based on IoT [38];

adaptive pre-crash control [39]; safe trajectory selection [76];

AV following another car driven by a human pilot (Trailing)

[40]; safe navigation [41]; heuristic optimization algorithm for

unsigned intersection crossing [42]; vehicle coordination [43];

maneuver classification [44]; learning to navigate from

demonstration [45]; AV movements optimization in

intersection [46]; learning and simulation of the Human Level

decisions involved in driving a racing car [47]; path tracking

[48]; and fuzzy-logic control approach to manage low level

vehicle actuators (steering throttle and brake) [49].

Six papers with research related to Fault Prevention were

found. These studies encompass researches related to the

preventing the occurrence or introduction of faults [50], such as

AI for security of wireless communication to ensure safety [51];

remote diagnosis, maintenance and prognosis Framework [52];

prediction of computational workload [53]; vehicle security

against cyber-attack [39][54]; and diagnosis of sensor faults

[55].

Four studies were found for each of the topics Human Factor

and Conceptual Model and Framework. The studies on human

factor cover important aspects to be considered in the

autonomous cars engineering due to the human-in-the-loop

factor, such as: safety, comfort, and stability based on the

human driver perception behavior [56]; design of real time

transition from assisted driving to automated driving under

conditions of high probability of a collision [57]; diagnosing

and predicting stress and fatigue of driver in semi-automated

vehicles [58]; and advances in driver-vehicle interface [59].

Considering the studies (4) proposing conceptual models and

frameworks, they have a considerable diversity of focus, such

as: ML and cloud-based framework proposed to address safety

and reliability-related issues [60]; AV conceptual model [61];

an interdisciplinary framework to extract knowledge from the

large amount of available data during driving to reduce driver’s

behavioral uncertainties [62]; and a proposition of an AV

highway concept to improve highway driving safety [63].

Considering the group of papers positioning AI as a potential

factor of decreasing the safety, the highest number of studies

was related to Fault Forecasting. In other words, those

0

5

10

15

20

25

1987 1992 1997 2002 2007 2012 2017

R² = 0,60323

-5

0

5

10

15

20

25

30

2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018

5

TABLE III

IMPACT OF AI-BASED SYSTEMS ON SAFETY AND ITS MAIN TOPICS AND REFERENCES

Category Codes #Hits %Papers References

CT.1 - Impact Decrease Safety Risks (Positive Impact on Safety) 48 81

CT.2 - Topics

Sensors and Perception 21 44

[18][75][35][36][21][25][27][29]

[23][22][34][20][44][32][31][33]

[26][28][30][24][19]

Navigation and Control 13 27 [38][39][76][40][41][47][42][43]

[48][44][45][46][49]

Fault Prevention 6 13 [39][51][52][53][54][55]

Conceptual Model and Framework 4 8 [60][61][62][63]

Human Factor 4 8 [56][57][58][59]

CT.1 - Impact Increase Safety Risks (Negative Impact on Safety) 11 19

CT.2 - Topics

Fault Forecasting 5 45 [64][65][66][67][68]

Ethics and Policies 4 36 [69][70][71][72]

Dependability and Trust 2 18 [73][74]

papers dealt with the limitations to estimate the present number

and future incidence of faults in AI-based systems, by executing

activities related to evaluation, testing, verification and

validation [50], such as: aspects (and limitations) related to

safety validation [64]; performance and safety verification

methodology [65]; test suites for AV [66]; end-to-end safety for

AV design [67]; and a framework to evaluate the impacts of

such a sophisticated system on traffic and the impact of

continuous increase in the number of highly automated vehicles

on future traffic safety and traffic flow [68].

There were four studies related to discussions about Ethics

and Policies. One of the studies discussed and performed

experiments on how distinct ethical frameworks adopted to

make decisions about AV crashes can affect the number of lives

endangered [69]. The other studies discuss the scope of AI on

AV with ethical aspects [70], ethics in AV design [71], and

moral values and ethical principles for autonomous machines

[72]. As can be seen, those studies are quite recent since the

oldest one was published in 2015.

Finally, 2 papers were found related to Dependability and

Trust. Dependability is an important concept in critical systems,

because it comprises attributes such as safety, security,

availability, reliability and maintainability, as well as how (the

mechanisms) to keep these systems attributes [50]. According

to [50], trust can be defined as accepted dependability. The

studies found are thus related to: safety issues [73] and current

mechanisms to ensure robust operation in safety-critical

situations facing the introduction of non-deterministic software

[74].

C. Techniques used (RQ.3) and problems they seek to address (RQ.4)

Aiming to answer RQ.3 and RQ.4, the sample papers were

classified into categories CT.3 (techniques) and CT.4

(problems) based on how their authors described the AI

technique used in the study. Then, some terminologies used to

define the codification for the categories CT.3 and CT.4 were

adapted based on the field literature [12][13][14][15][16][17],

when necessary.

Most reviewed papers reported the specific AI related

techniques used in the research. Some reported the use of more

than one technique, whereas others reported only the approach

used. Some papers (14 papers, 24%) were related to general

aspects of AI or ML techniques, without mentioning specific

techniques used or researched [68][71][72][61][64]

[67][69][74][70][65][63][58][62][19].

All the techniques found in the reviewed papers were mapped

considering the problem (CT.4) that they were solving. As a

result, Table VIII - placed in the Appendix - lists the techniques

found, the number of papers in which they were used, the main

problems they were seeking to address, and the references.

As can be seen, there is a considerable number of studies

(22%,13) that used techniques related to artificial neural

networks. Also, there is a reasonable number of studies

reporting the use of SVM (17%, 10). Some studies used Fuzzy

Logic (8%, 5), Bayesian Artificial Intelligence (e.g. Bayesian

Deep Learning, Naive Bayes Classifier-NBC, etc) (7%, 4),

Hidden Markov Based Models (e.g. Continuous Hidden

Markov Model-CHMM and Discrete Hidden Markov Model

DHMM) (7%, 4), Estimation Filters (e.g. Kalman Filter and

Particle Filters) (7%, 4), Nearest-Neighbour-Based Algorithm

(e.g. k-Nearest Neighbours - kNN) (7%, 4), Adaptive Boosting

(AdaBoost) (5%, 3), Ramer-Douglas-Peucker or Rameri

Douglas algorithm (5%, 3), Haar-like feature detector (5%, 3),

Histogram of Oriented Gradient (HOG) (5%, 3), Hough

Transformation (5%, 3), Optimization Heuristics (5%, 3),

Regression-Based Models (5%, 3) and Principal Components

Analysis (PCA) (3%, 2).

Analyzing Table VIII, it shows that each of the following

techniques were reported, in all the reviewed papers, only once:

Canny Edge Detection Algorithm, Case-based reasoning

(CBR), Channel Features, Clustering Algorithm k-mean,

Complex Decision Trees (CDT), Conditional Random Fields

(CRFs), Distributed Random Forest (DRF), Gaussian Mixture

Model (GMM), Linear Temporal Logic (LTL), Local Binary

Patterns (LBP), Neuroevolution of Augmenting Topologies

(NEAT), Novel Image Recognition Technique, Path Planning

Algorithms (A* and D*), Satisfiability Modulo Theories (SMT)

6

Solver, and Viterbi Algorithm. Thus, there is room for new

studies using techniques not yet used or under-represented by

the set of papers considered.

D. Reported findings (RQ.5)

Question RQ.5 is answered by CT.5 (findings), based on the

information about the findings reported on the sample papers.

Some papers did not report specific main findings in a

straightforward way because the propose frameworks or

approaches had not yet been tested or the results were still

incipient. Other papers described very specific findings that

would require a background section to support a proper

discussion. In those cases, only a higher level of abstraction of

the results is presented. Finally, because of space limitation,

only some specific examples are described here, while most of

the results are presented grouped around the main topic of

research. A complete list, oriented by the discussion presented

at Section IV, can be found on the Table IX presented in

Appendix.

The papers about topics related to Sensors and Perception

presented positive and promising results with the techniques

employed to address their research problems. In fact, this topic

already achieved significant results with the recent

developments in AI and sensor technologies. While AI had the

image and pattern recognition boosted by advancements such

as the new architectures of ANNs and new machine learning

techniques, sensor technologies have been boosted in the last

decades by the advancements in the robotics and mobile phone

industries. As a result, the papers demonstrated applications of

enhancements in the techniques or combination of techniques

and sensors in order to recognize and to detect important

elements and signals the human drivers need to handle to ensure

the proper operation of a vehicle. In this context, the findings

are positive for the application of ANNs to recognize turn signal

[18], road environment and signals [27][32][31][30], and

pedestrian [26][28], for example. Likewise, some papers

reported SVM has been applied successfully to detect road [75],

traffic light [35], and pedestrian [27].

The papers related to Navigation and Control also reported

positive and promising results. As presented previously, they

used diverse AI techniques to seek to address a broad range of

problems. For example, a hybrid AI architecture encompassing

ANN, CBR, and a hybrid Case-Based Planner (A* and D*

motion planner) was successfully tested to tackle the precrash

problem of intelligent control of autonomous vehicles [39],

while SVM was used to support a safest path planning in a

dynamic environment to avoid maneuvers too close to an

obstacle [41].

This SLR found 6 papers for the topic Fault Prevention. Each

of these papers used a distinct AI technique for the research

problems. One paper presented a preliminary result [53], and

another one proposed an approach but did not report results

[55]. All the others papers, related to the detection of cyber-

attack, presented promising positive results for the application

of ANNs [39], Estimation Filters [51], and Fuzzy-Logic [54],

for example. Also, preliminary positive results have been

reported on the use of a regression-based model to predict the

CPU patterns [53].

Two from the four remaining papers related to the topic

Human Factor, have presented preliminary positive results. One

presented promising results from using a regression-based

model to deal with selective attention mechanism [56], while

the other presented some examples of scenarios where the use

of Bayesian AI could avoid the collision when no action is taken

by the human driver [57]. The other 2 papers did not present

specific findings, due to their theoretical nature related to the

design considerations for the driving assistance system [59] and

human drivers monitoring to enhance the integration between

AVs and human drivers [58].

The papers proposing conceptual models and frameworks

did not present findings related to experimental results. Most of

them relied on general AI/ML instead of a specific technique

[61][62][63]. Also, besides the proposed approaches

themselves, they focused the discussions around the issues they

aimed to address, the theoretical background and future

potential problems to be addressed in the field.

The last three topics (Fault Forecasting, Ethics and Policies,

and Dependability and Trust) have papers more oriented to

theoretical discussions and propositions around the challenges

AVs are facing or will face related to safety topics, such as test

and validation [64], certification [67][74], autonomy assurance

and trust when non-deterministic and adaptive algorithms are

used [74] - crash assignment facing distinct ethical theories

[69], for example. In this context, most of them do not present

specific findings using experimental setups; instead, they

envision potential future solutions for the discussed challenges.

In other words, those papers try to shed an alert light on the

important topics that seem to be neglected by the AV

enthusiasts, trying to push the research agenda towards safety

engineering mindset.

As exceptions, 3 papers presented practical applications and

results. [69] presented some interesting findings using a simple

experimental simulated environment to test specific crash

scenarios under three ethical theories. They found that

understanding rational ethics is crucial for developing safe

automated vehicles. The results of their experiment indicate that

in specific crash scenarios, utilitarian ethics may reduce the

total number of fatalities that result from automated vehicle

crashes. [66] proposed an approach to describe test-cases for

validating autonomous vehicles using recordings of traffic

situations for creating a minimal test-suit that could help in the

certification process. Considering the example presented, they

show how minimalism is achieved by manually comparing the

test-cases. Although it is an interesting and promising approach,

there are no evidences that it could address a safety certification

processes requirement when considering non-deterministic

algorithms. Hence, the research was still preliminary. Finally,

although [73] presents an end-to-end Bayesian Deep Learning

architecture to reduce the risks of hard classifications by

adopting probabilistic predictions accounting for each model,

no findings from real experiments were presented.

E. Reported future studies (RQ.6)

Question RQ.6 is answered by CT.6 (future studies), based

7

on the collected information about future studies reported on

the sample papers. Some papers did not suggest future studies.

Other papers described intended future studies or works under

development. Those are frequently small incremental changes,

such as change of parameter or new test scenarios. Therefore,

they are not reported here since their specificities would require

a considerable background on the papers contents. That is out

of the scope of the systematic literature review.

The studies related to Sensors and Perception propose many

future studies, but mostly around improvements that would be

made in the future to address some of the limitations of the

presented research. Due to the space limitations, only some

examples are described here. [18] suggests additional research

on image recognition of low contrast images and vehicle images

with brake lamps. [35] suggests future work on traffic lights

detection under severe weather or night conditions. [34]

suggests more research on detecting speed bump during night

time. They also suggest research on speed bumper detection

when they have no pattern or marking. In addition to that, [34]

suggests research to improve the recognition capabilities to

distinguish zebra crossing from speed bump. [75] proposed

future research about road detection using road lane markers

that could be detected by LIDAR, while [21] proposed more

research focused on optimizing the lane detection and vehicle

recognition algorithms to reduce their computational costs.

Also considering the high computational costs, [27] proposed

using parallel computing to increase the speed of the image

recognition algorithms. Finally, according to [37], additional

research is needed on using the virtual environments for testing

because the authors believe their usage for training and testing

intelligent systems are becoming more relevant.

Most of the studies related to Navigation and Control suggest

future studies. The majority suggests extensions to the work

they presented. Here, few examples are presented. The study

proposing hybrid control architecture [39] suggests an

extension to consider the full kinematics and dynamic

limitations of the vehicle, while constantly acting to avoid

collisions and unsafe driving. The paper proposing an approach

using SVM to avoid maneuvers too close to an obstacle by

adding a safety margin [41] proposes future re-search to extend

it using a combination with the kinetic convex hulls2 to enable

the possibility of computing the solution ahead in time.

According to the authors, this would help to predict the position

and the width of the optimal margin. As a result, it would

improve the approach by adding the ability of reduce the

collision risk by preventing the AV from driving into a

dangerous situation. The study using Fuzzy Logic as the main

approach to control a semi-autonomous car 100-km experiment

[40] proposes future research using new sensors and filtering

methods for data fusion to reduce the risk on scenarios where

the GPS signal is lost. Finally, the study on AVs intersection

crossing [46] describes future work in which more types of

vehicles and more adjacent intersections would be included in

the simulations.

Most of the studies (4 of 6) related to Fault Prevention

2 Check [76] for more information about kinetic convex hulls.

suggest future studies. Half of the studies are related to security

aspects, while the other half is related to

diagnosis/prognosis/prediction. The study proposing a cyber-

attack detection system based on ANNs [39] suggests a future

study to apply the proposed approach to a real vehicle in

addition to the application of LSTM to detect online sensor

attack. The study proposing the use of Particle Filter and

Kalman Filter to secure connected vehicles against DoS attack

[51] proposes future work to assess the proposed security

scheme under many distinct scenarios, and also to execute tests

in real world set-ups. The study about predicting ADAS

remaining useful life for the prognosis of its safety critical

components using ANNs and other techniques, such as SVM

[52], proposes a considerably wide range of future studies, such

as using Least Square Support Vector Machine (LSs SVM);

using big data techniques to analyze the server data; studying

connected vehicle prognosis; using driver, vehicle and region

profile data to understand the impact on the environment and

driving style impact on the system lifespan; and more studies

on prognostics-enabled decision Making (PDM). Finally, the

work presenting the use of regression-based methods to predict

the CPU usage patterns of software tasks running on an AV [53]

suggest future work on the use of some regularization methods

for automatic feature selection, but also to particularly

investigate the effects of underestimating CPU utilization, and

how to handle underestimation of CPU utilization when it

happens, aiming to better understand how safe over (or under)

estimation of CPU utilization is in terms of reliable autonomous

driving.

The studies about Ethics and Policies on AVs basically

suggest more research on those topics. In the same way, most

of the studies tackling human-factor-related topics do not

propose future studies. As an exception, the paper proposing the

application of regression-based model for the selective attention

mechanism subject [56] proposed a future study to help to

reveal the mechanism of rear end collision accident to some

extent.

Half (2 of 4) of the studies related to Conceptual Model and

Framework do not suggest any future studies. However,

implicitly, the next steps would be the deployment of those

suggested approaches on experimental set-ups to collect real

results. The study proposing a framework to reduce the

uncertainty of a driver behavior prediction model [62] suggests

more studies focusing on the resilience and sustainability of the

system when deployed on a large scale in a complex system.

The papers about Fault Forecasting suggest some future

research. Among them, [64] suggests more research on safety

envelope mechanisms to describe a boundary within the state

space of the AVs rather than trying to prove that it will always

work correctly. Koopman, in another paper [67], suggest that

the accepted practices must be updated to create an end-to-end

design and validation process to address all the safety concerns

considering cost, risk, and ethical considerations. [66] proposes

more work on creating automated test-cases. [68] proposes

more studies based on the framework they proposed to evaluate

8

the impacts of AVs on traffic safety, specially using stochastic

simulations with random number seeds to achieve a broader

representative and a variety of traffic situations, as well as using

the proper statistical analysis techniques to ensure the statistical

validity of the results.

Finally, the 2 studies about Dependability and Trust also

present some suggestions of future studies. [73] asks for more

research on new concrete safety evaluation metrics. [74]

suggests more research on understanding the dependence of the

system components on AVs is needed to establish trust. They

also suggest that could be achieved by investigating the many

ways in which people, the system, and the environment

interrelate.

IV. SLR FINDINGS ORIENTED BY AN AV SYSTEM MODEL

In the previous section, the state of the art in the literature

about AI on AV safety was identified and investigated by means

of a SLR. Six research questions oriented the literature

identification, in which studies that include keywords related to

safety, AI and AV were considered. The resulting studies were

investigated and mapped into 6 categories: Impact (increase or

decrease safety risks), Topics (sensors and perception;

navigation and control; fault prevention; conceptual model and

framework; human factor; etc.), Techniques (general AI/ML;

ANN; SVM; etc.), Problem (AV validation; road detection;

collision avoidance; etc.), Findings, and Future Studies.

These results considered the AV as a system, but its specific

components and functions in an architectural point-of-view

were not considered. For deepening the understanding about the

state of the art of AI on AV safety it is necessary to show how

the presented works are applied/fitted on AV in an architectural

point-of-view. In other words, which of AV

modules/components and functions are already being

developed and which one could be more explored. In order to

achieve this goal, it is going to be considered the AV

architecture proposed in [78].

An automotive manufacturer consortium (CAMP-AVR) [78]

proposed a high-level architecture considering the main system

components demanded for the vehicle movement control, to be

used in the deployment of future Dynamic Driving Tasks

(DDT). Figure 4 (left) illustrates the model considering a

traditional vehicle (i.e. human operation with no automation

deployed), and Figure 4 (right) illustrates the introduction of

some level of machine automation (hybrid) in Sensors,

Controller and Actuators elements. While the diagram

considering the human operation can solely be mapped to the

SAE Automation Level 0 (no automation), the hybrid one

encompassing machine automation with human-in-the-loop can

be mapped to the SAE Automation Levels 1 to 4 (semi-

autonomous) [78].

In this context, a modified version of the semi-autonomous

model is proposed here (Figure 5) including the system

boundary. Also, the human related components were grouped

as one single component (human-in-the-loop), which interacts

with Machine Actuators, Machine Control, Machine Perception

and Environment. A single component represents a more

realistic approach facing the complexity added by the human in

the system and allows the examination of the user actions and

interactions as suggested by [79]. Also, it supports a necessary

human-centered and holistic view [80] to better support the

complexity of the human behavior and its interaction to the

system. It avoids the misconceptions of the too logical designs

from some engineering designs and helps to consider and accept

human behavior the way it is, not the way engineers would wish

it to be [81]. In fact, this is a necessary upgrade considering the

original model is derived by the classical view from the

automation engineering for industrial applications, where the

environment was under control of the system designer, the

human interactions had a considerable narrower scope, and its

potential impact to the whole system were much lower, when

compared to its application to the semi-autonomous vehicles.

As a result, the proposed DDT version (Figure 5) can be used

to map the selected scientific literature. Therefore, it can

provide a concise perspective on how the field literature covers

those main components and which the uncovered areas are.

Also, it can provide a good overview on the predominance of

the papers valence (increase or decrease) on safety.

Table IV shows how CT.1 (Impact) and CT.2 (Topic) codes

are mapped to the components of the modified semi-

autonomous system model, as well as the relationship between

CT.1 and CT.2. Most of the papers are related to machine

perception, followed by papers related to a broad system view.

Then, the next largest group of papers is related to the machine

control component. The remaining papers are related to the

human-in-the-loop aspects. An interesting aspect is that only

the studies with a broad system aspect were found to have both

CT.1 codes (increase and decrease system safety). Basically,

the studies focused on distinct components solely understand

AI can increase the safety risk. Therefore, there is a lack of

studies with a critical mindset that explore the potential

negative impacts of AI on the individual components. Finally,

no papers were found related to the vehicle, machine actuators

or environment.

9

Figure 4. Dynamic Driving Tasks Models: No Automation (left) x Semi-Automation (right) – Source: [78]

Figure 5. An adapted version of Dynamic Driving Task (DDT) Model

Table V shows how the wide range of AI techniques (CT.3

code) is mapped to the components of the modified semi-

autonomous system model. The AI techniques are grouped

around their scope: system-oriented (32%, 19) and component-

oriented (68%, 40). When a paper uses a combination of

techniques, for example, ANN and SVM, it results into a unit

added to the total number of papers using ANN and a unit added

to the total number of papers using SVM. In this context, most

of the studies (63%, 12) related to system-wide scope referred

to general AI/MI. Most of the studies (20%, 11) related to

machine perception used ANNs. In fact, ANN, SVM and HMM

(Hidden Markov Model) account for 48% of the studies related

to machine perception. Fuzzy logic (18%, 3) is the most widely

used technique in the machine control-related papers. Fuzzy

Logic, SVM, Optimization Heuristics and Ramer-Douglas-

Peucker or Ramert Douglas algorithm account for 53% of the

studies related to machine control. Finally, Bayesian Artificial

Intelligence techniques are used in most of the studies (29%)

related to human-in-the-loop.

Table VI shows the total count of each AI technique

occurrence over the sample papers. The sample papers have

different heterogeneity in the applied AI approaches. Besides

24% of the papers using generic AI/ML concepts, 49% of the

papers applied only one type of AI technique. Therefore, they

are homogeneous in terms of the applied AI technique. In those

studies, the most widely used techniques were Artificial Neural

Networks (28%, 8), Fuzzy Logic (14%, 4) and Support Vector

Machine (SVM) (10%, 3). The remaining 27% employed a

hybrid approach by combining multiple types of AI techniques.

Among those papers, the combination of Artificial Neural

Networks to other techniques (44%, 7), Support Vector

Machine to other techniques (SVM) (25%, 4) and Hidden

Markov-Based Models (e.g. Continuous Hidden Markov

Model-CHMM and Discrete Hidden Markov Model-DHMM)

to other techniques (13%, 2) were the most frequent hybrid

approaches found in the papers selected.

TABLE IV MODIFIED SEMI-AUTONOMOUS DDT SYSTEM MODEL X CT3 CODES

CT.3 - Topic Component of the Modified

DDT System Model (#Hits) %

CT.2 – Impact on Safety: Increase Safety (+)

Sensors and

Perception Machine Perception (21) 36

Navigation and

Control Machine Control (13) 22

Human Factor Human-in-the-loop (6) 10

Fault Prevention

System (+) (8)

32

Conceptual Model

and Framework

CT.2 – Impact on Safety: Decrease Safety (-)

Fault Forecasting

System (-) (11) Ethics and Policies

Dependence and

Trust

Many different combinations of ANNs with other techniques

were found (7 papers). As shown in Table VII most of those

papers are related to Sensors and Perception (3 papers) as well

as Navigation and Control (2 papers). Also, papers related to

Conceptual Model and Framework and Fault Prevention

employed hybrid approach (2 papers). The papers that used a

10

combination of models associated to Hidden Markov Based

Models were related to Navigation and Control as well as

Sensors and Perception. The paper that used Hough

Transformation combined to other models is related to

Navigation and Control. The paper that employed a

combination of techniques to propose a Novel Image

Recognition Technique is related to Sensors and Perception.

The paper using Regression- Based Models combined to other

techniques is related to AV Navigation and Control. Finally, all

the papers employing SVM combined to other techniques were

related to the topic Sensors and Perception. The same grouping

strategy applied to Table V (system-oriented and component-

oriented) can be applied here to evaluate the problems (CT.4).

System-level problems include 16 papers: AV Validation [64],

Machine-learning-based systems validation to the ultra-

dependable levels required for AV [67], Human and Machine

Driver Co-existence [60], Coexistence Human Machine

Controller [70], Driving Car Tasks Classification [61], Lack of

efficient Safety Performance Verification technique when

AI/ML is used [65], Crash assignment, especially between

automated vehicles and non-automated vehicles [69], Reduce

the uncertainty of a driver behavior prediction model [62],

Investigate three underexplored themes for AV research: safety,

interpretability, and compliance [73], How vehicle autonomy

technology can be used to benefit car drivers and also to propose

a concept of an autonomous highway vehicle which improves

highway driving safety [63], AV decisions in complex

dilemmas as a social agent [71], Hybrid (humans and machines)

collective decision-making systems [72], Autonomy assurance

and trust in Automated Transportation Systems [74], AV Test

[66] and, Evaluate the impacts of the number of highly

automated vehicles on future traffic safety and traffic flow [68].

TABLE V TECHNIQUES X DDT SYSTEM MODEL COMPONENT

DDT System Model Technique #Hits %Paper Accum.%

System-oriented

(System)

General AI/ML 12 63% 63%

Hough Transformation related approaches 2 11% 74%

Artificial Neural Networks 2 11% 84%

Optimization Heuristics 1 5% 89%

Estimation Filters (e.g. Kalman Filter and Particle Filters) 1 5% 95%

Linear Temporal Logic (LTL) 1 5% 100%

Component-

oriented

Machine

Perception

Artificial Neural Networks 11 20% 20%

Support Vector Machine (SVM) 8 15% 35%

Hidden Markov Based Models (e.g. Continuous Hidden

Markov Model-CHMM and Discrete Hidden Markov

Model-DHMM)

4 7% 43%

Estimation Filters (e.g. Kalman Filter and Particle Filters) 3 6% 48%

Histogram of Oriented Gradient (HOG) 3 6% 54%

Nearest-Neighbor Based Algorithm (e.g. k-Nearest

Neighbours - kNN) 3 6% 59%

Adaptive Boosting (AdaBoost) 3 6% 65%

Principal Components Analysis (PCA) 2 4% 69%

Haar-like feature detector 2 4% 72%

Fuzzy Logic 2 4% 76%

Viterbi algorithm 1 2% 78%

Bayesian Artificial Intelligence (e.g. Bayesian Deep

Learning, Naive Bayes Classifier-NBC, etc) 1 2% 80%

Regression Based Models 1 2% 81%

Hough Transformation related approaches 1 2% 83%

Ramer-Douglas-Peucker or Ramer-Douglas algorithm 1 2% 85%

Novel Image Recognition Technique 1 2% 87%

Gaussian Mixture Model (GMM) 1 2% 89%

General AI/ML 1 2% 91%

Complex Decision Trees (CDT) 1 2% 93%

Channel Features 1 2% 94%

Local Binary Patterns (LBP) 1 2% 96%

Clustering algorithm k-mean 1 2% 98%

Conditional Random Fields (CRFs) 1 2% 100%

Machine

Control

Fuzzy Logic 3 18% 18%

Support Vector Machine (SVM) 2 12% 29%

Optimization Heuristics 2 12% 41%

Ramer-Douglas-Peucker or Ramer-Douglas algorithm 2 12% 53%

Case-based reasoning (CBR) 1 6% 59%

11

Nearest-Neighbor Based Algorithm (e.g. k-Nearest

Neighbours - kNN) 1 6% 65%

Basic AI Path Planning algorithms such as A* and D* 1 6% 71%

Artificial Neural Networks 1 6% 76%

Regression Based Models 1 6% 82%

Distributed Random Forest (DRF) 1 6% 88%

Neuroevolution of Augmenting Topologies (NEAT) -

ANN + GA 1 6% 94%

Satisfiability Modulo Theories (SMT) Solver 1 6% 100%

Human-in-

the-loop

Bayesian Artificial Intelligence (e.g. Bayesian Deep

Learning, Naive Bayes Classifier-NBC, etc) 2 29% 29%

Regression Based Models 1 14% 43%

Haar-like feature detector 1 14% 57%

Canny Edge Detection Algorithm 1 14% 71%

Hough Transformation related approaches 1 14% 86%

General AI/ML 1 14% 100%

Considering the component-level problems, 21 papers

(36%) are related to dealing with algorithms and techniques

to deal with Machine Perception issues, such as: Vehicle

Cyber Attack [39], Turn Signal Recognition [18], Securing

connected vehicles against Denial of Service (DoS) attack [51],

Road Detection [75], Traffic Light Detection [35], Prediction of

advanced driver assistance systems (ADAS) remaining useful

life (RUL) for the prognosis of ADAS safety critical

components [52], Vehicle Detection and Counting [36],

predicts the CPU usage patterns of software tasks running on a

self-driving car [53], a safety warning and driver-assistance

system and an automatic pilot for rural and urban traffic

environments [21], reliable and robust obstacles detection

continues to be largely investigated and still remains an open

challenge, especially for difficult scenarios and, in general

cases, with loosened constraints and multiple simultaneous use-

cases [25], Pedestrian Detection [27], Road environmental

recognition and various object detection in real driving

conditions [29], Obstacle clustering and tracking [23]. For an

autonomous behavior, each truck must be able to follow the

vehicle ahead. Due to that, each vehicle must be able to

recognize the leading vehicle [22], Speed bump detection [34],

providing road safety to connected drivers and connected

autonomous vehicles [20], how to ”automate” manual

annotation for images to train visual perception for AVs [44],

Road Sign Classification in Real-time [32], Road Terrain

detection [31], Spatio-temporal situation awareness [33],

Pedestrian detection and movement direction recognition [26],

Pedestrian Trajectory Prediction [28], Road junction detection

[30], Cyber Attack in V2X [54], Learn from Demonstration

[45], Early detection of faults or malfunction [55], Road and

Obstacle Detection [24] and Enhance Image Understanding

[19].

The problems related to Machine Control were found in 17

papers (22%). Those problems include: Pre-Crash problem of

Intelligent Control of autonomous vehicles robot [39], Safe-

optimal trajectory selection for autonomous vehicle [76],

Driverless car 100 km experiment [40], Robot maneuvers too

close to an obstacle, which increases the probability of an

accident. Preventing this is crucial in dynamic environments,

where the obstacles, such as other UAVs, are moving [41],

Learning and simulation of the Human-Level decisions

involved in driving a racing car [47], Control intersection

crossing (all way stop) and optimizing it [42], How to prove the

correctness of an algorithm for Vehicle Coordination [43], Path

tracking [48], Drivers maneuver classification [44], AVs

intersections crossing optimization [46] and Manage low level

vehicle actuators (steering throttle and brake) [49].

Finally, the problems related to Human-in-the-loop new

DDT component (Figure 5) are present in 7 papers (10%).

Those problems include: Selective Attention Mechanism [56],

Developing remote controlled car with some automation to deal

with traffic light detection, obstacle avoidance system and lane

detection system to be driven from anywhere over a secured

internet connection [38], Collision avoidance when no action is

taken by driver to avoid the collision [57], Human drivers

monitoring system to ensure they will be able to take over

control within short notice [58] and, Design of driving

assistance system [59]. This seems to be an attention-point; this

problem category can be considered one serious challenge to

semi-autonomous vehicles (SAE Level 1 to Level4). Therefore,

more research is needed into this topic because only 6 papers

were found.

TABLE VI

HETEROGENEITY OF THE USED AI APPROACHES

Heterogeneity % Main Technique #Hits %Papers

Generic 24% General AI/ML 14 100%

Homogenous 49%

Artificial Neural Networks 8 28%

Fuzzy Logic 4 14%

Support Vector Machine (SVM) 3 10%

12

Regression Based Models 2 7%

Estimation Filters (e.g. Kalman Filter and Particle Filters) 2 7%

Bayesian Artificial Intelligence 2 7%

Optimization Heuristics 2 7%

Ramer-Douglas-Peucker or Ramer-Douglas algorithm 2 7%

Hough Transformation 1 3%

Satisfiability Modulo Theories (SMT) Solver 1 3%

Adaptive Boosting (AdaBoost) 1 3%

Linear Temporal Logic (LTL) 1 3%

Hybrid 27%

Artificial Neural Network combined to other techniques 7 44%

Support Vector Machine (SVM) combined to other techniques 4 25%

Hidden Markov Based Models (e.g. Continuous Hidden Markov Model-CHMM

and Discrete Hidden Markov Model-DHMM) combined to other techniques 2 13%

Hough Transformation related approaches combined to other techniques 1 6%

Regression Based Models combined to other techniques 1 6%

Novel Image Recognition Technique 1 6%

V. FINAL REMARKS

Machine Perception has more studies with practical results.

Considering the other components, few studies with practical

results from real deployments were found. Most of the papers

presented preliminary results. In fact, some papers start with a

promise and finish with more promises. Considering only 24%

of the total papers considered in this study were published by

journals, it is possible to conclude the field is not mature yet.

Some similar issues were studied in more than one paper

about Machine Perception, and distinct techniques were applied

to address them (for example, ANN and SVM applied to similar

issues as well issues as well distinct techniques applied to the

topic cyber-attack). Considering some of those techniques have

different working mechanisms, that fact can be an important

finding for the safety of autonomous cars as regards the need of

redundant components. Similar issues being addressed by

different techniques were not identified for Machine Control.

The papers related to Navigation and Control also reported

positive and promising results, although the level of maturity of

the achievements are clearly much lower than the sensors and

perception as well as far from what would be expected for an

autonomous vehicle considering the potential hazardous

situations it may face. In fact, most of the results presented are

preliminary.

Only few of the studies related to system described practical

results from real deployments. The papers proposing

conceptual models and frameworks bring important

contributions, but they are mostly not tested in real set-ups.

There is thus a lack of reported results from models and

frameworks that could build the foundation of AVs safety.

Also, few human-in-the-loop studies had practical results from

real deployments. However, they seem to be one of the most

important topics seeing that there will be more semi-

autonomous cars than fully autonomous ones for a while, and

they will co-exist. The human factor will thus be an important

variable in the system to be considered not only as the impacted

side of the safety, but as one of the sources of interactions

influencing the safety levels. The topic requires

multidisciplinary studies involving fields beyond engineering

and computer science, such as neurosciences. This shows the

field is not mature yet.

The amplitude and range of the reported future researches in

the papers reviewed suggest that there is an empty space for

new research into this field. For example, only few studies were

found about the three topics positioning AI as a potential source

of negative impact on safety - Fault Forecasting, Ethics and

Policies, and Dependability and Trust. When combined to the

other findings reported by the present study, it confirms the

impressions formed during an exploratory research of the

literature [1]. It reinforces the perception that the field of AI and

AV is not heavily influenced by the safety engineering culture

yet. In fact, the studies published about this current topic seem

to be more driven by computation-related domains, with no

tradition regarding safety culture, than other fields that are

much more connected to safety in critical systems [1].

Additional research is necessary for most of the studies

reviewed. They need to be extended to be tested in simulated or

real set-ups, new and broader scenarios, with new and more

data, and consider experimental designs whereby the results

from the proposed approach are compared to benchmarks and

alternative techniques. Many AI techniques have achieved

impressive results; however, it is still arguable whether the error

rates are suitable for real deployments in AVs under the light of

a (missing) hazard analysis. Therefore, additional studies with

improvements in those techniques are required. Finally, a

stronger influence of safety engineering on most of the studies

would benefit the field.

A research agenda must consider a serious safety agenda for

future studies, at system-level, component-level and AI

technique-level. In fact, there are some topics related to safety

concerns over AVs, which are critical-path to the development

of the field. Some of the suggested topics are related to the

challenges with validating machine-learning- based systems to

the ultra-dependable levels required for AVs; wider and deeper

studies about human-machine collaboration in the context of

AVs; autonomy assurance and trust in AVs; ethical and moral

decisions in the context of AVs; among other topics, from

13

Validating machine-learning-based systems to the ultra-

dependable levels required for AVs; and autonomy assurance

and trust in AVs seem to be the holy grail towards a fully

autonomous AV - SAE level 5 . They are also key topics for the

Safety Certification of non-deterministic control systems. In

those contexts, there are many gaps to be filled by future

researches, such as AVs software testing, Fault Injection

Testing for AI on AVs, Failure Modes and Effects Analysis

(FMEA) for AI on AVs, AI safeguards for AVs, AI safety

envelopes for AVs, AI redundancy for AVs (many possible

approaches, such as a hybrid connectionist and symbolic

architecture using causal inference), explainable AI for AVs, AI

fault forecasting. Finally, studies on V2X communication can

help autonomy assurance by providing channels for hardware

and software redundancy. Human-machine collaboration in the

context of AVs is another key topic with special impact on the

semi-autonomous vehicles (SAE levels 1 to 4). Investigations

on the best way humans and AVs can interact during normal

operations and facing hazardous situations are needed to meet

the adequate safety requirements the semi-autonomous vehicles

must have. Those studies must consider hybrid collective

decision-making systems to enable humans and machines to

work together and to agree on common decisions, as well as

how to deal with the lack of agreement in some situations.

Also, there is another important discussion arising in the

context of human-machine collaboration that must be

investigated. On the one hand, there are reports about advanced

driver assistant technologies that failed (such as Tesla

Autopilot) and the driver was not able to react in time to avoid

the accident. They ended-up in life losses and property

damages. On the other hand, there are reports about situations

in which the advanced driver assistant technologies saved the

drivers’ life by automatically taking the driver suffering a heart

attack to the hospital; fully controlling the car with a drunk

driver sleeping; and using a defensive lane change

TABLE VII HYBRID AI APPROACHES X TOPIC

Main AI Technique Topic AI Techniques Reference

Artificial Neural

Network

Conceptual Model

and Framework

HoughTransforms, HoughLines, LocalMaximaFinder,

Kalman filters and Convolutional Neural Network (CNN) [60]

Fault Prevention KNN, SVM Regression (SMO), ANN [52]

Navigation and

Control

CBR, ANN, fuzzy logic, Nearest-Neighbor Retrieval

Algorithm, Basic AI Path Planning algorithms such as A* and

D*

[39]

ANN combined to Genetic Algorithm - Neuroevolution of

Augmenting Topologies (NEAT) [47]

Sensors and

Perception

ANNs, AdaBoost, SVM, Hidden Markov Models (HMMs),

CRFs [30]

Clustering algorithm k- mean, ANN [31]

HOG, SVM, PCA, ANN [29]

Hidden Markov Based

Models

Navigation and

Control

GMM, Continuous Hidden Markov Model (CHMM),

Discrete Hidden Markov Model (DHMM) [45]

Sensors and

Perception

HMM, Viterbi algorithm, Adaboost trained Haar-like feature

detector [36]

Hough Transformation Navigation and

Control

Haar Feature Based Cascade Classifier, Canny edge detection

and Hough line transformation [38]

Novel Image

Recognition Technique

Sensors and

Perception Combination of mathematical techniques [34]

Regression Based

Models

Navigation and

Control (DRF) and Linear Regression (LR) [76]

Support Vector

Machine (SVM)

Sensors and

Perception

Haar, HOG, LBP, Chanel features, SVM [37]

k-Nearest Neighbours (kNN), Naïve Bayes classifier (NBC),

SVM [27]

Principal component analysis network (PCANet), SVM [35]

SVM, HOG [75]

maneuver to avoid being hit by a truck changing its lane. Some

players in the industry are pushing the automation evolution

steps towards full automation by requiring the human driver to

be a backup to the automated driver. Other players in the

industry believe the automated driver must be a backup to the

human driver. It looks like the second approach can be a

smoother and safer path towards a SAE level-5 automation.

Immersive environments for training and testing AVs

represent another research trend. As the underlying

technologies supporting AVs development evolve, higher

automation-levels become possible. Considering the potential

hazards until the AVs are well trained and fine-tuned, the

14

immersive technologies are becoming an important tool to

support the development, training and tests of fully autonomous

machines. Another broad topic requiring further research is

related to ethical and moral decisions in AVs. Some studies

only mention issues related to moral dilemmas while others

provide some simple experiments involving simulated

environments and/or human interviews. However, they

misinterpret important concepts and bring the discussions

around the decisions AVs must make when life losses are

involved, besides the moral and ethical perceptions from the

human perspective. All of them miss important points such as

statistical considerations and the societal result. In other words,

the discussions are not deep enough as regards situations such

as whether an AV should hit an old man or a child, while a true

safe machine control should consider all the probabilities

involved and select the one that minimizes the chances of life

losses instead of just picking an option. For example, the system

must consider small signals, such as which of the potential

victims is paying attention to the approaching AV and what

would their potential reaction be and effectiveness of it based

on the age and other metrics, as well, considering the multiple

scenarios, and the configuration of each, such as speed, region

of the car hitting which region of each victim, the potential

damages and the severity of the damages considering the

estimated weight and overall physical condition, to decide

based on the minimization of chances of life losses. This

approach will result into higher safety levels for society.

Finally, only 1 paper about autonomous truck was found.

Considering some specificities of autonomous truck and its

risks, at least a few more studies about the topic could be

expected.

REFERENCES

[1] A. M. Nascimento, L. F. Vismari, P. S. Cugnasca, J. B. Camargo, J. R. de Almeida, R. Inam, E. Fersman, A. Hata, and M. V. Marquezini.

Concerns on the differences between ai and system safety mindsets impacting autonomous vehicles safety. In Barbara Gallina, Amund

Skavhaug, Erwin Schoitsch, and Friedemann Bitsch, editors, Computer

Safety, Reliability, and Security, pages 481–486, Cham, 2018. Springer

International Publishing.

[2] Eusebio Scornavacca, Stuart J. Barnes, and Sid L. Huff. Mobile business research published in 2000-2004: Emergence, current status, and future

opportunities. CAIS, 17:28, 2006.

[3] Oswald Jones and Caroline Gatrell. Editorial: The future of writing and reviewing for ijmr. International Journal of Management Reviews,

16(3):249–264, 2014.

[4] Jane Webster and Richard T. Watson. Analyzing the past to prepare for the future: Writing a literature review. MIS Q., 26(2):xiii–xxiii, June

2002.

[5] Sara Abbaspour Asadollah, Daniel Sundmark, Sigrid Eldh, Hans HansS son, and Wasif Afzal. 10 years of research on debugging concurrent and multicore software: A systematic mapping study. Software Quality

Journal, 25(1):49–82, March 2017.

[6] Kai Petersen, Robert Feldt, Shahid Mujtaba, and Michael Mattsson.Systematic mapping studies in software engineering. In

Proceedings of the 12th International Conference on Evaluation and

Assessment in Software Engineering, EASE’08, pages 68–77, Swindon, UK, 2008. BCS Learning & Development Ltd.

[7] Yair Levy and Timothy J. Ellis. A systems approach to conduct an effective literature review in support of information systems research. Informing Science: The International Journal of an Emerging TransdisI

cipline, 9:181–212, January 2006.

[8] Jan vom Brocke, Alexander Simons, Bjoern Niehaves, Bjorn Niehaves, Kai Riemer, Ralf Plattfaut, and Anne Cleven. Reconstructing the giant: On the importance of rigour in documenting the literature search process.

In Susan Newell, Edgar Whitley, Nancy Pouloudi, Jonathan Wareham,

and Lars Mathiassen, editors, Information systems in a globalising world : challenges, ethics and practices ; ECIS 2009, 17th European

Conference on Information Systems, pages 2206–2217, Verona, June

2009. Universita` di Verona, Facolta` di Economia, Departimento de

Economia Aziendale.

[9] Lars Mathiassen, Mike Chiasson, and Matt Germonprez. Style composition in action research publication. MIS Q., 36(2):347–363, June

2012.

[10] Barbara Kitchenham. Procedures for performing systematic reviews. Keele university. technical report tr/se-0401, Department of Computer

Science, Keele University, UK, 2004.

[11] Eusebio Scornavacca Jr. and Stuart J. Barnes. M-banking services in japan: a strategic perspective. Int. J. Mob. Commun., 2(1):51–66, May

2004.

[12] Jason Brownlee. A tour of machine learning algorithms. Machine Learning Mastery, 2013.

[13] Abdullah Gani, Aisha Siddiqa, Shahaboddin Shamshirband, and Fariza Hanum. A survey on indexing techniques for big data: taxonomy and performance evaluation. Knowledge and Information Systems,

46(2):241–284, Feb 2016. [14] Mark Hall, Eibe Frank, Geoffrey Holmes, Bernhard Pfahringer, Peter

Reutemann, and Ian H. Witten. The weka data mining software: An

update. SIGKDD Explor. Newsl., 11(1):10–18, November 2009. [15] Baharum Baharudin, Lam Hong Lee, Khairullah Khan, and Aurangzeb

Khan. A review of machine learning algorithms for text-documents

classification. Journal of Advances in Information Technology, 1, 02 2010.

[16] Stuart Russell and Peter Norvig. Artificial Intelligence: A Modern Approach. Prentice Hall Press, Upper Saddle River, NJ, USA, 3rd edition, 2009.

[17] Ian H. Witten, Eibe Frank, and Mark A. Hall. Data Mining: Practical Machine Learning Tools and Techniques. Morgan Kaufmann Publishers Inc., San Francisco, CA, USA, 3rd edition, 2011.

[18] K. Yoneda, A. Kuramoto, and N. Suganuma. Convolutional neural network based vehicle turn signal recognition. In 2017 International Conference on Intelligent Informatics and Biomedical Sciences (ICIC

IBMS), pages 204–205, Nov 2017.

[19] Robert L Simpson. Strategic Computing Computer Vision: Taking Image Understanding To The Next Plateau. In Image Understanding and

the Man-Machine Interface, volume 758, pages 52–60. International

Society for Optics and Photonics, 1987. [20] B. Kihei, J. A. Copeland, and Y. Chang. Automotive doppler sensing:

The doppler profile with machine learning in vehicle-to-vehicle networks

for road safety. In 2017 IEEE 18th International Workshop on Signal Processing Advances in Wireless Communications (SPAWC), pages 1–

5, July 2017.

[21] Qing Li, Nanning Zheng, and Hong Cheng. Springrobot: a prototype autonomous vehicle and its algorithms for lane detection. IEEE

Transactions on Intelligent Transportation Systems, 5(4):300–308, Dec

2004. [22] Wilfried Woeber, Martin Kefer, W Kubinger, and D Szuegyi. Evaluation

of daylight and thermal infra-red based detection for platooning vehicles.

23rd DAAAM International Symposium on Intelligent Manufacturing and Automation 2012, 2:719–722, 01 2012.

[23] R. Dominguez, E. Onieva, J. Alonso, J. Villagra, and C. Gonzlez. Lidar based perception solution for autonomous vehicles. In 2011 11th In2 ternational Conference on Intelligent Systems Design and Applications,

pages 790–795, Nov 2011.

[24] Gianni Vernazza Daniele D. Giusto, Carlo S. Regazzoni. Multisensorial vision for autonomous vehicle driving. volume 1095, pages 1095 – 1095

– 8, 1989.

[25] M. Allodi, A. Broggi, D. Giaquinto, M. Patander, and A. Prioletti. Machine learning in tracking associations with stereo vision and lidar

observations for an autonomous vehicle. In 2016 IEEE Intelligent

Vehicles Symposium (IV), pages 648–653, June 2016. [26] Alex Dominguez-Sanchez, Sergio Orts-Escolano, and Miguel Cazorla.

Recognizing pedestrian direction using convolutional neural networks.

In Ignacio Rojas, Gonzalo Joya, and Andreu Catala, editors, Advances in Computational Intelligence, pages 235–245, Cham, 2017. Springer

International Publishing.

15

[27] Navarro Pedro J.; Fernndez C.; Borraz R.; and Alonso D. A machine learning approach to pedestrian detection for autonomous vehicles using high-definition 3d range data. Sensors (Basel), 17(1), 2016.

[28] Julian Bock, Jens Kotte, Till Beemelmanns, and Markus Klsges. Self- learning trajectory prediction with recurrent neural networks at intelligent intersections. In Proceedings of the 3rd International

Conference on Vehicle Technology and Intelligent Transport Systems -

Volume 1: VEHITS, pages 346–351. INSTICC, SciTePress, 2017. [29] J. Jeon, S. Hwang, and H. Moon. Monocular vision-based object

recognition for autonomous vehicle driving in a real driving

environment. In 2016 13th International Conference on Ubiquitous Robots and Ambient Intelligence (URAI), pages 393–399, Aug 2016.

[30] D. Habermann, C. E. O. Vido, F. S. Osrio, and F. Ramos. Road junction detection from 3d point clouds. In 2016 International Joint Conference on Neural Networks (IJCNN), pages 4934–4940, July 2016.

[31] F. Ebadi and M. Norouzi. Road terrain detection and classification algorithm based on the color feature extraction. In 2017 Artificial Intelligence and Robotics (IRANOPEN), pages 139–146, April 2017.

[32] S. Hamdi, H. Faiedh, C. Souani, and K. Besbes. Road signs classification by ann for real-time implementation. In 2017 International Conference on Control, Automation and Diagnosis (ICCAD), pages 328–332, Jan

2017.

[33] Eshed Ohn-Bar and Mohan Manubhai Trivedi. Are all objects equal? Deep spatio-temporal importance prediction in driving videos. Pattern

Recognition, 64:425 – 436, 2017.

[34] W. Devapriya, C. N. K. Babu, and T. Srihari. Advance driver assistance system (adas) - speed bump detection. In 2015 IEEE International

Conference on Computational Intelligence and Computing Research (ICCIC), pages 1–6, Dec 2015.

[35] Z. Chen and X. Huang. Accurate and reliable detection of traffic lights using multiclass learning and multiobject tracking. IEEE Intelligent Transportation Systems Magazine, 8(4):28–42, winter 2016.

[36] N. Miller, M. A. Thomas, J. A. Eichel, and A. Mishra. A hidden markov model for vehicle detection and counting. In 2015 12th Conference on Computer and Robot Vision, pages 269–276, June 2015.

[37] Antonio M. Lopez, Gabriel Villalonga, Laura Sellart, Germn Ros, David Vzquez, Jiaolong Xu, Javier Marn, and Azadeh Mozafari. Training my car to see using virtual worlds. Image and Vision Computing, 68:102 –

118, 2017. Automotive Vision: Challenges, Trends, Technologies and

Systems for Vision-Based Intelligent Vehicles. [38] M. R. T. Hossai, M. A. Shahjalal, and N. F. Nuri. Design of an iot based

autonomous vehicle with the aid of computer vision. In 2017

International Conference on Electrical, Computer and Communication Engineering (ECCE), pages 752–756, Feb 2017.

[39] B. M. Elbagoury, A. M. Salem, and L. Vladareanu. Intelligent adaptive precrash control for autonmous vehicle agents (cbr engine amp; hybrid a path planner). In 2016 International Conference on Advanced

Mechatronic Systems (ICAMechS), pages 429–436, Nov 2016.

[40] Jorge Godoy, Joshu Prez, Enrique Onieva, Jorge Villagr, Vicente Milans, and Rodolfo Haber. A driverless vehicle demonstration on motorways

and in urban environments. Transport, 30(3):253–263, 2015.

[41] Ondrej Linda, Todd Vollmer, and Milos Manic. Svm-inspired dynamic safe navigation using convex hull construction. 2009 4th IEEE Confer2

ence on Industrial Electronics and Applications, ICIEA 2009, 05 2009.

[42] I Zohdy and Hesham Rakha. Optimizing driverless vehicles at intersections. pages AM–00071, 01 2012.

[43] Mikael Asplund, Atif Manzoor, Melanie´ Bouroche, Siobhan` Clarke, and Vinny Cahill. A formal approach to autonomous vehicle coordination. In Dimitra Giannakopoulou and Dominique Mery´,

editors, FM 2012: Formal Methods, pages 52–67, Berlin, Heidelberg,

2012. Springer Berlin Heidelberg. [44] Erwin Jose Lopez Pulgarin, Guido Herrmann, and Ute Leonards.

Drivers’ manoeuvre classification for safe hri. In Yang Gao, Saber FalD

lah, Yaochu Jin, and Constantina Lekakou, editors, Towards Autonomous Robotic Systems, pages 475–483, Cham, 2017. Springer

International Publishing.

[45] Ge, F., Moore, W., & Antolovich, M. (2015, November). Learning from demonstration using gmm, CHMM and DHMM: A comparison. In

Australasian Joint Conference on Artificial Intelligence (pp. 204-217).

Springer, Cham. [46] I. H. Zohdy, R. K. Kamalanathsharma, and H. Rakha. Intersection

management for autonomous vehicles using icacc. In 2012 15th

International IEEE Conference on Intelligent Transportation Systems, pages 1109–1114, Sep. 2012.

[47] Francisco Gallego, Faraon´ Llorens, Mar Pujol, and Ramon´ Rizo. Learning human-level ai abilities to drive racing cars. In Proceedings of the 2005 Conference on Artificial Intelligence Research and

Development, pages 33–40, Amsterdam, The Netherlands, The

Netherlands, 2005. IOS Press. [48] Citlalli Gamez´ Serna, Alexandre Lombard, Yassine Ruichek, and

Abdeljalil Abbas-Turki. Gps-based curve estimation for an adaptive pure

pursuit algorithm. In Grigori Sidorov and Oscar Herrera-Alcantara,´ editors, Advances in Computational Intelligence, pages 497–511, Cham,

2017. Springer International Publishing.

[49] J.E. Naranjo, F. Jimnez, O. Gmez, and J.G. Zato. Low level control layer definition for autonomous vehicles based on fuzzy logic. Intelligent

Automation & Soft Computing, 18(4):333–348, 2012.

[50] Algirdas Avizienis, Jean-Claude Laprie, Brian Randell, and Carl Landwehr. Basic concepts and taxonomy of dependable and secure

computing. IEEE Trans. Dependable Secur. Comput., 1(1):11–33,

January 2004. [51] P. Sharma, Hong Liu, Honggang Wang, and Shelley Zhang. Securing

wireless communications of connected vehicles with artificial intelliw

gence. In 2017 IEEE International Symposium on Technologies for Homeland Security (HST), pages 1–7, April 2017.

[52] Mostafa Anwar Taie, Eman Magdy Moawad, Mohammed Diab, and Mohamed ElHelw. Remote diagnosis, maintenance and prognosis for advanced driver assistance systems using machine learning algorithms.

SAE Int. J. Passeng. Cars Electron. Electr. Syst., 9:114–122, 04 2016.

[53] Y. Seo, J. Kim, and R. Rajkumar. Predicting dynamic computational workload of a self-driving car. In 2014 IEEE International Conference

on Systems, Man, and Cybernetics (SMC), pages 3030–3035, Oct 2014. [54] Shirin Noei, Arman Sargolzaei, Alireza Abbaspour, and Kang Yen. A

decision support system for improving resiliency of cooperative adaptive

cruise control systems. Procedia Computer Science, 95:489 – 496, 2016. Complex Adaptive Systems Los Angeles, CA November 2-4, 2016.

[55] Newman R.M. Gaura, E. Microsensors, arrays and automatic diagnosis of sensor faults. In Technical Proceedings of the 2003 Nanotechnology Conference and Trade Show, volume 1, pages 276 – 279. TechConnect

Briefs, February 2003.

[56] Quanzhen Guan, Hong Bao, and Zuxing Xuan. The research of prediction model on intelligent vehicle based on driver’s perception.

Cluster Computing, 20(4):2967–2979, Dec 2017.

[57] Ata M. Khan. Design of real-time transition from driving assistance to automation function: Bayesian artificial intelligence approach. In Jan

Fischer-Wolfarth and Gereon Meyer, editors, Advanced Microsystems

for Automotive Applications 2014, pages 39–48, Cham, 2014. Springer International Publishing.

[58] Hamidur Rahman, Shahina Begum, and Mobyen Uddin Ahmed. Driver monitoring in the context of autonomous vehicle. In The 13th ScandiT navian Conference on Artificial Intelligence, November 2015.

[59] Ata Khan. Cognitive vehicle design guided by human factors and supported by bayesian artificial intelligence. In Neville A Stanton, editor, Advances in Human Aspects of Transportation, pages 362–372, Cham,

2018. Springer International Publishing.

[60] P. K. Gurudatt and V. Umesh. Novel architecture for cloud based next gen vehicle platform transition from today to 2025. In 2017 IEEE

International Conference on Consumer Electronics-Asia (ICCE-Asia),

pages 151–155, Oct 2017. [61] Silviu Ionita. Autonomous vehicles: from paradigms to technology. IOP

Conference Series: Materials Science and Engineering, 252(1):012098,

2017. [62] A. Rakotonirainy, O. Orfila, and D. Gruyer. Reducing driver’s

behavioral uncertainties using an interdisciplinary approach:

Convergence of quantified self, automated vehicles, internet of things and artificial intelligence. IFAC-PapersOnLine, 49(32):78 – 82, 2016.

Cyber-Physical & Human-Systems CPHS 2016.

[63] Akihiro Okuno, Atsushi Kutami, and Kenji Fujita. Towards autonomous cruising on highways. 08 1990.

[64] Philip Koopman. Challenges in Autonomous Vehicle Validation: Keynote Presentation Abstract. In Proceedings of the 1st International Workshop on Safe Control of Connected and Autonomous Vehicles,

page 3. ACM, 2017.

[65] D. Meltz and H. Guterman. Robil israeli program for research and development of autonomous ugv: Performance evaluation methodology.

In 2016 IEEE International Conference on the Science of Electrical

Engineering (ICSEE), pages 1–5, Nov 2016. [66] C. Wolschke, T. Kuhn, D. Rombach, and P. Liggesmeyer. Observation

based creation of minimal test suites for autonomous vehicles. In 2017

16

IEEE International Symposium on Software Reliability Engineering

Workshops (ISSREW), pages 294–301, Oct 2017. [67] P. Koopman and M. Wagner. Autonomous vehicle safety: An

interdisciplinary challenge. IEEE Intelligent Transportation Systems

Magazine, 9(1):90–96, Spring 2017. [68] M. Bahram, Z. Ghandeharioun, P. Zahn, M. Baur, W. Huber, and F.

Busch. Microscopic traffic simulation based evaluation of highly

automated driving on highways. In 17th International IEEE Conference on Intelligent Transportation Systems (ITSC), pages 1752–1757, Oct

2014.

[69] Wesley Kumfer and Richard Burgess. Investigation into the role of rational ethics in crashes of automated vehicles. Transportation Research

Record: Journal of the Transportation Research Board, 2489:130–136,

01 2015. [70] G. Lugano. Virtual assistants and self-driving cars. In 2017 15th

International Conference on ITS Telecommunications (ITST), pages 1–

5, May 2017. [71] Ebru Dogan, Raja Chatila, Stephane Chauvier, Katherine Evans, Petria

Hadjixenophontos, and Jerome Perrin. Ethics in the design of automated

vehicles: The avethics project. In EDIA@ECAI, 2016. [72] John Tasioulas Kristen Brent Venable Brian Williams Joshua Greene,

Francesca Rossi. Embedding ethical principles in collective decision

support systems. In Proceedings of the Thirtieth AAAI Conference on Artificial Intelligence (AAAI-16), 2016.

[73] Rowan McAllister, Yarin Gal, Alex Kendall, Mark van der Wilk, Amar Shah, Roberto Cipolla, and Adrian Weller. Concrete problems for autonomous vehicle safety: Advantages of bayesian deep learning. In

Proceedings of the Twenty-Sixth International Joint Conference on

Artificial Intelligence, IJCAI-17, pages 4745–4753, 2017. [74] R.; Lacher, A.; Grabowski and S. Cook. Autonomy, trust, and

transportation. In AAAI Spring Symposium Series, 2014.

[75] Keisuke Kazama, Yasuhiro Akagi, Pongsathorn Raksincharoensak, and Hiroshi Mouri. Fundamental study on road detection method using

multi-layered distance data with hog and svm. Journal of robotics and

mechatronics, 28(6):870–877, dec 2016. [76] H. A. Najada and I. Mahgoub. Autonomous vehicles safe-optimal

trajectory selection based on big data analysis and predefined user

preferences. In 2016 IEEE 7th Annual Ubiquitous Computing, Electronics Mobile Communication Conference (UEMCON), pages 1–

6, Oct 2016.

[77] Mohammad Ali Abam and Mark de Berg. Kinetic sorting and kinetic convex hulls. In Proceedings of the Twenty-first Annual Symposium on

Computational Geometry, SCG ’05, pages 190–197, New York, NY,

USA, 2005. ACM. [78] Jerry Engelman Charles Green Charles Kawashima Steve Kiger Danil

Prokhorov Levasseur Tellis Barbara Wendling Andy Christensen, AnP

drew Cunningham. Key considerations in the development of driving automation systems. Gothenburg, Sweden, 2015.

[79] Neil R. Storey. Safety Critical Computer Systems. Addison-Wesley Longman Publishing Co., Inc., Boston, MA, USA, 1996.

[80] Salvendy G, editor. Handbook of human factors and ergonomics. John Wiley & Sons; 2012 May 24

[81] Norman, D. (2013). The design of everyday things: Revised and expanded edition. Basic books.

APPENDIX

TABLE VIII

CT.3 X CT.4

Technique

(CT.3) Hits Papers Addressed Problem (CT.4) References

General AI/ML 14 24%

AV Validation; Challenge with validating machine-learning based

systems to the ultra-dependable levels required for autonomous

vehicle; Coexistence Human Machine Controller; Driving Car Tasks

Classification; Lack of efficient Safety Performance Verification

technique when AI/ML is used; Crash assignment, especially between

automated vehicles and non-automated vehicles; Reducing the

uncertainty of a driver behavior prediction model; Integration between

automatic vehicle and human driver; How t h e vehicle autonomy

technology can be used to benefit car drivers and t o improve highway

driving safety by a concept of an autonomous highway vehicle; AV

decisions in complex dilemmas as a social agent; Hybrid (humans and

machines) collective decision making systems (work together and

agree on common decisions); Autonomy assurance and trust

(CERTIFICATION PROA CESS) in Automated Transportation

Systems; Evaluating the impacts of the number of highly automated

vehicles on future traffic safety and traffic flow; Enhancing Image

Understanding.

[64],[67],

[70],[61],

[65],[69],

[62],[58],

[63],[71],

[72],[74],

[68],[19]

Artificial

Neural Networks

(ANN)

13 22%

Vehicle Cyber Attack; Turn Signal Recognition; Pre-crash issues of

Intelligent Control of autonomous vehicles robot; Real-time Road

Sign Classification; Road Terrain detection; Spatio-temporal situation

awareness; Pedestrian detection and movement direction recognition;

Pedestrian Trajectory Prediction; Road junction detection; Early faults

or malfunction detection; Prediction of advanced driver assistance

systems (ADAS) remaining useful life (RUL) for the prognosis of

ADAS safety critical components; Road environmental recognition and

various objects detection in real driving conditions; Human and

Machine Driver Co-existence;

[39],[18],

[32],[31],

[33],[26],

[28],[30],

[55],[52],

[29],[60]

Support

Vector Machine

(SVM)

10 17%

Road Detection; Robot maneuvers too close to an obstaR cle; Road

environmental recognition and various object detection in real driving

conditions; Drivers maneuver classification; Traffic Light Detection;

[75],[41],

[29],[20],

[44],[35],

17

Prediction of adc vanced driver assistance systems (ADAS) remaining

useful life (RUL) for the prognosis of ADAS safety critical components

Pedestrian Detection; How to ”automate” manual annotation for

images to train visual perception for AVs Road junction detection;

[52],[27],

[37],[30]

Bayesian

Artificial

Intelligence

4 7%

Collision avoidance when no action is taken by driver; Safety,

interpretability, and compliance; Pedestrian Detection; Design of

driving assistance system;

[57],[73],

[27],[59]

Fuzzy Logic 5 8%

PreCrash problem of Intelligent Control of autonomous vehicles robot;

Driverless car 100 km experiment Cyber Attack in V2X; Manage

low level vehicle actuators (steering throttle and brake); Road and

Obstacle Detection;

[39],[40],

[54],[49],

[23]

Hidden

Markov

Based Models

4 7% Vehicle Detection and Counting; Road junction detection; Learn from

Demonstration;

[36],[29],

[45]

Estimation Filters 4 7%

Human and Machine Driver Co-existence; Securing connected vehicles

against Denial of Service (DoS) attack; Reliable and robust obstacles

detection;

[60],[51],

[24]

Nearest

Neighbour-Based

Algorithm

4 7%

Pre-crash problem of Intelligent Control of autonomous vehicles

robot; Pedestrian Detection; Providing road safety to connected

drivers and connected autonomous vehicles;

[39],[26],

[19]

Adaptive

Boosting

(AdaBoost)

3 5% Vehicle Detection and Counting; Leading vehicle recogV nition in

platooning; Road junction detection;

[36],[21],

[29]

Ramer-Douglas

Peucker

or Ramer-

Douglas

algorithm

3 5% Obstacle clustering and tracking; Path tracking; [22],[48]

Haar-like

feature

detector

3 5%

Developing remotecontrolled car with some automation to deal with

traffic light detection, obstacle avoidance system and lane detection

system to be driven from anywhere over a secured internet

connection; Vehicle Detection and Counting; How to ”automate”

manual annotation for images to train visual perception for AVs;

[38],[36],

[37]

Histogram

of Oriented

Gradient

(HOG)

3 5%

Road Detection; Road environmental recognition and various objects

detection in real driving conditions; How to ”automate” manual

annotation for images to train visual perception for AVs;

[75],[28],

[37]

Hough

Transfor-

mation

3 5%

Road Detection; Road environmental recognition and various object

detection in real driving conditions; How to ”automate” manual

annotation for images to train visual perception for AVs;

[60],[38],

[20]

Optimiza-tion

Heuristics 3 5%

Control intersection crossing (all way stop) and op- timization;

Autonomous vehicles intersections crossing optimization; Human and

Machine Driver Co-existence;

[46],[42],

[60]

Regression-

Based

Models

3 5%

Selective Attention Mechanism; Safe-optimal trajectory selection for

autonomous vehicles; Predicts the CPU usage patterns of software tasks

running on a self-driving car;

[56],[79],

[53]

Principal

Componen-ts

Analysis (PCA)

2 3% Traffic Light Detection; Road environmental recognition and various

object detection in real driving conditions; [35],[28]

Canny Edge

De- tection

Algorithm

1 2%

Developing remote- controlled car with some automation to deal

with traffic light detection, obstacle avoidance system and lane

detection system to be driven from anywhere over a secured internet

connection;

[38]

Case-based

reasoning

(CBR)

1 2% Pre-crash problem of Intelligent Control of autonomous vehicles robot; [39]

Channel

Features 1 2%

How to ”automate” manual annotation for images to train visual

perception for AVs; [37]

18

Clustering

Algo- rithm k-

mean

1 2% Road Terrain detection; [30]

Complex

Decision Trees

(CDT)

1 2% Providing road safety to connected drivers and connected autonomous

vehicles; [19]

Conditional

Random

Fields (CRFs)

1 2% Road junction detection; [29]

Distributed

Random Forest

(DRF)

1 2% Safe-optimal trajectory selection for autonomous vehicle; [79]

Gaussian

Mixture Model

(GMM)

1 2% Learn from Demonstration; [45]

Linear

Temporal

Logic (LTL)

1 2% AV Test; [66]

Local Binary

Patterns (LBP) 1 2%

How to ”automate” manual annotation for images to train visual

perception for AVs; [37]

Neuroevo-lution

of Augmen-

ting

Topologies

(NEAT)

1 2% Learning and simulation of the Human-Level decisions involved in

driving a racing car; [47]

Novel

Image

Recogni-tion

Technique

1 2% Speed bump detection; [34]

Path Planning

Al- gorithms

(A* and D*)

1 2% PreCrash problem of Intelligent Control of autonomous vehicles robot; [39]

Satisfiability

Modulo

Theories

(SMT) Solver

1 2% How to prove the correctness of an algorithm for Vehicle Coordination; [43]

Viterbi

Algorithm 1 2% Vehicle Detection and Counting; [36]

TABLE IX

FINDINGS ON PAPERS ORIENTED BY THE DISCUSSION

Issue Suggested

Approach AI

Technique Findings Reference

DDT System Model Component - System

Human and Machine

Driver Co-existence

Continuously monitor

the driving behavior of

the neighboring vehicles,

sensor behavior and

processor behavior of the

ego vehicle regardless of

the vehicle being

autonomous or not

ANN (Hybrid)

The authors have foreseen and proposed

the solutions for future problems, which

would occur while the autonomous vehicles

are a part of driving. All the three described

architectures have addressed the safety

related problems. The architectures are

based on the availability of the resources for

vehicles. The first two architectures address

the safety failure due to the human

ignorance and autonomous vehicle

behavior. Third architecture addresses the

way of securing the failed vehicle due to

system failure. All the architectures rely

highly on the efficient connectivity and

[60]

19

computer vision algorithms.

Investigate three

under-explored

themes for AV

research: safety,

interpretability, and

compliance

End-to-end Bayesian

deep learning

architecture to propagate

uncertainty throughout

the AV framework. In

this case, standard deep

learning makes hard

predictions, whereas

Bayesian deep learning

outputs probabilistic

predictions accounting

for each model’s

ignorance about the

world.

ANN (Hybrid)

3 critical themes for a smooth adoption of

AV systems by society were highlighted.

Hard decisions are dangerous. Soft

(uncertain) classifications should be

propagated through to the decision layer.

This enables the AV to act more cautiously

in the event of greater uncertainty. We also

discussed the themes. Authors argument

interpretability and compliance as ways to

build trust and mitigate fears which

passengers might otherwise reasonably

have about unfamiliar black-box AV

systems. Also, they discussed about the

importance of clear metrics to evaluate each

component’s probabilistic output based on

their ultimate effect on the vehicle’s

performance.

[73]

AV Validation Safety Envelopes General

AI/ML No findings, only suggested approaches. [64]

Challenge with

validating machine-

learning based

systems to the ultra-

dependable levels

required for

autonomous vehicle

Safety certification

strategy addressing the

cross-disciplinary

concerns of safety

engineering, hardware

reliability, software

validation, robotics,

security, testing, human-

computer interaction,

social acceptance, and a

viable legal framework.

General

AI/ML

The paper only points out the challenge

with validating machine-learning based

systems to the ultra-dependable levels

required for autonomous vehicle fleets, and

how that challenge relates to a number of

other areas. It does not provide any

particular finding.

[67]

Coexistence Human

Machine Controller

In-car Virtual

Assistants

General

AI/ML

It is too early to assess whether

carmakers' optimistic vision of in-car virtual

assistants will match users' experience or

follow a similar destiny of "Clippy the

Paperclip" from Microsoft Word.

Proprietary systems are predominant and

there is no common framework addressing

ethical principles, liability, data protection,

privacy and security on many of the

technologies associated to AVs. Due to

these shortcomings, allowing a fully

autonomous approach could generate more

drawbacks than benefits. At least in a first

phase, it seems more appropriate to apply

AI-based virtual assistants to autonomously

execute tasks and take decisions (i.e. replace

humans) on safety-related functions only, in

line with the requirements defined by

international standards.

[70]

Driving Car Tasks

Classification

Classification of the

tasks that take place

during the driving of the

vehicle and its modeling

from the perspective of

traditional control

engineering and artificial

intelligence

General

AI/ML

The major issues realted to safety and and

the efforts to make sure the technologies

involved are robust are discussed: test the

safety of the ADAS, standardization, the

development of models and algorithms, the

appropriate constructing solutions for

implementation, and ethical issues. No

specific findings are presented.

[61]

Lack of efficient

Safety Performance

Methodology to

generate an estimation of

General

AI/ML

Detailed methodology was proposed to

deal with the issue by means of statistical [65]

20

Verification

technique when

AI/ML is used

probability of fatal

mishap of an

autonomous UGV

navigation algorithm

based on Statistical

Testing in a MonteCarlo

manner in a Simulated

Environment

testing via simulation. Demonstration was

still a work in process.

Crash assignment,

especially between

automated vehicles

and nonautomated

vehicles

The integration of

three ethical theories—

utilitarianism, respect for

persons, and virtue

ethics—with vehicle

automation is used in a

simple crash scenario

where an automated

vehicle must choose

between three crash

types on the basis of a

randomly assigned

ethical theory to

understand the outcomes

of distinct ethical

frameworks

General

AI/ML

The results of the experiment indicated

that in specific crash scenarios, utilitarian

ethics may reduce the total number of

fatalities that result from automated vehicle

crashes, although other ethical systems may

be useful for developing rules used in

machine learning. The experiment

demonstrates that understanding rational

ethics is crucial for developing safe

automated vehicles.

[69]

Reduce the

uncertainty of a

driver behaviour

prediction model

Proposes a Data

Analysis Framework to

exploit AI, quantified

self, internet of things

and automated driving to

build a computational

driver behavioural

model aming to reduce

the uncertainty of a

driver behaviour

prediction model and be

used monitor, predict

and control a

transportation system.

General

AI/ML

It is very hard to predict due to the fluidity

and interactions of the driving factors

determining the driver performance.

[62]

How vehicle

autonomy

technology can be

used to benefit car

drivers and also to

propose a concept of

an autonomous

highway vehicle

which improves

highway driving

safety

Conceptual discussion General

AI/ML

No specific findings were presented by

the conceptual discussion [63]

AV decisions in

complex dilemmas

as a social agent

Proposition of a

framework for an ethics

policy for the artificial

intelligence of an AV

General

AI/ML

The ethics of automated vehicles is

becoming a major issue from legal, social,

and vehicle control perspectives. AV will

have to make decisions

that might eventually harm an agent and that

these decisions should not contradict the

interests of the end users or the principal

stakeholders. An ethics policy for

automated vehicles is needed and the

proposed framework (AVEthics) is only the

beginning of a long path.

[71]

21

Humans and

machines will often

need to work

together and agree on

common decisions.

Shared moral values

and ethical principles

General

AI/ML

Hybrid collective decision-making

systems will be in great need [72]

Autonomy assurance

and trust

(certification

process) in

Automated

Transportation

Systems

Framework

Proposition for the

discussion around the

topic

General

AI/ML

Authors explored some of the unique

challenges that autonomous transportation

systems present with regard to traditional

certification approaches such being non-

deterministic and employing adaptive

algorithms. Authors discussed the concept

of multiparty trust and how it can be

extended to a framework illustrating the

relationships between disparate roles. Two

thought experiments showed that building

and

maintaining trust in the perception and

judgment of increasingly autonomous

systems will be a challenge for the

transportation community.

[74]

Evaluate the impacts

of the number of

highly automated

vehicles on future

traffic safety and

traffic flow

Framework to

evaluate the impacts of

AV on traffic and the

impact of continuous

increase in the number of

highly automated

vehicles on future traffic

safety and traffic flow

General

AI/ML

The results of impact evaluation in this

study show that the increase in the

penetration rate of the highly automated

vehicle together with proper adjustment of

model parameter may result in considerate

improvements of safety in traffic in terms of

defined indicators. The developed

methodology allows to compare traffic

efficiency and safety measures with

different penetration rates in various

scenarios by means of microscopic traffic

simulation.

[68]

AV Test

Creation of Minimal

Test-Suites with Test-

cases for the validation

of AVs using recordings

of traffic situations

LTL

The process of test-case derivation can be

applied was demonstrated. According to

the authors, the derivation of test-cases

categorizes the recordings automatically

and allows test engineers to specify test

inputs. The test-case descriptions use the

Linear Temporal Logic (LTL) and allow the

execution of continuous behaviors, which

may also react to the behavior of the tested

vehicle. According to the authors, as the

traffic recordings can also be used for

machine learning algorithms, the

contributes to the discussion of their safety

certification. They also state the approach

is flexible as it can be extended if new traffic

situations are supposed to be covered by

testing.

[66]

DDT System Model Component – Human-in-the-loop

Collision avoidance

when no action is

taken by driver to

avoid the collision

Real time transition

from assisted driving to

automated driving under

conditions of high

probability of a collision

if no action is taken to

avoid the collision

Bayesian

Artificial

Intelligence

Systems can be designed to feature

collision warnings as well as automated

active safety capabilities. The high-level

architecture of the Bayesian transition

model seems promising. Example scenarios

illustrate the function of the real-time

transition model.

[57]

Design of driving

assistance system

Discussion about

Design considerations

are advanced in order to

Bayesian

Artificial

Intelligence

No specific findings, only discussions

about the important considerations to be

taken into account when designing AVs

[59]

22

overcome issues in in-

vehicle telematics

systems

Integration between

AV and human

driver

un-obstructive human

driver monitoring

approaches to ensure

they will be able to take

over control within short

notice

General

AI/ML

One of the most essential parts of the

autonomous driving system is to monitor

driver’s physical and mental state to avoid

unexpected traffic accidents. There are

some solutions for non-contact

measurement of vital signs, such as HR, RR

include laser Doppler, microwave Doppler

radar, and thermal imaging. Several AI

approaches that have been applied in

classifying non-contact physiological

sensor signals in different other domains

could be possible to investigate in

classifying driver’s signals.

This paper shows that the assessment of

non-contact physiological parameters

presents a greater challenge and few

attempts have been made to adopt it for the

driving situation.

[58]

Develop remote

controled car with

some automation to

deal with traffic light

detection, obstace

avoidance system

and lane detection

system to be driven

from anywhere over

a secured internet

connection

Traffic Light

Detection: Haar Feature

Based Cascade

Classifier; Lane

Detection: Canny Edge

detection and Hough line

transformation was used

Hough

Transformation

related

approaches

combined to

other

techniques

Low-cost remote-controlled car

prototype with basic automated functions

and using basic off-the-shelf components

with promising results. It proved cheap and

useful prototypes can be built for research.

Results with the proposed road detection

techniques proved to be highly efficient for

a road with clearly visible lane market.

Canny Edge detection proved to have low

error rate.

[38]

Selective Attention

Mechanism Weber–Fechner law

Regression

Based Models

Besides, the model is consistent with the

famous Weber–Fechner law. The Weber–

Fechner law says that all people’s feeling,

including visual feeling, auditory feeling

and so on all comply with the fact that the

feeling is not proportional to the strength of

the corresponding physical quantity but

proportional to the logarithm of the

corresponding physical quantity.

[56]

DDT System Model Component – Machine Control

Pre-crash problem of

AV Intelligent

Control

AV Intelligent

Adaptive Control

architecture using an

hybrid AI architecture:

CBR Engine for

Adaptive control (high

level) + hybrid Case-

Based Planner (A* and

D* motion planner)

Artificial

Neural

Network

combined to

other

techniques

Its is flexible to be integrated to lower

levels of vehicles controller and path

planners as (A* & D*). It is an ongoing

reasearch. Some limited and embrionary

experimental are mentioned and the authors

claim present research ideas for different

pre-crash scenarios and cases such as

intersection safety and some general cases.

The paper also discusses approaches to

integrate basic kinematics of AVs features

and presents future prospective on the

possibility of integration of high-level

intelligent controller with lower-levels

mechanical and kinematics features of

vehicles or robotics in general concepts.

[39]

Learning and

simulationf the

human-level

Use Neuroevolution

of Augmenting

Topologies (NEAT) and

Artificial

Neural

Network

Pilots' learning curve is irregular, due to

the characteristics of the problem, but

presents a good positive tendency which

[47]

23

decisions involved in

driving a racing car

a online videogame

prototype as a test-bed

environment. NEAT is a

combination of ANN

and Genetic Algorithm

(GA)

combined to

other

techniques

leads them to acquire fruitful abilities in just

50 generations with a population of 120

individuals. Pilots easily learn how to turn

following soft cruves, but still have big

poblems ientifying and steering hard ones,

which even make them crash into track

limits sometimes. The paper presents

individuals habing only 2 output neurons:

one for turning and the other one for

throttling/braking. Therefore, for this

reatively young ANNs, is very dificult to

acquire the high-level behaviours that have

to completely change the sign of the output

of the second neuron, whenevver a sharp

curve is near.

AV experiment (100

km)

AV following a

manually driven car

(trailing)

Fuzzy Logic

The authors state this paper is one of the

first communications fully describing the

control system and techniques required to

perform an experiment with autonomous

vehicles on open roads. It introduced a

different control approach for controlling

autonomous vehicles on urban and

motorway environments. A method for

online adjustment of the CACC fuzzy

controller is described and implemented,

coping with the most relevant disturbances

and uncertain parameters, such as road

slopes, passenger weight, or gear ratio. The

experiment successfully proves the

capability of the developed system to drive

more than one hundred kilometres

autonomously. A public demonstration of

the described system was conducted in June

2012, comprising a 100-km route through

urban and motorway environments. It was

able to cope with such gaps as motorway

overpasses, traffic signals, etc.The authors

stated the tracking results obtained with the

CACC system were very precise, with the

distance error being kept to less than 1

metre. Likewise, the lateral control was able

to maintain the vehicle on the path of the

leader with acceptable errors for both

scenarios. However, the localization system

needs to be improved to allow longer GPS

gaps. Also, the presence of a 900-metre-

long tunnel forced the deactivation of the

autonomous system while the vehicle

passed through.

[40]

Manage and control

low level vehicle

actuators (steering

throttle and brake)

Control schema to

manage low level

vehicle actuators

(steering throttle and

brake) based on fuzzy

logic

Fuzzy Logic

The proposed automatic low-level

control system has been defined,

implemented and tested in a Citroen C3

testbed vehicle, whose actuators have been

automated and can receive control signals

from an onboard computer where the soft

computing-based control system is running.

The preliminary results are confirming the

potential of the proposed technique.

[49]

24

Control intersection

crossing (all way

stops) and optimize it

Use a simulation to

model and simulate the

scenario. Developed a

heuristic optimization

algorithm for driverless

vehicles at unsignalized

intersections using a

multi-agent system.

Optimization

Heuristics

Although, the proposed research was still

in its initial stages, it presented some

significant time savings compared to an

AWSC intersection control. It showed that

by applying the proposed algorithm on only

four crossing vehicles, the total delay was

reduced by approximately 35 seconds,

which is equivalent to a 65 percent

reduction in the total intersection delay.

[42]

Autonomous

vehicles

intersections

crossing

optimization

Proposition of a new

tool for optimizing the

AVs movements through

intersections -

Cooperative Adaptive

Cruise Control (CACC)

Optimization

Heuristics

A simulation with one vehicle type and a

single intersection was performed. Also, all

vehicles in the simulation were assumed to

have CACC system to send/receive

information and follow speed advices as

directed. The preliminary results are

promissing and encourage future studies

where the authors plan to use simulations

with more types of vehicles and a greater

number of adjacent intersections.

According to the author, the results from

this research also warrant studies with

regard to incorporating non-CACC vehicles

into the system and studies pertaining to

tackling unexpected system changes,

pedestrian movements etc.

[46]

Path tracking cutting

corners using

traditional geometric

algorithms

A curve safety sub-

system for path tracking

based on the Pure Pursuit

algorithm and a dynamic

look-ahead distance

definition (based on

vehicle current speed

and lateral error). A sub-

system for path tracking

where an algorithm that

analyzes GPS

information offline

classifies high curvature

segments and estimates

the ideal speed for each

one.

Ramer-

Douglas-

Peucker or

Ramer-

Douglas

algorithm

Experimental results showed

improvements in comfort and safety due to

the extracted geometry information and

speed control, stabilizing the vehicle and

minimizing the lateral error

[48]

Safe-optimal

trajectory selection

for autonomous

vehicle

Use Big Data Mining

approach for crash

prediction and ETA.

Regression

Based Models

combined to

other

techniques

A Big Data based method and algorithm

has been presented to find the safest-optimal

trajectory for fully autonomous vehicles.

The method proposed relies strongly on the

results obtained from Big Data prediction

system which predicts accidents, ETA, and

clearance time. The algorithms for checking

and calculating the optimal trajectory are

very lightweight and straightforward. The

simulations using the available data are

promising.

[76]

How to prove the

correctness of an

algorithm for

Vehicle

Coordination

Use a distributed

coordination protocol

and an intersection

collision avoidance

(ICA) case study + Z3

Theorem prover +

Satisfiability Modulo

Theories (SMT) solver

Satisfiability

Modulo

Theories

(SMT) Solver

Paper presented a formalisation of the

distributed coordination problem

encountered by intelligent vehicles while

contending for the same physical resource.

It formalised a coordination protocol and an

intersection collision avoidance case study

in the SMT-lib language and proved system

safety using the Z3 theorem prover. The two

[43]

25

to prove correctness and

safety of a vehicular

coordination problem

main conclusions are: (1) The responsibility

approach to distributed coordination is a

suitable abstraction for formal reasoning on

system safety. The core of this approach is

that every entity is responsible for making

sure that it does not enter an unsafe state

with respect to any other entity. This can be

contrasted with the other approaches where

consensus is required between all nodes,

decisions are made by a central manager, or

where each pair of nodes negotiates

independently, all of which seem

problematic from a scalability point of

view; (2) The automatic verification of

collaborative vehicular applications with

the help of SMT solvers is at least plausible.

Some cases were found where the model

could not be verified and increasing the

detail and scale of the model would

certainly enlarge this problem. However,

there are certainly domain-specific

approximations that can be made to

alleviate some of these problems.

Robot maneuvers too

close to an obstacle

increases the

probability of an

accident. Preventing

this is crucial in

dynamic

environments, where

the obstacles, such as

other UAVs, are

moving

SVM-Inspired

Dynamic Safe

Navigation Using

Convex Hull

Construction an

algorithm for a fast

construction of a

maximum margin

between sets of obstacles

and its maintenance as

the input data are

dynamically altered

Support Vector

Machine

(SVM)

MMS-CH algorithm for calculating the

safest path in dynamic environment was

presented. It used the construction of

convex hulls over the input data to eliminate

data points irrelevant for the solution and to

use the boundary of the hulls to search for

the optimal separation margin between sets

of obstacles. The tests showed the algorithm

performs well in dynamic scenarios where

the input data might be altered by insertion

or deletion of data points. The

preprocessing phase of the MMS-CH

algorithm can recognize whether the change

in the data set does or does not require any

recalculation of the previous solution and

thus prevents unnecessary computations.

[41]

Drivers’ manoeuver

classification

Motion tracking (i.e

skeletal tracking) data

gathered from the driver

whilst driving to learn to

classify the manoeuvre

being performed

(Kinnect)

Support Vector

Machine

(SVM)

Preliminary results show that skeletal

tracking data can be used in a driving

scenario to classify maneuvers.

[44]

DDT System Model Component – Machine Perception

Recognize leading

vehicle in a convoy

Object detection using

Thermal infrared

classifiers and visible

light classifiers

Adaptive

Boosting

(AdaBoost)

Thermal infrared classifiers and visible

light classifiers were compared and

evaluated. Both approaches perform very

well. However, the accuracy of the visible

light classifiers cannot be reached by

thermal infra-red classifiers. But because of

the good performance of the thermal infra-

red classifier under all weather conditions,

the performance of the thermal infra-red

classifier is acceptable.

[22]

Prediction of

advanced driver

assistance systems

ML Classification

techniques

Artificial

Neural

Network

SVM shows best ML classification

performance (low errors and correlation

coefficient near 1) in prognosis of the

[55]

26

(ADAS) remaining

useful life (RUL) for

the prognosis of

ADAS’ safety

critical components

combined to

other

techniques

ADAS systems under the given

experimental assumption and Neural

Networks have the worst classification

performance. This work just proposes a

framework for a new area of research in

prognostics for automotive domain.

Road environmental

recognition and

various object

detection in real

driving conditions

Single monocular

camera for autonomous

vehicle in real driving

conditions

Artificial

Neural

Network

combined to

other

techniques

Pedestrian detection algorithm with GPU

were 6 times faster than CPU. Traffic sign

and traffc light recognition are 2 to 3 times

faster than pedestrian detection. However,

when the days are dark or there is backlight,

it was hard to separate the objects from

background.

[27]

Road Terrain

detection

Color feature

extraction + Clustering

algorithm k- mean +

ANN

Artificial

Neural

Network

combined to

other

techniques

Color Feature Extraction was used to

classify the Road Terrain with a Neural

Network (NN). 7666 images were used for

classification and results were promising.

[31]

Road junction

detection

3D point clouds

approach

Artificial

Neural

Network

combined to

other

techniques

The performance of ANNs, SVMs and

AdaBoost were compared for the second

step of the method, and of HMMs and CRFs

for the last. AdaBoost was considered the

best classifier, as it managed to learn the

training set without overfitting, generalizing

well to the test set. On a frame-by-frame

analysis, subsequent use of CRF and HMM

do not seem to improve from the results

obtained by AdaBoost itself. However, both

methods removed a lot of the classification

noise, generating an output that allows to

more clearly detect the start and end of a

road junction.

[30]

Vehicle Cyber

Attack Detection System

Artificial

Neural

Networks

Paper aimed to address the problem of

attack detection and identification when the

majority of multiple sensors was attacked in

an automotive CPS. LSTM and GRU

detected and identified attacks by

considering sequential information with real

data. It was demonstrated that the accuracy

of LSTM is the highest among data-based

methods (i.e., Neural Network, SVM,

simple RNN, GRU and LSTM). The

accuracy of LSTM followed the accuracy of

GRU. Especially, LSTM and GRU had

superior ability to detect coinstantaneous

attacks. LSTM and GRU showed high

performance in identification of Class 2, 3,

4, 5 and 6. Although calculation time of

GRU is faster than that of LSTM, it is no

matter to detect the attacks of the sensors on

a general computer to use a CPU.

[39]

Turn Signal

Recognition

Image Recognition

and Timming (95% and

82.2% accuracy)

Artificial

Neural

Networks

This paper proposed the flushing light

detection for preceding vehicles. The results

show that the obtained classifier detects turn

signals with an accuracy of 95(%).

Moreover, the proposed method is capable

of recognizing an appropriate frequency of

flushing light with an accuracy of 82.2(%)

for sequential driving data.

[18]

27

Road Sign

Classification in

Real-time

Use ANN (2 steps:

MLP + SLP)

Artificial

Neural

Networks

A novel approach based on two modules

was presented. The first module consists of

classifying the road sign's shape using MLP.

The shapes are classified in four classes:

triangular, inverted triangular, circular and

octagonal shapes. The accuracy of the MLP,

however, is improved when using only six

features values for increasing the speed of

the algorithm and minimizing the memory.

The second module is reserved to the

identification of the contents of recognized

shapes: the circular and triangular signs via

a simple SLP. As for the octagonal sign and

upside down triangular, they have a unique

indication which are the obligation to stop

and give way. A Performance Factor was

introduced in order to make a subjective

comparison between our proposed approach

and the other methods available in the

literature, which revealed that our proposed

system outperforms most of the other

approaches. Regarding running time, the

current software implementation takes

relatively a real time.

[32]

Spatio-temporal

situation awareness Deep Learning

Artificial

Neural

Networks

Given a driving video, the research aimed

to model which of the surrounding vehicles

are most important to the immediate driving

task. Employing human-centric annotations

allowed for gaining insights as to how

drivers perceive different on-road objects.

Although perception of surrounding agents

is influenced by previous experience and

driving style, we demonstrated a consistent

human-centric framework for importance

ranking. Extensive experiments showed a

wide range of spatio-temporal cues to be

essential when modeling object-level

importance. Furthermore, the importance

annotations proved useful when evaluating

vision algorithms designed for on-road

applications and autonomous driving.

[33]

Pedestrian detection

and movement

direction recognition

Deep Learning

Artificial

Neural

Networks

Paper presented a method to differentiate

the motion of pedestrians in real life

environments. Using a novel input-filtered

image based on the post-processing of static

recorded video frames, it could successfully

distinguish three different pedestrian

movement directions. Additionally, it has

been proved how CNNs can impressively

perform in such a task by training them with

a specialized dataset. Moreover, it has been

demonstrated how the results can be

enhanced even further by searching for the

best hyper-parameters once the CNN has

been fine-tuned for the specific problem, in

this case tuning the momentum and weight

decay CNN parameters. Paper also

presented an evaluation of the current state-

of-the-art CNNs, with ResNet being the

best-performing CNN for the image

[26]

28

recognition problem used (94% accuracy in

the validation set and 79% in the test set).

Pedestrian

Trajectory Prediction

Self-learning

Trajectory Prediction

Artificial

Neural

Networks

Results show that the LSTM prediction

model is superior to a constant velocity

Kalman Filter for pedestrian prediction

even on small datasets. Also, it was showed

that the prediction model can adapt to

changes in the pedestrian walking path

using only a small part of the new data. By

that, the size of the dataset can be kept rather

small although depicting the pedestrian’s

movement patterns

[28]

Early detection of

faults or malfunction

On-chip sensor

diagnosis

Artificial

Neural

Networks

The paper discusses the suitability and

feasibility of enhancing the reliability of

microsensor by adding an on chip self-

diagnosis capability. The approach used AI

techniques and sensors with no accessible

internal signals are taken as an example.

Some common acceleration sensor faults

are considered, and an indication is given of

the manner in which these faults can be

detected and isolated, either on an

individual sensor basis or based on

cooperative work within a sensor network.

The design requirements for such self-

diagnosable measument systems are set and

further practical implementation issues are

raised.

[55]

Securing connected

vehicles against

Denial of Service

(DoS) attack

Augment message

authentication with

Particle Filter (PF) and it

to Kalman Filter (KF).

Estimation

Filters (e.g.

Kalman Filter

and Particle

Filters)

Particle filter significantly reduces

communication overhead while keeping the

same detection level of spoofed messages

when compared to Kalman filter in VANET

applications. Stimulating different scenarios

with Context adaptive beacon verification

along with Kalman and particle filter on

University of Massachusetts Dartmouth and

State Road (Dartmouth) proved that it can

detect and prevent spoofed attacks and help

reducing the computational overhead. But,

the Current method of securing the

connected vehicle with filters leave the

burden of privacy protection on VANET.

The practice makes the autonomous cars the

target of attack because of the number of

spoofed messages missed by context

adaptive beacon verification is around 11%

(41 out if 46 were detected) which leaves

the undetected rate too high to be replaced

by conventional verification method. KF is

good when road was linear and lags when

the path is non-linear. PF is good for both.

KF saves upto 86.5% while Particle Filter

can save 85.94% computational overhead

for the same scenario. Detect around 76%

(24% missed) spoofed beacons with

Kalman Filter and 89% (11% missed)

spoofed beacons with Particle Filter.

[51]

Reliable and robust

obstacles detection

An innovative multi-

dimensional structure

based on association

Estimation

Filters (e.g.

Kalman Filter

The presented system was able to track

and fuse obstacles coming from a laser and

a stereo camera. The approach has been

[25]

29

costs originating from a

classifier provides an

optimal solution to the

association problem with

respect to the total

association cost.

and Particle

Filters)

compared with other state of the art

algorithms, showing better results in all the

considered metrics. Moreover, the system

uses less computational resources and thus,

fixed the platform, may work at higher

frame rate compared to other solutions,

making it more appropriate for automotive

applications. The system has demonstrated

a correct reconstruction of the dynamic

world surrounding the vehicle, proving to

be able to help the driver in the assessment

of critical situations. In particular, the

developed algorithm provides a stable,

robust and reliable detection, classification

and tracking of the multiple targets coming

from different sources. Moreover, the

proposed approaches were seen to

outperform the state-of-the-art approaches

on a public dataset.

A fault tolerant and reliable system requires

sensors redundancy and complementarity.

Common approaches rely on object level

fusion. It has been introduced a medium-

level fusion which take advantage from both

the approaches. The fusion is performed at

object level but preserving the low-level

information; in this way it is guaranteed a

real-time processing exploiting all available

information.

Cyber Attack in V2X Fuzzy Detector Fuzzy Logic

To address security issues of a system of

connected vehicles (CVs), a fuzzy detector

was also introduced that detects possibility

of a cyber threat and takes proper actions in

response to the specific attack. Results show

the designed system can detect any

adversary access to the system and can

prevent subsequent crashes by adjusting the

safe following distance.

[54]

Road and Obstacle

Detection Sensor Fusion Fuzzy Logic

A high-level data-fusion strategy has

been devised, which is based on the

identification and representation of the

descriptive and procedural knowledge

required. Such a strategy yields better

recognition results by merging the various

hypotheses generated by the separate

channels and solving possible conflicts

through a fuzzy representation of

knowledge when compared to a benchmark.

In addition, the data-fusion system has

performed a more accurate segmentation

process by using goal-driven low-level

procedures, according to which the image

regions have been assigned relative fuzzy

memberships to the object to be recognized.

[24]

Enhance Image

Understanding

Develop generic

technology that will

enable the construction

of complete, robust, high

performance image

understanding systems

General

AI/ML

This paper provided an overview of the

technical and program management plans

being used in evolving the proposed

technology, but no results were presented.

[19]

30

to support a wide range

of DoD applications

Learn from

Demonstration

Use a Semi-automated

mine robot

Hidden

Markov Based

Models (e.g.

Continuous

Hidden

Markov

Model-CHMM

and Discrete

Hidden

Markov

Model-

DHMM)

combined to

other

techniques

In this paper, three methods were

compared based on three trajectories in the

low noise and noisy environments. The

GMM based method had the best

performance in a low noise environment. In

practice, there’s always unexpected noise

around a robot, implying the GMM based

method was not practical for real

environments. The CHMM based method

was suitable for turning trajectories, while

The DHMM based method was more robust

for straight trajectories.

[45]

Vehicle Detection

and Counting

Hidden Markov

Model + Viterbi

algorithm + Adaboost

trained Haar-like feature

detector Approach

Hidden

Markov Based

Models (e.g.

Continuous

Hidden

Markov

Model-CHMM

and Discrete

Hidden

Markov

Model-

DHMM)

combined to

other

techniques

The proposed method has been shown to

give significantly better vehicle volume

counts than both multiple targets moving

object tracking and VDL on a dataset of

over 88 hours of video. On this testing set,

the proposed method achieved a median 5-

minute-bin error of 0.0686 for this counting

task while the multiple target motion

tracking and VDL implementations had

median errors of 0.0957 and 0.2290

respectively. The proposed method was also

more reliable having fewer and less severe

occurrences of 5-minute-bin errors

throughout the testing set.

[36]

Safety-warning and

driver-assistance

system and an

automatic pilot for

rural and urban

traffic environments

Adaptive randomized

HT (RHT) for robust and

accurate detection of

lane markings without

manual initialization or a

priori information under

different road

environments

Hough

Transformation

In this paper, a prototype was

demonstrated, and tasks of lane detection

detailed. Preliminary experimental results

in different road scene and a comparison

with other methods have proven the validity

of the proposed method

[21]

Speed bump

detection

Use Camera and

image recognition

Novel Image

Recognition

Technique

The average performance of the system

considering only speed bump with proper

marking is 85%.

[34]

Obstacle clusteering

and tracking

Lidar + split-and-

merge algorithm

Ramer-

Douglas-

Peucker or

Ramer-

Douglas

algorithm

Paper presents a robust platform for

implementing a perception system for

ground vehicles using a LIDAR sensor and

two cameras has been designed. Tests were

performed using this platform and different

implementations, and the results were

checked with the real-world scenes,

demonstrating the technique validity.

[23]

Predicts the CPU

usage patterns of

software tasks

running on a self-

driving car

Methods for learning

the patterns of tasks’

CPU utilizations in given

driving contexts

Regression

Based Models

A feature vector was designed to

represent the internal and external states of

a self-driving car and five regression

methods were used to predict the CPU

usage patterns of four software tasks

running on a self-driving car. Through

testing with the actual driving data, the

results showed a regression method could

be used to predict a software task’s dynamic

[53]

31

CPU utilization.

Providing road safety

to connected drivers

and connected

autonomous vehicles

Observing the

Doppler profile

Support Vector

Machine

(SVM)

The paper presented a collision and

driving scenario classification system based

on the Doppler profile which could

potentially decouple the safety benefits of

V2V communications from relying on SM

content. The Doppler profile in V2V

networks showed rich data about the

vehicles and their environments and could

be exploited to potentially provide a reliable

collision avoidance service directly from

the radio front-end. No experimental results

were presented.

[20]

Road Detection

Road area detection

method using a support

vector machine (SVM)

and histogram of

oriented gradient (HOG)

features and 3D lidar

Support Vector

Machine

(SVM)

combined to

other

techniques

Classifier to differentiate road areas from

other areas using a 3D Lidar with machine

larning techniques. Range data of lidar

changes in the layer direction but not in the

rotational direction. HOG features of the

reaos concentrate in the same bin. In

contract the features of the non-road areas

are distributed among several bins

representing various directions. Found

differences between the histograms for the

roas planes and the other areas. In real world

data, same tendences of HO features. Error

rate of 8.51% using SVM. Area up to 10m

ahead of the vehicle can be identified

correctly.

[75]

Traffic Light

Detection

Two-stage

preprocessing using

Principal Component

Analysis (PCA)

followed by SVM

Support Vector

Machine

(SVM)

combined to

other

techniques

Paper presents a system that can detect

multiple types of green and red traffic lights

accurately and reliably. Color extraction

and blob detection were applied to locate the

candidates with proper optimization. A

classification and validation method using

PCANet was then used for frame-by-frame

detection. Multiobject tracking method and

forecasting technique were succesfully

employed to improve accuracy and

produced stable results.

[35]

Pedestrian Detection

Use High-Definition

3D Range Data (from a

LIDAR)

Support Vector

Machine

(SVM)

combined to

other

techniques

An exhaustive analysis of the

performance of three different machine

learning algorithms have been carried out:

k-Nearest Neighbours (kNN), Naïve Bayes

classifier (NBC), and Support Vector

Machine (SVM). Each algorithm was

trained with a training set comprising tool

277 pedestrians and 1654 no pedestrian

samples and different kernel functions:

kNN with Euclidean and Mahalanobis

distances, NBC with Gauss and KSF

functions and SVM with linear and

quadratic functions. LOOCV and ROC

analysis were used to detect the best

algorithm to be used for pedestrian

detection. The proposed algorithm has been

tested in real traffic scenarios with 16

samples of pedestrians and 469 samples of

non-pedestrians. The results obtained were

used to validate theoretical results. An

overfitting problem in the SVM with

[27]

32

quadratic kernel was found. Finally, SVM

with linear function was selected since it

offered the best results. A comparison of the

proposed method with five other works that

also use High Definition 3D LIDAR to

carry-out the pedestrian detection,

comparing the AUC and Fscore metrics.

Conclusions are the proposed method

obtains better performance results in every

case. Pedestrian detection has traditionally

been performed using machine vision and

cameras, but these techniques are affected

by changing lighting conditions. 3D LIDAR

technology provides more accurate data

(more than 1 million points per revolution),

which can be successfully used to detect

pedestrians in any kind of lighting

conditions

How to "automate"

manual annotation

for images to train

visual perception for

AVs

Training visual

models using photo-

realistic computer

graphics

Support Vector

Machine

(SVM)

combined to

other

techniques

The experiments showed that virtual-

world data is effective for training vision-

based pedestrian detectors which can be

adapted to operate in real scenarios. The

different adaptation procedures have shown

to provide adapted detectors that improve

those trained only on virtual data, as well as

those trained using only the real-world data

available for the adaptation (which

constrained to save a ∼90% annotation effort along the experiments).

[37]