2
ATR for UAS Surveillance
Introduction
Unmanned aerial systems (UAS) have become universal for both military and civilian
applications including surveillance, reconnaissance, target acquisition, and aerial mapping.
Equipping UAS platforms with automated target recognition (ATR) capabilities can greatly
enhance their utility by enabling intelligent autonomous detection, classification, and tracking of
objects of interest during flight missions. However, developing robust and accurate ATR systems
poses substantial algorithmic, hardware, and systems integration challenges. This paper provides
a comprehensive survey and assessment of the state-of-the-art in ATR technologies and their
integration with UAS to perform automated surveillance and target acquisition functions.
ATR relies on advanced computer vision and machine learning algorithms applied to
imagery and sensor data collected by onboard sensor systems. Research and development of
ATR capabilities has been an active domain for over 30 years, initially driven by military
applications for automated threat detection and identification. More recently, civilian UAS
applications including search and rescue, wildfire monitoring, and border security have also
spurred advancement of ATR techniques. However, real-time ATR from aerial platforms entails
significant challenges including processing bandwidth limitations, power constraints, integration
with navigation and control systems, and optimizing tradeoffs between competing performance
metrics.
This paper will present an in-depth review of key aspects of ATR systems including
algorithms, sensors, architectures, and integration approaches. Both template-based and machine
learning ATR algorithms will be examined, along with core sensor modalities such as electro-
optical, infrared, and radar imaging. ATR processing requirements will be analyzed considering
factors like onboard versus offboard architectures, power and size limitations, and real-time
3
ATR for UAS Surveillance
performance needs. Additionally, techniques for integrating ATR with UAS navigation and
flight control to enable autonomous target tracking and risk terrain avoidance will be discussed.
Finally, ATR system performance benchmarks and tradeoffs including detection accuracy, false
alarm rates, and processing latency will be evaluated based on field testing results and
operational datasets.
This paper aims to provide a technical survey and comparative assessment of the
capabilities and limitations of modern ATR technologies and their integration with UAS
platforms for intelligent surveillance applications. Both existing ATR capabilities and directions
for future enhancement will be examined.
Research
ATR Algorithms
Machine learning algorithms such as deep Convolutional Neural Networks (CNNs),
region based CNNs, and Graph Convolutional Networks (GCNs) have become the dominant
approach for aerial image classification and Automated Target Recognition (ATR) (Gao et al.,
2020; Zhang et al., 2019). GCNs in particular have achieved state-of-the-art results by
incorporating spatial context and feature extraction layers tailored to hyperspectral imagery. Pre-
training on large labeled datasets like ImageNet improves generalization, while fine-tuning on
domain-specific data boosts detection accuracy. Despite success on benchmarks, real-world
performance lags due to diversity of target types, orientations, backgrounds, and imaging
conditions.
Hybrid approaches combine machine learning with template matching and heuristic rules
to improve robustness. For example, infrared small target detection algorithms utilize target-
background collaboration and competition schemes to isolate objects based on thermal contrast
4
ATR for UAS Surveillance
(Wang & Wang, 2020). Dual-domain features from visible and infrared spectra improve
classification over single modalities. Physics-based modeling and segmentation preprocessing
help focus the search space for machine learning stages.
ATR Sensors
Electro-optical sensors provide high-resolution visible spectrum imagery for identifying
targets based on shape, texture, color, and contextual clues (Zhang et al., 2019). However,
performance suffers under low light and foggy conditions. Infrared sensors detect thermal
signatures of targets, improving visibility at night and in dusty environments, but have lower
image resolution (Wang & Wang, 2020). Radar provides all-weather capability and direct motion
tracking by measuring target velocity and direction. Yet radar has limited ability to classify
target types and can suffer from clutter.
Multispectral and hyperspectral sensors collect information across hundreds of narrow
bands, enabling detailed spectral profiles for identification. However, they require sophisticated
processing of massive datasets. LIDAR and stereo cameras add 3D structural cues but have
lower range than radar. Fusing data from multiple complementary modalities can overcome
limitations of individual sensors (Ariante et al., 2022). This improves classification accuracy but
requires additional processing and data integration complexity.
ATR Processing Requirements
Onboard processing provides fast and reliable ATR without wireless throughput
bottlenecks, but faces size, weight, power, and computing constraints, especially on small UAS
platforms (Politi et al., 2024). Specialized AI acceleration chips like GPUs and TPUs can enable
advanced neural networks within tight resource limits. Optimized inference engines leverage
quantization, pruning, efficient architectures like MobileNets, and model compression to reduce
5
ATR for UAS Surveillance
computation and memory needs. Offboard processing enables virtually unlimited algorithms at
the cost of latency, bandwidth, and dependence on the communication link. A hybrid approach
balances onboard detection with offboard classification.
Real-time ATR requires not only high accuracy but also low latency below 100ms. This
is critical for closed-loop integration with UAS flight controls and navigation to enable
capabilities like obstacle avoidance and dynamic target tracking (Ariante et al., 2022). High
reliability and determinism are also important for safety-critical applications. Redundant sensing
combined with health monitoring improves robustness.
ATR Integration with UAS
Tight integration of ATR subsystems with navigation, control, and planning components
allows UAS to autonomously adapt trajectories based on detected threats and prioritize targets
(Politi et al., 2024). For example, integrating radar target tracking and electro-optical
classification can enable real-time path planning around dangerous weather while visually
identifying the highest priority ground targets. Matching field of view with vehicle dynamics
optimizes vantage points during maneuvers. An augmented reality interface provides intuitive
situation awareness to human operators, highlighting recognized targets, uncertainties, and flight
path predictions.
Advances in sense-and-avoid technologies are critical to enable routine beyond visual
line of sight (BVLOS) operations with ATR-equipped UAS. Fusion of computer vision, lidar,
radar, and acoustic sensors provides comprehensive detection of aircraft, terrain, and obstacles
(Ariante et al., 2022). A ground control system with rotating lidar can even supply backup
navigation data to land autonomously without GPS.
Evaluation of ATR Systems
6
ATR for UAS Surveillance
Standard datasets like VOT2020, DOTA, xView, VisDrone, and CARPK quantify ATR
performance in terms of detection accuracy, false alarm rates, and processing latency. State-of-
the-art deep neural networks achieve over 90% detection on favorable benchmarks but still
struggle with complex real-world data. False alarms must be minimized for safety-critical UAS
applications as incorrect classifications can lead to catastrophic autonomous behavior. End-to-
end system latency is measured to ensure real-time constraints are met for closed-loop control.
Ultimately, progress relies heavily on field testing and operational experience across
diverse conditions. Larger, more realistic labeled datasets will expand the domains covered.
Continued growth in onboard processing power and algorithms will enable ATR systems to
match the capabilities of human pilots. Reliable all-weather, all-domain automated target
recognition remains an open grand challenge.
Conclusion
This paper has provided a comprehensive overview of the current state of automated
target recognition (ATR) technologies and their integration with unmanned aerial systems (UAS)
for intelligent surveillance applications. ATR has seen significant advancement driven by
machine learning approaches applied to imagery from sensors across multiple modalities
including electro-optical, infrared, and radar. Tight integration of ATR subsystems with UAS
navigation and control enables autonomous capabilities like dynamic target tracking and
trajectory adaptation.
However, real-world performance continues to lag benchmarks due to the diversity and
complexity of operational environments. Key remaining challenges include improving
generalization across target types and imaging conditions, reducing false alarms, meeting
stringent real-time latency requirements, and enhancing reliability. More sophisticated
7
ATR for UAS Surveillance
algorithms, larger datasets, sensor fusion, and extensive field testing will be critical to realize the
full potential of ATR to provide autonomous all-weather, all-domain surveillance from UAS
platforms.
Safety and reliability will need to be rigorously validated before full autonomy can be
supported, especially for beyond visual line of sight operations in civilian airspace. But the rapid
pace of progress in artificial intelligence, sensors, and autonomous systems points to a future
where ATR-enabled UAS can take on an increasing share of dull, dirty, and dangerous
surveillance missions to aid human operators. With continued research and maturation, ATR
promises to be a transformative technology enabling a new era of intelligent remote sensing
across a wide range of commercial and military applications.
8
ATR for UAS Surveillance
Reference
Ariante, G., Ponte, S., Papa, U., Greco, A., & Del Core, G. (2022). Ground control system for
UAS safe landing area determination (SLAD) in urban air mobility operations.GSensors
(Basel, Switzerland),G22(9), 3226.Ghttps://doi.org/10.3390/s22093226
Gao, J., Ma, Y., Zhang, A., & Plaza, A. (2020). Graph convolutional networks for hyperspectral
image classification. IEEE Transactions on Geoscience and Remote Sensing, 59(7),
5966-5978. https://doi.org/10.1109/TGRS.2020.3041316
Politi, E., Purucker, P., Larsen, M., Reis, R. J. D., Rajan, R. T., Penna, S. D., Boer, J.,
Rodosthenous, P., Dimitrakopoulos, G., Varlamis, I., & Höß, A. (2024). Enabling
technologies for the navigation and communication of UAS operating in the context of
BVLOS.GElectronics (Basel),G13(2), 340.Ghttps://doi.org/10.3390/electronics13020340
Wang, C., & Wang, Z. J. (2020). Infrared small target detection based on target-background
collaboration and competition. Infrared Physics & Technology, 106, 103213.
https://doi.org/10.1016/j.infrared.2020.103213
Zhang, Y., Liu, Q., & Wang, X. (2019). Road extraction from high-resolution remote sensing
imagery using deep learning. Remote Sensing, 11(13), 1580.
https://doi.org/10.3390/rs11131580
Powered by TCPDF (www.tcpdf.org)
