4 pages total due Monday evening PST

Proposal Draft

Summary

Detecting real-time traffic congestion is the most crucial step in Intelligent Transportation System (ITS). Numerous congestion detection technologies, such as inductive loop detectors, traffic radar sensors, and video surveillance cameras, have been proposed and demonstrated acceptable performances. However, due to the inflexibility and high maintenance cost required for deploying and operating inductive loop detectors and traffic radar sensors, video-based traffic sensing technique has been gaining its popularity in detecting traffic congestion. Congestion detection function embedded in the video-based system relies on video analytics (VA). VA is a video image processing technique which captures the changes of individual image pixels. Since the quality of video image is affected by environmental factors such as weather (e.g., snow, rain), and light condition (e.g., sun glare, car headlight during the night time), VA often produces undesirable accuracy for the congestion detection, thereby resulting in numerous false alarms for the operators in the traffic management centers (TMCs). Therefore, in practice, the TMC operators often turn off the congestion detection function and conduct manual detection by watching video footages all the time, which is inefficient and time consuming particularly for the large number of video cameras.

Introduction

The significance of my research lies in improving the state-of-the-practice of real-time traffic congestion detection without any extra cost to build infrastructure. The research will develop and evaluate an innovative VA-free real-time traffic congestion detection method. Powered by Artificial Intelligence integrated with Machine Learning techniques, the proposed method will conduct human-like traffic congestion detection in an automatic and robust manner by bypassing the implementation of video image processing. The proposed method will enable to cover the large number of video footages in real-time. Thus, it will dramatically reduce the time and effort for the TMC operators to detect traffic congestion and improve the accuracy of the detected congestion.

Project Description

Thanks to the accelerated development of computing power and machine learning algorithm, training a large data set in a comparatively short amount of time was made possible. In this research we pursue an autonomous traffic congestion detection method, more specifically, detection of the sudden speed drop. We approach this problem by using machine learning techniques, looking for the patterns of traffic congestion information from historical unlabeled video footage data, then reshape them and feed them into a constructed neuro-network. The model used will be learned offline by using Google TensorFlow™. The research difficulties arise when attempting to do this in real-time under changing illumination and weather conditions as well as heavy traffic congestions. We will propose a novel approach to use a semi-supervised clustering technique. In detail, our plan could be separated into multiple tasks as follows:

· Task 1: Construct a TensorFlow-based machine learning model by using historical traffic congestion information

· Task 2: Conduct machine training process and validate training results with testing data

· Task 3: Perform Proof-of-Concept (POC) tests to evaluate the performance of the proposed congestion detection method.

Plan of Work

I will develop a software package to conduct the real-time congestion detection as one of the final deliverables. The software package will include 1) the program source code of the congestion detection algorithm, 2) trained neuro-networks, and 3) a graphical user interface (GUI)-based computer software to provide potential users (e.g., TMC operators) with intuitive operation of the proposed method. In addition, a final report summarizing the POC test results will be added to the final deliverables. With the final product from our research, it is expected that TMC operators adopt our product to improve the efficiency and accuracy of real-time traffic congestion detection.

Intended Impact

Government’s agencies (e.g., state and city department of transportation) responsible for operating TMCs would be primary stakeholders for the products of THE research. In addition, transportation data providers in private sectors who feed real-time traffic data to TMCs, such as INRIX, HERE, TRANSCOM, will be potential customers of our research products. Once adopted in the market, our research product will affect the improvement of the efficiency and accuracy of real-time traffic congestion detection.

Conclusion

In this research, I attempt to address the problem of real-time traffic congestion detection. By using a direct perception approach, mapping a single image directing to a congestion indicator, I design a system which can detect congestion independent of location, time and weather. I demonstrate that the use of FFT and WT with a convolutional neural network can produce high accuracy across multiple conditions in new locations. These results are promising but still exploratory. Future research steps in this area include 8 the creation of a larger dataset and training for larger networks. This set should include tens of thousands to hundreds of thousands of images from 50 or more locations. This would allow for more thorough testing. Additionally, a convolution neural network with more layers may improve results. This model would require more data and computational power then was available for this research. Considering the success of such models in other domains and the promising results in this study, the performance of a larger convolution neural network would be interesting to investigate.

Budget

Intelligent Transportation Systems (ITS) Laboratory in the Department of Civil and Environmental Engineering will provide the research team with primary resources. Thus, the overall budget should not exceed $20,000.

Professor’s comments:

Very good start. 

The most important part of a proposal is the summary which needs some work. Most of what you have should be in an introduction and some of what is in our introduction shold be in the summary.  Please check in the text for what should be in a summary (a brief discussion of what you are going to do, not why or the need for it.)

Put in a more detailed budget as if it were the real thing. Make up some costs for equipment or what ever you might need.  Pay yourself and an aide to keep track of data. Add a timeline section in which you list what you will do each week or month of the summer. 

Be sure to give credit to any sites you took information from.

Add  a cover page and a table of contents for the final proposal.