Date of Award
Doctor of Philosophy (PhD)
Dr. Wayne Sarasua, Committee Chair
Dr. Ronnie Chowdhury, Committee Member
Dr. Jennifer Ogle, Committee Member
Dr. Joshua Levine, Committee Member
In transportation engineering, sufficient, reliable, and diverse traffic data is necessary for effective planning, operations, research, and professional practice. Using aerial imagery to achieve traffic surveillance and collect traffic data is one of the feasible ways that is facilitated by the advances of technologies in many related areas. A great deal of aerial imagery datasets are currently available and more datasets are collected every day for various applications. It will be beneficial to make full and efficient use of the attribute rich imagery as a resource for valid and useful traffic data for many applications in transportation research and practice. In this dissertation, a traffic surveillance system that can collect valid and useful traffic data using quality-limited aerial imagery datasets with diverse characteristics is developed. Two novel approaches, which can achieve robust and accurate performance, are proposed and implemented for this system. The first one is a computer vision-based approach, which uses convolutional neural network (CNN) to detect vehicles in aerial imagery and uses features to track those detections. This approach is capable of detecting and tracking vehicles in the aerial imagery datasets with a very limited quality. Experimental results indicate the performance of this approach is very promising and it can achieve accurate measurements for macroscopic traffic data and is also potential for reliable microscopic traffic data. The second approach is a multiple hypothesis tracking (MHT) approach with innovative kinematics and appearance models (KAM). The implemented MHT module is designed to cooperate with the CNN module in order to extend and improve the vehicle tracking system. Experiments are designed based on a meticulously established synthetic vehicle detection datasets, originally induced scale-agonistic property of MHT, and comprehensively identified metrics for performance evaluation. The experimental results not only indicate that the performance of this approach can be very promising, but also provide solutions for some long-standing problems and reveal the impacts of frame rate, detection noise, and traffic configurations as well as the effects of vehicle appearance information on the performance. The experimental results of both approaches prove the feasibility of traffic surveillance and data collection by detecting and tracking vehicles in aerial video, and indicate the direction of further research as well as solutions to achieve satisfactory performance with existing aerial imagery datasets that have very limited quality and frame rates. This traffic surveillance system has the potential to be transformational in how large area traffic data is collected in the future. Such a system will be capable of achieving wide area traffic surveillance and extracting valid and useful traffic data from wide area aerial video captured with a single platform
Zhao, Xi, "Traffic Surveillance and Automated Data Extraction from Aerial Video Using Computer Vision, Artificial Intelligence, and Probabilistic Approaches" (2016). All Dissertations. 1811.