An effective methodology for visual traffic surveillance

by Lai, Hon-seng

Abstract (Summary)
(Uncorrected OCR) Abstract of thesis entitled An Effective Methodology for Visual Traffic Surveillance submitted by LAI Hon Seng for the degree of Doctor of Philosophy at the University of Hong Kong in January 2000 This thesis presents a methodology for Automatic Visual Traffic Surveillance. The methodology has been developed upon the concept of Unconstrained Detection Regions, which enables it to perform multiple vehicle detection, tracking and traffic parameter estimation in real time. The goal of this research was to develop a suitable methodology for automatic visual traffic surveillance to perform multiple vehicle detection, tracking and traffic parameter estimation in real time as well as being able to tackle scene changes automatically, and detect and handle vehicle occlusion. The proposed methodology is able to first, tackle unpredictable outdoor environment changes, noises, camera actions and vibration automatically based on several novel algorithms. Second, detect vehicle occlusion by monitoring the changes in vehicle model dimensions and handle occlusion using a novel algorithm, so that vehicle can be tracked individually. Third, estimate traffic parameters of individual vehicles or road conditions. Fourth, achieve real time response by employing efficient algorithm to reduce the computation requirement. It has been tested on a number of freeway surveillance image sequences and has shown to be successful in performing the above correctly. In essence, the methodology employs a modular architecture, which consists of four modules: (1) feature preserving noise filtering; (2) vehicle extraction; (3) vehicle tracking; and (4) traffic information estimation, each of which was studied and novel algorithms were proposed to tackle their respective problems. Specifically, to remove noise in digital images and to minimize the feature degradation, a feature preserving noise filtering method has been developed. In it, pixels are classified into either corrupted or uncorrupted and only those corrupted pixels are filtered. Our evaluation has shown that the new method can remove noise effectively, preserve image features, has faster processing speed and can be applied iteratively. In vehicle extraction, a background subtraction approach has been considered, from which, vehicles are extracted by subtracting a stationary background from the image sequence. A scoreboard-based background estimation method has been developed to record the variation of each pixel for selecting different estimation algorithms. Our evaluation has shown that it is fast and accurate. From the background, road parameters are extracted automatically using an edge-based approach in conjunction with two edge discrimination heuristics to determine the road lanes, centerline and lane direction. Our evaluation indicated that it is fast, automatic and can work with straight roads with multiple lanes. Complex vehicle occlusion that appears in visual surveillance is tackled by a generalized deformable modeling method and a vehicle occlusion detection and handling method. In principle, a vehicle is fitted by a 2D projection of a 3D cuboid wire frame with parameterized vertices. By monitoring the changes in the model dimensions and its Area ratio, occlusion is detected successfully every time, and the trajectory is deduced. When occlusion occurs, vehicles are tracked separately by replicating the vehicle trajectory or splitting the occluded vehicle model. Trials on real-world image sequences have proven the effectiveness of these methods. Traffic parameters can be estimated from the trajectories including count, speed, traffic saturation and speed ratio. Potentially, other traffic parameters may also be obtained from the trajectories including lane change frequency and congestion index. The research presented in this thesis has contributed to a number of aspects in automatic visual traffic surveillance research. First, robustness in extraction and detection can be achieved by using background estimation and subtraction, appropriate extraction of vehicles, road parameters estimation and simplified deformable modeling of the vehicles. Second, accuracy in parameter estimation can be achieved by using feature preserving noise filtering, accurate background estimation, modeling and speed estimation. Third, computation speed can be reduced by using fast algorithm for filtering, background estimation and modeling. Fourth, occlusion can be detected and managed by deformable modeling, which presents a simple and effective way for dealing with this frequently occurred scenario.
Bibliographical Information:


School:The University of Hong Kong

School Location:China - Hong Kong SAR

Source Type:Master's Thesis

Keywords:traffic engineering data processing


Date of Publication:01/01/2000

© 2009 All Rights Reserved.