A Fast Probabilistic Method for Vehicle Detection and Tracking with an Explicit Contour Model

This dissertation, "A Fast Probabilistic Method for Vehicle Detection and Tracking With an Explicit Contour Model" by Wai-sing, Boris, Yiu, 姚維勝, was obtained from The University of Hong Kong (Pokfulam, Hong Kong) and is being sold pursuant to Creative Commons: Attribution 3.0 Hong Kong License. The content of this dissertation has not been altered in any way. We have altered the formatting in order to facilitate the ease of printing and reading of the dissertation. All rights not granted by the above license are retained by the author. Abstract: Abstract of thesis entitled AFastProbabilisticMethodforVehicleDetectionandTracking withanExplicitContourModel submitted by YIUWaiSing, Boris for the degree of Master of Philosophy at the University of Hong Kong in August 2005 Autonomous traffic surveillance systems have to detect and track every vehicle within their fields of view in real time. One major challenge for such systems is to handle different views of various classes of vehicles. This is not a trivial problem, as the appearances of dif- ferent vehicles may look substantially different in an image, and simply stating presence is not sufficient. Such systems require accurate vehicle boundaries in the images for construct- ing an informative, distinctive and concise model for a vehicle, and for providing a strong cue for locating the vehicle during visual tracking. The systems should, therefore, detect and recognise the boundary of each vehicle. Existing methods either make restrictive assumptions on the scene or focus on a par- ticular class of vehicles in a manner that cannot be extended to a wide variety of vehicles without consuming too much computing power. In this thesis, a generic '2.5D' shape model is introduced to describe the general shapes of vehicles, and used for detecting and tracking various classes of vehicles efficiently. The proposed model encompasses most approximated 2D projections of a 3D cuboid in variable dimensions. Based on such a generic model, a probabilistic template fitting framework is developed to determine the best contour of a po- tential vehicle in images using an efficient dynamic programming algorithm. This allows traffic surveillance systems to detect various classes of vehicles efficiently. A track is then initiated for every detected vehicle automatically. To handle the image shape deformation of the vehicle during tracking, the generic 2.5D shape model is further extended to explicitly describe how the image shape of a target vehicle would vary with its motion. Such an enhanced shape model consists of a concise set of parameters that offers a reasonable level of detail for real-time surveillance applications. This low dimensional model can be integrated into any Bayesian tracking framework, and the resultant tracking algorithm can predict the next state of a target more accurately and thus make more relevant measurements from the images. The shape model also relates the shape to the velocity of the target to improve robustness by monitoring any state inconsistency. The method was evaluated on various traffic videos, and it worked effectively in realistic conditions, including curved and inclined roads as well as straight roads, and achieved good accuracy and fast performance. The approach of explicitly modelling shape deformation in a low-dimensional model was shown to be effective in a real-time surveillance system. DOI: 10.5353/th_b3505717 Subjects: Computer vision Image processing Automobiles - Tracking Bayesian statistical decision theory Algorithms