A Methodology for Resolving Multiple Vehicle Occlusion in Visual Traffic Surveillance

This dissertation, "A Methodology for Resolving Multiple Vehicle Occlusion in Visual Traffic Surveillance" by Chun-cheong, Pang, 彭俊昌, 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 A Methodology for Resolving Multiple Vehicle Occlusion in Visual Traffic Surveillance submitted by PANG Chun Cheong for the Degree of Doctor of Philosophy at The University of Hong Kong in September 2005 This thesis proposes a novel methodology for resolving N vehicle occlusion in visual traffic surveillance. The methodology has been developed upon the philosophy that a vehicle can be represented by a generalized deformable model (GDM) which is a rectangular model that can fit onto any type of vehicle. By using the GDM as a building block, three models are proposed to describe the projection behavior of a single GDM, namely, the Vertex-Shape Model (VSM) which models the relationship between the camera viewing position and the number of projected vertices of the GDM; the Contour Description Model (CDM) which models the shape of the GDM contour as well as the direction of lines on the GDM contour for all the projection cases in the VSM; and the Resolvability Model (RM) which models the threshold number of GDM lines allowed to be occluded such that the original shape of the GDM is unrecoverable. To apply the proposed models to resolve an N-vehicle occluded cluster, the number of vehicles (N) and their individual dimension are calculated as follows. First, we approximate the composite contour by the tangential slope curvature points detected on the contour, and then generate a description based on the direction of lines on the approximated contour with respect to the vanishing points of the road. Second, the approximated contour is segmented into N sub-segments with reference to the CDM, which together represent the partial GDM of the N occluded vehicles. N corresponds to the number of vehicles occluded in the cluster. Third, a resolvability index is then assigned to each vehicle based on the number of occluded lines of the partial GDM, which indicates whether it is possible to recover the original shape of that GDM. Fourth, the GDM of each vehicle is then completed based on the existing lines of the partial GDM. Fifth, by projecting the resolved GDM from 2D image coordinates to 3D world coordinates, dimension of the vehicle can be estimated. The proposed methodology has been evaluated on 267 sets of real world monocular traffic image sequences taken on a busy road containing 3074 different sized vehicles that are occluded in the images. From the evaluation, it is found that the calculation of N is 100% accurate. Moreover the proposed method can successfully compute the resolvability index of each vehicle, and resolve the GDM of individual vehicles. The percentage resolvability of the vehicles is found to be 93.82%. The average accuracy for width, length and height estimation is 94.66%, 93.75% and 95.31% respectively. The result has shown that the proposed method offers significant improvement in count accuracy over all other known methods. It also accurately classifies the severity of occlusion according to the resolvability index and is able to produce a fairly accurate estimation of vehicle dimension when the vehicle is resolvable. In general, it can potentially be generalized to resolve occlusion in the context of computer vision, such as human counting, crowd analysis, object counting, bin-picking and robotic vision. DOI: 10.5353/th_b3627575 Subjects: Traffic congestion Image processing Co