Robust person and vehicle tracking for intelligent visual surveillance
Mangyu Kong · The University of Queensland · 2008
The problem of visual inspection of outdoor environments (e.g., airports, railway stations, roads,etc.) has received growing attention in recent times. The work presented in this thesis is acomponent of a larger project to apply intelligent Closed-Circuit Television (CCTV) to enhancethe counter-terrorism capability for the protection of mass transport systems. The purpose ofintelligent surveillance systems is to automatically perform surveillance tasks by applying camerasin place of human eyes. Recently, with the development of video hardware, the video surveillancesystem is becoming more widely applied and is attracting more researchers to develop fast androbust algorithms.In this thesis, we describe the proposed pedestrian classification and tracking system that isable to track and label multiple people in an outdoor environment such as a railway station. Wepropose an approach that combines blob matching with particle filtering to track multiple peoplein the scene, i.e., the proposed method selects the successful features of blob matching and particlefiltering for tracking. In our proposed method, the system can easily track persons even when theyare partially occluded by each other and can track them correctly after merging. In addition, anovel appearance model derived from the colour information from both the moving regions and theoriginal input colour image is proposed to track people in the event of poor foreground extraction.Additionally, the proposed appearance model also includes spatial information of the humanbody in both vertical and horizontal directions, making location more accurate. In the objectclassification stage, hierarchical chamfer matching combined with the particle filter is applied toclassify commuters in the railway station example into several classes. Based on single cameratracking, we extend our work to multiple camera-based people tracking using a two-level trackingscheme that includes image level tracking and particle filter-based ground level tracking.In addition, a novel method to extract cars from moving regions including shadow area, basedon shape and colour information, is proposed. Chamfer template matching score, and non-shadowregion edge score, are applied as the shape information; while the shadow confidence score(SCS) is used as the colour information. Experimental results show that the proposed methodis significantly better than the approaches where only the colour information is considered.