Tracking people across multiple cameras in the presence of clutter and noise
As video surveillance systems become more and more pervasive in our society, it is evident that simply increasing the number of cameras does not guarantee increased security, since each operator can only attend to a limited number of monitors. To overcome this limit, automatic video surveillance systems (AVSS, computer-based surveillance systems that automate some of the most tedious work of security operators) are being deployed. One such task is tracking, defined by the end users in this project as "keeping a selected passenger always visible on a surveillance monitor". The purpose of this work was to develop a single-person, multi-camera tracker that can be used in real time to follow a manually-selected individual. The operation of selecting an individual for tracking is called tagging, and therefore this type of tracker is known as a tag and track system. The developed system is conceived to be deployed as part of a large surveillance network, consisting of possibly hundreds of cameras, with possibly large blind regions between cameras. The main contribution of this thesis is a probabilistic framework that can be used to develop a multi-camera tracker by fusing heterogeneous information coming from visual sensors and from prior knowledge about the relative poisitioning of cameras in the surveillance network. The developed tracker has been demonstrated to work in real time on a standard PC independently of the number of cameras in the network. The developed tracker has been demonstrated to work in real time on a standard PC independently of the number of cameras in the network. Quantitative performance evaluation is carried out using realistic tracking scenarios.