Adaptation in Non-Parametric State Estimation with Application to People Tracking

The employment of visual sensor networks in surveillance systems has brought in as many challenges as disadvantages. While the integration of multiple cameras into a network has the potential advantage of fusing complementary observations from sensors and enlarging visual coverage, it also increases the complexity of tracking tasks and poses challenges to system scalability. The research work in this thesis addresses the problem of building an efficient and scalable multi-camera tracking system that (i) adapts, in real time, to the dynamics of the monitored scenario and (ii) attempts to maximize usage and sharing of available sensing and processing resources. To perform reliable tracking, a preliminary step is fast and accurate people detection, for which we propose a locus-based probabilistic occupancy map (LPOM). The LPOM computes the probability of targets being in the map by using only motion information plus calibration data. To make the tracking system more scalable, we present a decentralized multi-camera multi-people tracking framework with a three-layer architecture, in which we formulate the overall task (i.e. tracking all people using all available cameras) as a vision based state estimation problem and aim to maximize utility and sharing of available sensing and processing resources. By exploiting the geometric relations between sensing geometry and people's positions, our method is able to dynamically and adaptively partition the overall task into a number of nearly independent subtasks with the aid of information theory, each of which tracks a subset of people with a subset of cameras (or agencies). The method hereby reduces task complexity dramatically and helps to boost parallelization and maximize the system's real time throughput and reliability while accounting for intrinsic uncertainty induced, e.g., by visual clutter, occlusion, and illumination changes. Moreover, we propose a preliminary information theoretical framework, in which we apply task-driven polling strategies that control the sensing process to minimize the number of observations to be processed while maximizing their expected impact on estimation process, and use parameter adaptation techniques to reallocate computational resources dynamically and opportunistically according to task complexity and measured evidence. For each proposed approach we carry out a number of experiments and demonstrate its efficiency and advantages.