Probabilistic Multi-Person Tracking for Ambient Intelligence

Monitoring human activities with visual sensors is still a challenge, especially when multiple targets are involved. Occlusions, if not properly handled, are a major source of failure. Indoor environments with complex topology require the use of sensor networks whose effective management is by itself a difficult problem. Situated in the context of Ambient Intelligence, this thesis is concerned with both algorithmic and architectural aspects of distributed monitoring systems. The algorithmic approach pursued is that of non–parametric Bayesian filtering, a probabilistic state estimation framework whose multi target formulation allows physically–based modeling of the occlusion process within an appearance based, background independent observation model. While unaffordable in its plain, joint formulation, a novel, efficient, probabilistically sound solution is proposed and its robustness experimentally verified. Its MonteCarlo implementation provides a suitable framework for tackling dynamic resource allocation within a distributed modular architecture. Representation size of probabilistic estimates is adapted to conveyed uncertainty, while active environment sampling aims at minimizing uncertainty. Independent sensor agencies are synergically reconfigured by a supervisor process with global environment knowledge. The proposed system is then adaptive to its own performance, environment and sensing and computing infrastructure, providing a scalable solution to the problem of visual monitoring of crowded, topologically complex environments