Automatic Estimation of Movement Statistics of People

Automatic analysis of how people move about in a particular environment has a number of potential applications. However, no system has so far been able to do detection and tracking robustly. Instead, trajectories are often broken into tracklets. The key idea behind this paper is based around the notion that people need not be detected and tracked perfectly in order to derive useful movement statistics for a particular scene. Tracklets will suffice. To this end we build a tracking framework based on a HoG detector and an appearance-based particle filter. The detector is optimized by learning a scene model allowing for a speedup of the process together with a significantly reduced false positive rate. The developed system is applied in two different scenarios where it is shown that useful statistics can indeed be extracted.