Spatio-temporal crowd density model in a human detection and tracking framework

Recently signicant progress has been made in the eld of person detection and tracking. However, crowded scenes remain particularly challenging and can deeply aect the results due to overlapping detections and dynamic occlusions. In this paper, we present a method to enhance human detection and tracking in crowded scenes. It is based on introducing additional information about crowds and integrating it into the state-of-the-art detector. This additional information cue consists of modeling time-varying dynamics of the crowd density using local features as an observation of a probabilistic function. It also involves a feature tracking step which allows excluding feature points attached to the background.

This process is favourable for the later density estimation since the influence of features irrelevant to the underlying crowd density is removed. Our proposed approach applies a scene-adaptive dynamic parametrization using this crowd density measure. It also includes a self-adaptive learning of the human aspect ratio and perceived height in order to reduce false positive detections. The resulting improved detections are subsequently used to boost the eciency of the tracking in a tracking-by-detection framework. Our proposed approach for person detection is evaluated on videos from dierent datasets, and the results demonstrate the advantages of incorporating crowd density and geometrical constraints into the detection process. Also, its impact on tracking results have been experimentally validated showing good results.