Gate and Common Pathway Detection in Crowd Scenes Using Motion Units and Meta-Tracking

This paper proposes a new approach for analysing crowded video scenes. The proposed approach decomposes the scene motion dynamics into a graph of interconnected atomic elements of coherent motions named Motion Units (MUs). Different MUs cover scene's local regions with different size and shape, which can even overlap. MUs relationships are analysed to discover the scene entrances and exits. Dominant motion pathways are then discovered by meta-tracking of particles injected at the scene entrances and driven through MUs using their linear dynamical systems until reaching scene exits. A prototype is developed such that; MUs are constructed by tracklet clustering, MU's motion pattern is represented by a linear model, and the MUs relationships are defined by the continuation likelihood among their mean tracklets. The prototype was evaluated on the challenging New York Grand Central Station scene, as well as other crowded scenes, and it managed to outperform the state of the art approaches.