Matching gradient descriptors with topological constraints to characterise the crowd dynamics
Buchao Zhan, Paolo Remagnino, Sergio A. Velastín, François Brémond, Monique Thonnat · 2006
Understanding complex crowd scenes involves many dimensions: level of clutter, density of pedestrians, global/detail dynamics of the scene, etc. Socio-dynamics has tackled the problem by providing prototypes of crowd behaviour based on human observations and then explaining them by certain physical models. We are interested in automatic learning the factors involved in complex scenes for different dimensions and deriving a physical model from the real world via computer vision methods. In this paper we propose a solution to extract the motion of complex crowd which could be very difficult for conventional trackers. The method is based on a matching of local gradient descriptor supported by local topology constraints. Spatial pyramid and temporal smoothing are employed for optimizing the algorithm. Dynamics can then be accumulated over time so as to derive a higher level understanding of the scene.