Macroscopic analysis of crowd motion in video sequences
Arie Nakhmani, Amit Surana, Allen Tannenbaum · 2014
We use dynamic active contours driven by optimal mass transport optical flow to detect crowd behaviors, in particular crowd merging, splitting and collision events. The overall framework is variational, and thus one could very naturally formulate functionals which include geometric active contours together with optical flow ideas. This allows to fuse temporal and intensity distribution information explicitly into a single framework. From a networking point of view, we consider here macroscopic models as opposed to microscopic or mesoscopic approaches to crowd dynamics. Our experiments show high detection rate of macro crowd behaviors with complicated real world scenarios.