Assessment of Automated Crowd Behaviour Analysis Based on Optical Flow

In visual surveillance, camera streams are often used to keep an eye on dense crowds. The examination of this data is mostly done manually by observers. When analysing multiple cameras some assistance is desirable. Computer vision methods can be used to assist observers in detecting crowd behaviours. Methods based on optical flow are particularly interesting since they can examine high density crowds with cluttering and (partial) occlusion without increasing computing costs. Not many methods can detect specific behaviour of dense crowds without the need of a learning stage. One promising method by Solmaz et al. uses the Jacobian stability of the optical flow field in the scene to detect five behaviour patterns viz. blocking, bottlenecks, fountainheads, rings and lanes. The method is implemented and a demo program is written with which experiments are performed on several datasets. The detection of three out of five behaviour patterns turn out to be promising, for the latter two improvements are proposed