Direction estimation of crowd flow in surveillance videos

Estimation of density and direction of crowd flow from surveillance video has attracted much research attention recently in the area of computer vision. Crowd density in a video sequence can be considered as a global entity to estimate the direction of interest of the crowd. Though there are methods to compute the density and direction, a combined approach will give more insights into the problem. In this work, we have compared the simple approaches to find the density of the crowd from a given video sequence and then extended it to the computation of the global movement. Keypoint descriptors extracted from the scene are used to compute the dense areas which is further used to define the direction of the flow. Feature extraction methods like simple blob detector, SIFT, SURF, MSER are used for the estimation of direction. The proposed method is useful in the analysis of crowd behaviour and can bring out crowd semantics.