Real-time abnormal situation detection based on particle advection in crowded scenes

This paper presents a real-time abnormal situation detection method in crowded scenes based on the crowd motion characteristics including the particle energy and the motion directions. The particle energy is determined by computation of optical flow derived from two adjacent frames. The particle energy is modified by multiplying the foreground to background ratio. The motion directions are measured by mutual information of the direction histograms of two neighboring motion vector fields. Mutual information is used to measure the similarity between two direction histograms derived from three adjacent frames. The direction probability distribution for each frame can be directly estimated from the direction histogram by dividing the entries by the total number of the vectors. A metric for all the video frames is computed using normalized mutual information to detect the abnormal situation. Both the modified particle energy and mutual information of direction histograms contribute to the detection of the abnormal events. Furthermore, the dynamic abnormality is measured to detect the dynamical movement associated with severe change in the motion state according to the spatio-temporal characteristics. In experiments, we will show that the proposed method detects the abnormal situations effectively.