Kullback-Leibler threshold computing method based on clustering motion patterns

Abstract

In order to cluster the moving targets of the video in dense crowded scene, according to the video regulation reticulation, it can acquire the spatio-temporal motion patterns of every grid by spatio-temporal gradient. Using the symmetric K-L(Kullback-Leibler) divergence as a distance measure, the clustering of the spatio-temporal motion patterns could be finished. The accuracy of the clustering plays an important role in the target detection, and the K-L threshold is the key for the accuracy of the clustering. Different K-L threshold will lead to different clustering effect. This paper proposes a method based on the dichotomy combining power of the motion pattern to determine the K-L threshold accurately and quickly.