Detecting abnormal motion of pedestrian in video

Abstract

Visual analysis of human motion in video sequences has attached more and more attention to computer visions in recent years. In order to identify pedestrian movement in intelligent security monitoring system, moving body is detected and the boundary is extracted. According to the distance between contour points and the centroid, an exclusive 2-D (dimension) matrix is formed. In order to reduce computational cost affine transformation is proposed to normalize the matrix. And then the normalized matrix compares with the standard sequence which based formerly. The result is a vector, and then computes the standard deviation of the vector. A support vector machine (SVM) is presented to classify. In the realization of the system, first of all, a sequence of motive human images and unwrapped curve are proposed. And then the minimal standard deviation which is the difference between the standard and capture images is selected. Finally another compare between the neighbour and next frame can determine abnormal or not. Therefore, we can recognize some abnormal behaviors and then alarm, so that it becomes intelligible in nature. The results show that the new algorithm has better performance.