Frequency Domain Analysis of Human Motions in Surveillance Video

A method to distinguish different human motions including walking, running, and wandering in the surveillance video is proposed in this paper. First of all, a block-based background extraction method is used to construct the background image. Second, the moving object can be detected by the use of RGB-based motion detection method and then the shadow removal scheme. Finally, a temporal signal representing the dynamic area of minimal rectangular regions covering the moving object is determined and then serves as a feature. By applying the discrete Fourier transform on the temporal signal, the various human motion statuses can be efficiently recognized according to the maximal magnitudes and the corresponding frequencies. The experimental results show that the accuracy achieves 90% accuracy in average for all the test video sequences.