Low-Rank Sparse joint Representation for Moving Object Detection in Video

Abstract

For the video sequences with fixed cameras, it is a reasonable assumption that the fixed background has low-rank characteristic, and the dynamic foreground has sparse characteristic. A new motion detection method based on low-rank and sparse joint representation is proposed in this paper. The ideas of the proposed method are described as follows: The noise of video sequence is removed by image preprocessing. The optical flow between continuous video sequences is estimated, which is used to generate a binary motion mask as a movement weight matrix. An optimization model with low-rank background and sparse foreground is established based on the idea of subspace learning theory. The background and foreground of each frame are obtained by using the ADMM-BCD iterative algorithm. Experimental results show that the proposed method is super to the other same sort of moving detection methods. The proposed method has perfect effect on slow moving target detection.