Incremental Moving Object Detection Using Saliency Detection And Subspace Learning

Moving object detection is the basis of intelligent video analysis.For videos captured by static cameras: The background component of the images are correlated with each other,so the background matrix is low-rank;besides,the foreground is relatively smaller than the whole image,then the foreground is considered sparse;what's more,the foreground has connectivity in spatial.The state-of-art moving object detecting methods based on matrix decomposition mainly utilize the properties of low-rank,sparsity,and connectivity.Generally,they seek the best solution by alternating algorithm.Even though the state-of-art methods have achieved some effects,but there are still some problems: on the one hand,what moves are not only the meaningful objects(for example,people),but also parts of background(such as waving trees,and water ripples),which will cause some false alarms.On the other hand,it is difficult to detect the moving objects whose textures are not rich.In fact,the salient object detection can clearly divide the foreground and background,and it is robust to the dynamic background which is generated by the meaningless moving objects.To improve the false alarms caused by the dynamic background and missed alarms caused by the not rich texture of the moving objects,an incremental moving object detection in video based on saliency map and subspace learning is proposed.The proposed algorithm form a saliency map constraint term of the background,and incorporate it with background reconstruction term,foreground sparsity term,and foreground connectivity term,finally form an optimization problem.Experimental results on both the Perception Test Images Sequences dataset and Wallflower dataset demonstrate that the proposed method is effective in reducing false and missed problems.Besides,the proposed method belongs to the incremental moving object detection method,so it satisfies the real time demand of moving object detection in video.