Incrementally Detecting Moving Objects With Sparsity And Connectivity

Moving objects detection is crucial for intelligence video analysis.For videos captured by fixed cameras,Moving objects detection can be regarded as a separate problem between background and foreground.The background is typically assumed to be spanned by a low dimensional subspace and the foreground is the outliers.However,in real world cases,many cases would violate the assumption.On the one hand,dynamic backgrounds like waving trees and water ripples,as well as noises would not lie in the low dimensional subspace,which would lead to increase the false positive rate.On the other hand,camouflage and lingering foreground objects would also significantly increase the difficulty of accurate foreground detection,which would lead to increase the miss positive rate.Thus,it is a challenging task to robustly and efficiently detect moving object in video.DEtecting Contiguous Outliers in the LOw-rank Representation(DECOLOR)is the state-of-the-art batch-based moving objects detection algorithm.It not only made the assumption of background but also made full use of the sparsity and connectivity of foreground,which obtained a good detection result.However,it could not deal with the videos frame by frame and its efficiency was very low.Grassmannian Robust Adaptive Subspace Tracking Algorithm(GRASTA)is the most popular incrementally moving objects detection algorithms in recent years.Moreover,it could process the videos frame by frame and it was efficent.However,it simply imposed the sparsity constraint on the foreground and was no longer sufficient.Therefore,this paper proposes a new algorithm which further takes into account the foreground characteristics including both sparsity and smoothness.An efficient alternative minimization algorithm is proposed to seek the optimal solutions.Experimental results on public databases demonstrate the effectiveness of the proposed method.