Methods For Moving Objects Detection Under Dynamic Interference And Intermitted Motion

Moving objects detection is a research hotspot in computer vision and pattern recognition,which is widely used in security monitoring,intelligent traffic,action recognition,and human-computer interaction.The purpose of moving objects detection is to detect moving objects in a video quickly and accurately based on the correlation between video frames.Generally,moving objects detection can be converted into an optimization problem according to the low rank of the background matrix and the sparseness of the moving objects matrix.At present,many methods can effectively detect objects which move continuously in a static background.However,in the actual scene,the background usually contains various types of dynamic interference,and the detected objects do not always move continuously,but often move intermittently.For human,these problems are easy to deal with,but for computers,they are challenging.Therefore,the problem of dynamic interference and intermitted motion in moving objects detection is researched,and the improved moving objects detection methods named IMNTV-RPCA and IMONet based on low-rank sparse representation and convolutional neural networks respectively are proposed.First,an improved detection model MNTV-RPCA is proposed by introducing mixed norm and total variation to Robust Principal Component Analysis(RPCA).It decomposes the residual matrix S,so it can detect pure moving objects from scenes containing dynamic interference.Further,in order to improve the convergence speed of the algorithm,we remove the residual matrix S in MNTV-RPCA and replace the nuclear norm with the low-rank approximation in Robust Orthogonal Subspace Learning(ROSL)to propose the improved model named IMNTV-RPCA.Compared with MNTV-RPCA,it has faster detection speed and comparable accuracy.The experimental results show that MNTV-RPCA and IMNTV-RPCA can effectively detect objects moving continuously under dynamic interference.Secondly,because MNTV-RPCA and IMNTV-RPCA are based on background modeling,they can suppress the effects of dynamic interference,but they cannot achieve ideal results when detecting objects moving intermittently.To solve this problem,deep learning is applied to moving objects detection,and an end-to-end Moving Objects Network(MONet)is designed.MONet is no longer based on the background modeling,but segments moving objects directly from the scene according to the correspondence between pixels.Further,in order to overcome the shortcoming of MONet that it can not detect complete objects,we introduce the low-level motion feature into the MONet decoding network,and propose Improved MONet(IMONet).The experimental results show that MOnet and IMONet can both detect objects moving continuously under dynamic interference and effectively detect objects moving intermittedly.