Research On Moving Object Detection Method Based On RPCA Theory

Moving Object Detection(MOD)is a critical step for video analysis applications.When processing complex video sequences,including camera jitter,illumination change,bad weather,etc.,the MOD methods based on background subtraction often fail to accurately detect the ob-jects due to the background modeling's terrible performance.Recently,the MOD schemes based on Robust Principal Component Analysis(RPCA)have aroused extensive attention.Al-though it has been applied in MOD successfully,some problems still need to be solved in the classical RPCA model:(1)Classic RPCA model can't deal with non-Gaussian modeling error efficiently;(2)Classic RPCA model can not detect the moving objects accurately;(3)Classic RPCA cannot fully utilize the spatial-temporal correlation among the frame sequence.The main works and innovations of this thesis are summarized as follows:(1)To constraint non-Gaussian modeling error effectively,this thesis proposes a modeling error constraint method based on the maximum entropy criterion(MCC).Thel₂-norm is re-placed by MCC,which can deal with non-Gaussian noise efficiently,making the RPCA model more robust to non-Gaussian error.Then,a foreground modeling method based on the Lapla-cian scale mixture(LSM)model is proposed.The LSM can impose moving objects in complex scenes effectively.Using LSM model instead of thel₁-norm,the object detection performance of RPCA is significantly improved.Finally,the MCC and LSM are combined to model the mod-eling error and foreground,respectively,and a novel moving object detection method,called Hyper RPCA,is proposed.Experimental results show that Hyper RPCA can model the non-Gaussian error effectively,and detect the objects effectively in the complex scene.(2)This thesis studies the Spatial-Temporal Laplacian mixed scale(STLSM)model,and a novel moving object detection method based on the STLSM model is proposed.There is an influential temporal correlation between adjacent frames and a great spatial correlation between pixels in the same frame for the video frame sequence.Classic RPCA cannot fully exploit the spatial and temporal correlation.In the STLSM model,the foreground estimation information obtained in the previous frame is used as a priori of the current frame's object model.Besides,the non-zero mean Laplace model constrains the temporal correlation of foreground compo-nents.In current frame,pixels with local similarity are modeled by the same STLSM model.Experimental results show that the object detection performance of the STLSM model is im-proved through the effective use of spatial-temporal correlation.