Concerning the fact that common background subtraction methods ignore the spatio-temporal continuity of foreground and the disturbance of dynamic background to foreground extraction, an improved background subtraction model was proposed based on Tensor Robust Principal Component Analysis (TRPCA). The improved tensor nuclear norm was used to constrain the background, which enhanced the low rank of background and retained the spatial information of videos. Then the regularization constraint was performed to the foreground by 3D Total Variation (3D-TV), so as to consider the spatio-temporal continuity of object and effectively suppress the interference of dynamic background and target movement on the foreground extraction. Experimental results show that the proposed model can effectively separate the foreground and background of videos. Compared with High-order Robust Principal Component Analysis (HoRPCA), Tensor Robust Principal Component Analysis with Tensor Nuclear Norm (TRPCA-TNN) and Kronecker-Basis-Representation based Robust Principal Component Analysis (KBR-RPCA), the proposed algorithm has the F-measure values all optimal or sub-optimal. It can be seen that, the proposed model effectively improves the accuracy of foreground and background separation, and suppresses the interference of complex weather and target movement on foreground extraction.

In order to solve the problems of too many parameters, long training time and slow detection speed of the existing human skeleton key point detection models, a detection method combining the human skeleton key point detection model called Convolutional Pose Machines (CPMs) and the lightweight convolutional neural network model called SqueezeNet was proposed. Firstly, the CPMs with 4 stages (CPMs-Stage4) was used to detect the key points of the human images. Then, the Fire Module network structure of SqueezeNet was introduced into CPMs-Stage4 to reduce the model parameters greatly, and thus to obtain a new lightweight human skeleton key point detection model called SqueezeNet15-CPMs-Stage4. The verification results on the extended Leeds Sports Pose (LSP) dataset show that, compared with CPMs, SqueezeNet15-CPMs-Stage4 model has the training time reduced by 86.68%, the detection time of single image reduced by 44.27%, and the detection accuracy of 90.4%; and the proposed model performs the best in training time, detection speed and accuracy compared with three reference models improved VGG-16, DeepCut and DeeperCut. The experimental results show that the proposed model achieves high detection accuracy with short training time and fast detection speed, and can effectively reduce the training cost of the human skeleton key point detection model.

In view of the fact that the low rank and sparse methods generally regard the foreground as abnormal pixels in the background, which makes the foreground detection precision decrease in the complex scene, a new foreground detection method combining weighted Schatten- p norm with 3D Total Variation (3D-TV) was proposed. Firstly, the observed data were divided into low rank background, moving foreground and dynamic disturbance. Then 3D total variation was used to constrain the moving foreground and strengthen the prior consideration of the spatio-temporal continuity of the foreground objects, effectively suppressing the random disturbance of the anomalous pixels in the discontinuous dynamic background. Finally, the low rank performance of video background was constrained by weighted Schatten- p norm to remove noise interference. The experimental results show that, compared with Robust Principal Component Analysis (RPCA), Higher-order RPCA (HoRPCA) and Tensor RPCA (TRPCA), the proposed model has the highest F-measure value, and the optimal or sub-optimal values of recall and precision. It can be concluded that the proposed model can better overcome the interference in complex scenes, such as dynamic background and severe weather, and its extraction accuracy as well as visual effect of moving objects is improved.