Research On Human Pose Estimation Method Based On Improved Deep Neural Network

Human body pose estimation aims to locate the key points of human body parts and connect them correctly to form a human body skeleton to describe the current human body pose information.In the field of computer vision,human pose estimation,as the basis for understanding high-level human behavior in images and videos,is still an important and challenging task,and has broad application prospects in many fields such as smart security,smart medical treatment,fashion entertainment,and so on.In this paper,by analyzing the research status of human body pose estimation methods at home and abroad,combined with the relevant knowledge of deep neural network,we conducted in-depth research on how to effectively improve the accuracy of image-based human pose estimation and training efficiency.Specific research contents are as follows:Aiming at the low accuracy of human pose estimation under different appearances,viewing angles,occlusions,cluttered backgrounds and inherent geometric ambiguities,this paper proposes a human pose estimation method based on Composite Residual Block Stacked Hourglass Network(CRSH).Firstly,by improving the original residual block(RB),a large receptive field residual block(LRFRB)and a multi-scale residual block(MSRB)are proposed.Secondly,an hourglass subnetwork is designed based on RB,LRFRB and MSRB.Then,four hourglass subnetworks are cascaded,and intermediate supervision is set at the end of each hourglass subnetwork to form the final composite residual block stacked hourglass network.Finally,a human pose estimation algorithm based on CRSH is designed.Through experiments on multiple data sets,the results prove that the network effectively improves the accuracy of human pose estimation tasks.Aiming at the problem that the large number of parameters in CRSH causes the training process to consume too much resources and time,this paper proposes a human pose estimation method based on the Lightweight-Composite Residual Block Stacked Hourglass Network(LW-CRSH).Firstly,the depth separable convolution is used in LRFRB and MSRB to reduce the amount of model parameters.Secondly,based on the idea of channel separation and reorganization,Channel split modules and Channel shuffle modules are added to reduce the number of transmission channels while enhancing feature fusion.Then,in the first half of LW-CRSH,Lightweight-Pyramid Residual Block(LW-PRB)is used to further reduce the amount of parameters and calculations.Finally,the human body pose estimation algorithm based on LW-CRSH is given.Through experiments on multiple data sets,the results prove that the network can effectively improve the training efficiency while ensuring the accuracy of prediction.