Research On 2D Human Pose Estimation Based On Deep Adversarial Neural Networks

In the research of the human body pose estimation,recognizing images containing the human body and estimating the two-dimensional(2D)pose of the human body is a very important basic work in the field of machine vision research.Nevertheless,due to the complexity of the limbs in the image,the camera angle and other objective reasons,the human body in the image has varying degrees of distortion and occlusions;At the same time,in the actual application process of human body pose estimation,real-time and easy transplantation need to be guaranteed,and there are also certain requirements on the model complexity of the pose estimation network,which makes pose estimation be a very challenging Task: When the standard deep neural network(DNN)faces with more complex human poses and objective scenes,the estimation accuracy is often not high enough.;however,some posture estimation network models have extremely deep structure to guarantee estimation accuracy,resulting in high complexity.In this paper,based on the deep adversarial neural network to study the 2-Dimension human pose estimation,aiming at the above-mentioned problems,the following work is mainly completed:(1)Carry out a depth study of the basic network in the work of this article: Stacked Hourglass Network(SHN),and conduct a recurrence experiment.In the recurrence experiment,the output of target net is extracted and analyzed in different structural stages and different test sets.In the experiment,the percentage of correct keypoints,PCK and PCKh,were 89.0% and 88.5%,which confirmed the rationality and effectiveness of the SHN structure design.At the same time,it also found the shortcomings of SHN in posture estimation research: for the human body in the image joints are occluded,limb twisted,and limb feature information is not obvious.The accuracy of its pose estimation will be greatly reduced,which makes the performance of SHN pose estimation limited.The PCKh values of the two occlusion test sample sets are only 71% and 59%,respectively.(2)Apply the generative adversarial network to human pose estimation.SHN is subjected to two generation adversarial trainings,and the penalty reward mechanism,boundary parameter equalization mechanism,and body geometric constraints are incorporated into the second adversarial model to obtain the final post estimation model.The method was tested on a public data set,and the results showed that the method can effectively improve the accuracy of the pose estimation of SHN in complex situations,the PCK and PCKh values reached 94.8% and 92.2%,respectively the mean square estimation error(MSE)of the test samples was better reduced,and the PCKh estimates of the two occluded test sample sets reached 81% and 78%,respectively.however,due to the difference in pose estimation requirements and structural module design,some complexity related indicators such as average processing time,FLOPs and param number of the SHN model obtained from adversarial training in this chapter are somewhat increased compared with other SHN.(3)In response to the problems in(2),in order to reduce the size and complexity of the SHN model while keeping the pose estimation accuracy constant.Based on the experimental results of(1),reducing the number of stacks and the channels of the network is used in SHN for light weight treatment,thereby constructing a lowcomplexity SHN network;the residual module is replaced with the hourglass residual module to obtain larger receptive fields and local details;partial nearest neighbor upsampling is replaced with bilinear interpolation up-sampling to improve the quality of the feature image after up-sampling.The results of the ablation experiment and the adversarial training experiment show that the structure improvement of the network can effectively reduce the model size and complexity of the optimized target network SHN,complexity related indicators compared with the network in(2)have been greatly reduced;at the same time,the pose estimation accuracy of SHN has not been greatly affected.The PCK and PCKh values are 93.7% and 91.5%,respectively.The mean square estimation error(MSE)of the test sample is still lower than the original SHN,and the PCKh estimates for the two occlusion test sample sets are 77% and 73%,respectively.