Research On Occluded Human Pose Estimation Based On GAN And Hourglass Network

In the field of contemporary artificial intelligence and natural human-computer interaction,human pose estimation,as the basis of human behavior recognition and analysis technology,is a key step for computers to understand human behavior in images and videos,and has been widely concerned by many scholars at home and abroad.In the field of the human body posture estimation,as the human body in the real life often interact with the surrounding environment,within the scope of the camera view often produce condition,let the body and the environment,it is difficult to distinguish between the shade problem not only in body posture estimation algorithm to predict the image point accuracy of obscured area drastically,And the accuracy of key points in the surrounding area will be affected to a certain extent.Compared with the traditional method of artificial feature design,CNN can extract more accurate and robust convolutional features than artificial features,and has the ability to implicitly save and model the spatial location relationship between key points.Therefore,this paper chooses the human pose estimation method based on deep learning to study the occlusion problem.Based on the existing public data set occlusion samples and occlusion annotation is limited,this paper proposes a simulated occlusion method,and uses the simulated occlusion to evaluate the impact of human occlusion on the prediction of human key points.The experimental results of simulated occlusion are analyzed in detail.In this paper,two improved methods are proposed based on the fusion of stacked hourglass network and generative adversarial network to solve the occlusion problem.(1)The SE channel attention mechanism is added to the basic residual module of the fusion network,and different weights are added to different feature channels in the CNN so that the network can pay more attention to the channel where the features of the visible region are located.Network output prediction(2)in fusion heat map parts add a shade attention branch,will predict heat map as the key point of spatial attention with color image after histogram normalization multiplication as input,through the block classification output current state key to keep out the visibility of the vector,and the visibility vector coding of each channel is visible through the connection layer weights,Finally,the visible weight is multiplied to each feature channel of the convolution feature.Experimental results show that the prediction of the two improved methods is more consistent with the topology of human body.Among them,the improved method based on SE attention achieved 92.1% and 72.9% accuracy in the verification set of MPII dataset for visible key points and occlusion key points,respectively,and improved by0.5% and 1.9% in parallel comparison with hourglass network.The improved method based on occluded attention achieves 91.7% and 71.8% accuracy of visible key points and occluded key points,respectively,and the parallel comparison with hourglass network improves by 0.1% and 0.8%.