Research On Robust Human Pose Estimation Algorithm For Complex Scenes

Human pose estimation means detecting and combining the key-points of the human body in order to show the posture in the image / video.As an important technical support to understand human behavior,it has important research significance and wide application value.In complex and changeable realistic scenes,factors such as occlusion,camera angle,and multi-scale bring about the negative performance of the human pose estimation algorithm.Therefore,the analysis of human pose with higher degrees of freedom still faces many challenges.This thesis is based on deep learning and attention mechanism,to study how to make full use of deep features to improve the model's performance of the human pose estimation for realistic complex scenes.The main work of this thesis is summarized as follows:1)A human pose estimation algorithm based on channel attention is proposed.Aim-ing at the problem of that human scales are different in images,a multi-resolution fea-ture extraction backbone network is constructed.Starting from the high-resolution sub-net,gradually increase the subnets from high-resolution to low-resolution,and the multi-resolution subnets are parallel to extract the multi-resolution features of the human body.Not only is essential extract feature information,but also needs to focus on some impor-tant and specific information in the images.On the basis of this backbone network,this chapter further introduces channel attention,and connects a branch network to calculate the weight on each convolutional layer in parallel to emphasize important feature informa-tion and suppress relatively secondary information.The method proposed in this chapter is verified on the COCO 2017 dataset.The average accuracy AP is 73.2%,and the per-formance is better than some SOTA human pose estimation models,such as SCARB and Simple Base Line.2)A human pose estimation algorithm based on channel frequency domain enhance-ment is proposed.In computer vision,the ability of neural network models extract feature information will affect the final results,and rich feature information can often bring better results.The global average pooling of traditional channel attention will suppress part of the feature information,so that the generated feature map contains only part of the in-formation in the original image.To solve this problem,this chapter converts the depth feature information of the image from the spatial domain to the frequency domain,and the discrete cosine transform operation is used to replace the traditional global average pooling,thereby retaining more depth feature information,constructing a frequency do-main enhanced channel attention to estimate human pose based on this algorithm.The method presented in this chapter is verified on COCO 2017 dataset,the average accuracy of 73.4% can be achieved.3)A human pose estimation algorithm guided by position and channel attention is proposed.For human pose estimation,the relationship among key-points and their posi-tion information are also very important,especially when some key-points are occluded,only considering the local information of key-points may case estimation errors.The in-troduction of channel attention only considers the importance of feature maps.For this reason,this chapter introduces position attention on the basis of channel attention to model the position information of the key-points of the human body in the image,constructs the joint representation of spatial long-distance features and interactive features,and designs the combination of channel and positional attention.The performance of jointly guided human pose estimation algorithm is more robust.The method proposed in this chapter is verified on the COCO 2017 dataset.The average accuracy AP is 73.7%,and the perfor-mance is better than some SOTA human pose estimation models,such as PRTR,RMPE and Pose Fix.