Research On Human Pose Estimation Method Based On High Resolution Net

In the field of computer vision,human pose estimation can help people better understand human pose and behavior,and can also provide strong support for behavior recognition and detection.Relevant studies show that the human pose estimation method based on deep learning has shown more excellent results,which makes the human pose estimation gradually become hot research issues in computer science and other fields.With the wide application of deep learning technology,the human pose estimation method based on deep learning has significantly improved the effect of human pose estimation model.However,due to the influence of various scenes,human movement,clothing and other factors,the human body image has some problems,such as incomplete feature extraction and insufficient data of node location.At the same time,due to the great difference in the characteristics of different joints between each other,the accurate position of the output becomes more difficult.In order to further improve the accuracy and ability of the human pose estimation model,this thesis will be carried out from two aspects: joints estimation and feature fusion.The specific work is as follows.(1)Construct high-resolution network RDSCANet based on residual down-sampling and channel attention mechanism.In response to the problem of information loss in the down-sampling method of high resolution network,the residual down-sampling module is designed to transmit context information to the feature map,realize feature reuse and maintain the high resolution of the feature map,and reduce the loss of joints information in the process of down-sampling.At the same time,the channel attention mechanism is introduced to focus on the feature information of the global channel dimension to enhance the effective feature information in view of the impact of the increased context information on network learning.The experimental results on COCO2017 dataset show that combining the channel attention mechanism with the residual down-sampling module can improve the accuracy of human pose estimation.(2)Propose a high-resolution network MSDFFNet based on multi-scale dense feature fusion.A dense feature sampling module is designed to address the insufficient ability of the RDSCANet network to extract image features.The module uses convolutional kernels of different sizes for dense feature sampling of input images to obtain more effective feature information of human joints,thereby improving the robustness of the model and optimize the detection results in different scenes or occlusion in human pose estimation.In order to avoid the problem of information loss when fusing features at different scales,a multi-scale feature fusion module is introduced,which is based on Soft Pooling to design a fast pyramid pool structure.The experimental results on COCO2017 dataset show that the MSDFFNet model can make full use of the effective information in the feature map to improve the accuracy of human pose estimation.