Human Pose Estimation Based on Improved HRNet Model

There exist some issues such as occlusions, variable human body poses, complex backgrounds in the human pose images, so there are still challenges in the task of human body pose estimation. By adding a new attention mechanism module and reweighting the last feature maps by the original HRNet, We propose an improved HRNet model. The ability of the model is enhanced to learn spatial and semantic information. The experiments on the COCO dataset and MPII dataset show that our model could detect some key points that are missed or detected incorrectly by the original network, and the accuracy is also increased.