Human Pose Estimation of Diver Based on Improved Stacked Hourglass Model

In this paper, a network structure is proposed for the task of single person pose estimation in a complex environment. This method improves the stacked hourglass model, achieves the feature extraction on most scales, and raises the detection accuracy of human key points. In the hourglass module, we use convolution operation to complete the upsampling to get more semantic information. When the responses of the two residual elements are added, we replace the identity mapping in the residual element with the 1×1 convolution element module to improve the phenomenon of variance explosion. We conducted model evaluation experiments on MPII and LSP data sets, and the results showed that the average detection accuracy of key points was improved by 0.2% and 0.8% respectively through our improvement on the stacked hourglass model.