An Improved Human Pose Estimation Algorithm Based on High-Resolution Net Model

The convolution kernel used in the High-Resolution Net (HRNet) model for the initial extraction of features is single in size, which does not fully combine the context information, resulting in weak feature extraction ability, and the convolution operation mainly focuses on the extraction of local features, which has limitations in capturing long-distance features and inter-channel features. In order to solve the above problems, a multi-scale feature fusion module is introduced, and a dilated convolution and 1×1 convolution are introduced for multi-scale feature extraction, and then the weights of the fused features are adaptively redistributed through the concatenation of Channel Attention(CA) and Spatial Attention(SA) modules. By introducing the channel prior convolutional attention(CPCA) module, and by constructing the CPCA-based bottleneck module CPCA-Bottleneck, the attention weights can be dynamically distributed in the channel and spatial dimensions, and the feature extraction ability can be further improved. Experimental results show that this method can effectively alleviate the above problems and improve the accuracy of human key point prediction.