Lightweight human pose estimation method based on HRNet

In the field of human pose estimation, HRNet (high-resolution network) has excellent performance. In order to make it deployable in mobile devices with weak computing power, we designed it to be lightweight. Through similarity analysis on the feature maps output by each convolutional layer, we found that some channels are redundant. Partial convolution can be used to reduce computational redundancy. We redesigned the feature extraction module of HRNet using partial convolution. Partial convolutional module reduced computational complexity and avoided frequent memory access, allowing the network to extract spatial features more effectively. Furthermore, we calculated the similarity based on the feature maps generated by each convolutional layer, adjusted the number of channels in each module of the network, and deleted redundant channels. Compared with HRNet, our lightweight high-resolution network improves FPS by 54% and 67% on CPU and GPU respectively, making it more suitable for use in mobile devices.