Research on human pose estimation algorithm for occlusion scene
Hui Gao, Ning Wu, Zhigang Lv, Xiaoyan Li, Ruohai Di, Yuanqiang Lin · 2023
This paper proposes a human pose estimation algorithm for obscured scenes with poor prediction results caused by self or external objects. Firstly, a lightweight upsampling operator CARAFE is introduced to optimize the traditional nearest neighbor interpolation to reduce the localization error caused by interpolation and thus improve the localization accuracy of key points. Secondly, by building an improved multiscale feature fusion module WASP, the network can fully utilize and fuse different resolution features to ensure the adequacy of feature information. The experimental results show that the detection accuracy AP of this algorithm can reach 76.7 %, which is 0.8 % higher than the original algorithm, and effectively improves the performance of human key point detection in occlusion scenes.