Human Pose Recognition Based on KAN's Improved HRNet
Jun Hou, Harry Liu, Yinjing Du, Shun Peng, Ruizhe Bian · 2025
In order to ensure the further safety of workers in circuit secondary maintenance, this paper aims at the problem of low accuracy of human posture estimation, and proposes a new network KHRNet for network optimization using HRNet as the backbone network to improve the accuracy of model detection. First of all, faced with the problems of poor fitting effect, weak theory interpretability and high operation cost of deep neural network, KHRNet network is formed by adding KAN difference module to HRNet. HRNet is known for its high-resolution feature extraction capability. KAN is introduced into its high-low resolution branch to optimize feature expression and improve the nonlinear mapping capability of the model, thus enhancing the accuracy of attitude estimation. The experimental results show that compared with the benchmark model HRNet method, the improved network model significantly improves the detection accuracy of human pose estimation, the average accuracy of COCO in the public data set is up to85.8%, and the AP in MPII in the public data set is up to90.1%.