An improved human pose estimation model based on DEKR
Jiarui Luo, Peng Han, Jian Qiu, Dongmei Liu, Miao Chen, Kaiqing Luo · 2024
Human pose estimation in crowded scenes has always been a challenging task in bottom-up multi-person pose estimation. To improve the accuracy of pose estimation in dense crowds, we propose an improved bottom-up human pose estimation model called H-DEKR, which is based on Disentangled Keypoint Regression for Bottom-Up Human Pose Estimation (DEKR). The model first enhances the coarse/fine-grained feature extraction abilities of the backbone (HRNet) by introducing different structures of Polarized Self-attention (PSA). Then, Pyramid Convolution (PyConv) is introduced to extract multi-scale information, alleviating the problem of uneven human scales. Results show that our model based on HRNet-W32 achieves accuracy of 67.1% on the CrowdPose dataset, which is 1.4% higher than the DEKR, respectively. Therefore, the proposed model in this paper is able to improve the accuracy of human pose estimation in dense crowds.