Research on human pose estimation technique based on depth-separable convolution
Yi-Hao Jiang, Hongfang Lv, Ran Ren · 2023
Traditional human pose estimation suffers from low real-time performance and accuracy. This paper suggests a solution to the mentioned challenges by presenting the Lean-OpenPose network, a lighter and more efficient version derived from the OpenPose research. By introducing the ECA attention mechanism in MobileNetV3, it replaces VGG19 in OpenPose as the new feature extraction network, optimizing subsequent multi-stage network structures with the idea of depth-wise separable convolution. The experimental findings reveal substantial enhancements in comparison to the original OpenPose network. The number of parameters is reduced by 76.14%, the floating-point calculation amount is reduced by 81.17%, and the fps performance sees an impressive improvement of 96.9%. It can quickly and accurately identify human keypoints, providing feasibility for deployment on edge terminal devices.