Lightweight Open Pose Based Body Posture Estimation for Badminton Players

Abstract

In this paper, we optimize the bottom-up human pose estimation network OpenPose in the context of badminton player's body pose, replace the VGG19 backbone network in this model with a lightweight network MobileNet with less number of parameters to achieve a lightweight model, and introduce a polarized self-attention mechanism ( Polarized Self-Attention (PSA) is introduced at the front and back ends of the MobileNet backbone network to achieve a reasonable overhead while keeping the input features as high resolution as possible. The results show that the mAP0.5 (%) and mAP0.75 (%) of the optimized model are 85.79% and 70.57%, respectively. Compared with the original OpenPose model, the detection speed is significantly improved, and the FPS of the optimized model is improved by 64.28%, although the average accuracy is slightly reduced. Finally, the experiments show that the optimized lightweight OpenPose model has good estimation accuracy and estimation speed in estimating the body pose of badminton players and can be applied in embedded devices.