GolfPose: Golf Swing Analyses with a Monocular Camera Based Human Pose Estimation

With the rapid developments of computer vision and deep learning technologies, artificial intelligence takes a more and more important role in sports analyses. In this paper, to attain the objective of automated golf swing analyses, we propose a lightweight temporal-based 2D human pose estimation (HPE) method, called GolfPose, which achieves improved performance than the state-of-the-art image-based HPE methods. Unlike traditional image-based methods, our temporal-based method, designed for efficient and effective golf swing analyses, takes advantage of the temporal information to improve the estimation accuracy of fast-moving and partially self-occluded keypoints. Furthermore, in order to make sure the golf swing analyses can run on mobile devices, we optimize the model architecture to achieve real-time inference. With around 10% of the parameters and half of the GFLOPs used in the state-of-the-art HRNet, our proposed GolfPose model can achieve 9.16 mean pixel error (MPE) in our golf swing dataset, compared with 9.20 MPE for HRNet. Furthermore, the proposed temporal-based method, facilitated with golf club detection(GCD), significantly improves the accuracy of keypoints on the golf club from 13.98 to 9.21 MPE.