GolfPoseNet: Golf-Specific 3D Human Pose Estimation Network

This paper aims to develop a robotic golf trainer using a wheeled robot and create a model that accurately recognizes the user's golf motion. Since existing 3D pose models show limitations in golf motion recognition [1], [2], we propose an optimized recognition model by combining various methods and preprocessing golf datasets to overcome these problems. We used HRNet to estimate 2D pose from video frames and fed these data into a transformer-based model to generate three hypotheses through residual linking and embedding layers. These were integrated to finally design a structure to predict 3D pose, and batch normalization, dropout, and LeakyReLU were used to improve the stability and nonlinearity handling of the model. Experimental results show that the proposed model effectively learns temporal information from multi-view 2D pose data without 3D annotation and performs well even when 3D data is scarce. The proposed method can be used in various applications such as human-robot interaction and sports motion analysis including golf, and further research on multiple datasets and real-world environments is expected to improve further and generalize the model's performance.