OpenPose Based Posture Analysis of Basketball Players

Authors:

Abstract
Human pose estimation has a wide range of applications in the field of computer vision, especially in sports event analysis where its importance is increasingly prominent. This study aims to optimize the basketball player detection model based on OpenPose, improving its performance in real-world scenarios. Firstly, we replace the VGG19 network in Openpose with a lightweight MobileNet structure to increase computational efficiency and reduce model parameters. Secondly, we introduce the Pyramid Spatial Attention (PSA) module into MobileNet to better capture local and global contextual information. Lastly, we adopt the Adaptive Wing Loss as the loss function, making the model more stable and accurate in localization. We pre-train the model on the COCO2017 dataset and conduct experiments on basketball player detection tasks. The experimental data show that, compared to the baseline model, our full model has significant improvements in Average Precision (AP), AP50, and running speed (FPS) metrics. Specifically, the full model achieves an AP of 70.0%, AP50 of 88.0%, and FPS of 41. The research results of this study can provide more accurate pose estimation information for basketball game analysis, contributing to the improvement of game strategy formulation and player performance evaluation.