An enhanced YOLOv8-pose network for joint image-based tennis action classification and human pose estimation

Abstract This study addresses joint single-frame player-instance tennis action classification and human pose estimation from individual RGB images. We propose an enhanced YOLOv8-Pose network that jointly predicts player bounding boxes, action labels, and 17 human keypoints. The network combines an enhanced StarNet backbone for efficient multi-scale representation, a motion-relevant attention module that emphasizes action-discriminative body configurations visible in one image, and a composite pose loss for accurate keypoint localization. On the primary four-class Tennis Action dataset, the proposed model obtains Pose AP50 and AP50:95 values of 0.975 and 0.746, respectively, together with a matched-instance action accuracy of 0.990, matching coverage of 1.000, 3.08 million parameters, 8.71 GFLOPs, and 51.7 FPS on an NVIDIA GeForce RTX 5070 GPU. To determine whether the pose-estimation gain persists under a second independent training setting, the model is separately trained, validated, and tested on the two-class tennis-player-pose dataset using its official split, where it achieves Pose AP50 and AP50:95 values of 0.820 and 0.549. The results indicate that the proposed architecture maintains a favorable accuracy–efficiency trade-off in both independently trained dataset settings.