TAR-YOLO: A Novel Deep Learning Model and Dataset for Tennis Action Recognition

With the growing popularity of tennis globally, there is an increasing demand for intelligent systems capable of accurate action recognition and timely feedback, with potential applications in real-time broadcasting, AI-assisted coaching, skill evaluation, and injury prevention. Traditional approaches, which often rely on manual observation and delayed correction, struggle to meet the needs of fine-grained skill development. This paper presents the Tennis Action Recognition You Only Look Once Detection Network (TAR-YOLO), a novel, pose-driven action recognition model based upon the YOLO11 architecture, addressing challenges such as occlusion, pose deformation, and multi-view consistency. In this research, two novel components RES-Head and DSAM are proposed, while SPD-Conv and Slide Loss are integrated into this model. These four key architectural improvements significantly enhance the performance of TAR-YOLO, with RES-Head enabling multi-scale feature fusion, DSAM enhancing attention-based representation of key motion cues and deformable action patterns, SPD-Conv improving feature extraction, and Slide Loss addressing sample imbalance through dynamic gradient reweighting during training. A custom dataset, TAR-Det, specifically designed for tennis pose estimation and action classification, is also constructed. Experimental results show that TAR-YOLO achieves a Precision of 95.4%, Recall of 93.7%, mAP_0.5 of 96.2%, mAP_0.5:0.95 of 93.5%, FLOPs of 16.9, and FPS of 89.3 on the TAR-Det dataset, confirming its effectiveness in complex and dynamic tennis action recognition tasks.