A lightweight pose-aware YOLO framework for real-time tennis stroke recognition and training analysis

Real-time tennis action recognition is important for sports video analysis and training review, yet many pose-aware models remain computationally expensive for routine use. This article proposes HL-TAR-YOLO, a lightweight pose-aware YOLO framework for fine-grained tennis stroke recognition. Built on YOLO11-Pose, the framework integrates DualConv for efficient feature extraction, SimAM for parameter-free spatial refinement, DySample for dynamic upsampling, and Slide Loss for adaptive optimization under sample difficulty imbalance. The proposed model was evaluated on TAR-Det, a pose-annotated tennis action dataset introduced in our previous work, and compared with representative lightweight and pose-aware baselines in terms of precision, recall, mAP, parameter count, FLOPs, and inference speed. Experimental results show that HL-TAR-YOLO achieves 95.2% precision, 93.9% recall, 94.3% mAP_0.5, and 91.0% mAP_0.5:0.95, while using only 2.6 M parameters and 6.9 GFLOPs and running at 122.3 FPS under the reported GPU evaluation setting. Additional CPU inference analysis showed lower single-frame latency than YOLO11s-Pose. These results suggest that HL-TAR-YOLO provides an effective accuracy-efficiency trade-off for resource-constrained tennis stroke analysis, while temporal modeling, external validation, and real-device evaluation remain important directions for future work.