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

Area of Science:

• Computer Vision
• Sports Science
• Artificial Intelligence

Background:

• Growing global tennis popularity necessitates intelligent systems for action recognition and feedback.
• Traditional methods lack the precision for fine-grained skill development in tennis.
• Challenges include occlusion, pose deformation, and multi-view consistency in tennis actions.

Purpose of the Study:

• To develop a novel, pose-driven action recognition model for tennis.
• To address limitations of existing methods in recognizing complex tennis actions.
• To introduce the Tennis Action Recognition You Only Look Once Detection Network (TAR-YOLO).

Main Methods:

• Developed TAR-YOLO based on the YOLO11 architecture.
• Proposed novel components: RES-Head for multi-scale feature fusion and DSAM for enhanced attention.
• Integrated SPD-Conv for improved feature extraction and Slide Loss for sample imbalance.
• Constructed a custom dataset, TAR-Det, for tennis pose estimation and action classification.

Main Results:

• TAR-YOLO achieved high performance on the TAR-Det dataset.
• Key metrics include 95.4% Precision, 93.7% Recall, and 96.2% mAP_0.5.
• Demonstrated efficiency with 16.9 FLOPs and 89.3 FPS.

Conclusions:

• TAR-YOLO effectively recognizes complex and dynamic tennis actions.
• The model shows significant potential for applications like AI-assisted coaching and real-time broadcasting.
• Architectural improvements enhance accuracy and efficiency in sports action recognition.