Automated vision-language framework for kinematic profiling and performance diagnostics in competitive swimming

Area of Science:

• Sports Science
• Computer Vision
• Artificial Intelligence

Background:

• Manual kinematic analysis in swimming is time-consuming and not immediate enough for elite training.
• Current methods lack the efficiency needed for real-time feedback during high-level athletic preparation.

Purpose of the Study:

• To develop an automated vision-language framework for real-time kinematic analysis in swimming.
• To provide immediate coaching diagnostics and performance profiling for swimmers.

Main Methods:

• Utilized YOLOv11 for object detection with a splash-injection strategy to manage aquatic occlusion.
• Extracted kinematic metrics (velocity, distance) using homography transformation.
• Integrated the DeepSeek-V3 Large Language Model for automated pedagogical feedback and report generation.

Main Results:

• Achieved a mean Average Precision (mAP@0.5) of 94.64% in challenging dynamic water conditions.
• Successfully tracked swimmers through water turbulence.
• Synthesized quantitative kinematic data into natural language assessments of pacing and fatigue.

Conclusions:

• The open-source framework significantly decreases the manual workload associated with performance analysis.
• Combines computer vision and automated reporting for a scalable, objective tool.
• Suitable for daily swim training and technical evaluation in competitive swimming.