Automatic Estimation of Football Possession via Improved YOLOv8 Detection and DBSCAN-Based Team Classification

Area of Science:

• Computer Vision
• Sports Analytics
• Deep Learning

Background:

• Traditional sports analytics rely on manual data, which is time-consuming and subjective.
• Automating football analytics requires accurate and objective methods for tracking game events.

Purpose of the Study:

• To develop a deep learning framework for precise football possession estimation from broadcast video.
• To eliminate the need for manual annotations and event-based data in sports analytics.

Main Methods:

• Utilized YOLOv8-P2S3A and YOLOv8-HWD3A for object detection (football and players).
• Employed DBSCAN clustering for unsupervised team identification based on jersey colors.
• Integrated Norfair multi-object tracking and a temporal refinement module for possession duration accuracy.

Main Results:

• Achieved high validation average precision for football (79.4%) and player detection (71.1%).
• The system demonstrated superior possession estimation with a root mean square error (RMSE) of 4.87.
• Outperformed baseline models like YOLOv10n (RMSE: 5.12) and YOLOv11 (RMSE: 5.17).

Conclusions:

• The proposed framework significantly enhances precision, efficiency, and automation in football analytics.
• Offers practical value for coaches, analysts, and sports scientists in professional settings.
• Demonstrates the effectiveness of deep learning for objective sports data analysis.