Overview
About the Project
This project is a computer vision system that analyzes basketball match videos to enable automatic statistics and score tracking.[cite: 1] In addition to detecting players and the ball, the developed system can determine which team the players belong to and who is currently in control of the ball.[cite: 1] Thanks to its user-friendly interface, the desired video can be uploaded to the system, and the analysis process can be easily initiated with resolution and hardware preferences.[cite: 1]
Key Features
- Player and Ball Tracking: Using YOLO Object Tracking, a unique ID is assigned to each object detected throughout the video stream, and continuous trajectory tracking of the objects is performed.[cite: 1]
- Dynamic Team Separation: Players are grouped according to feature similarities using the K-Means algorithm.[cite: 1] To increase the success of the algorithm, virtual crop areas focusing only on jersey color were created.[cite: 1]
- Score and Ball Handler Detection: The player closest to the center of the ball is assigned as the ball handler.[cite: 1] To attribute the score to the correct team during a shot, a memory mechanism that records past ball contacts was integrated into the system.[cite: 1]
- Custom Model Training: The initial model, which was insufficient, was retrained using the YOLOv8s architecture with a new dataset of 12,000 images prepared on Roboflow and supported by data augmentation techniques.[cite: 1]
User Interface
The project features a control panel that includes operations such as video uploading, model selection, resolution adjustments (e.g., 1280), and selecting the hardware unit to be used (e.g., CPU).[cite: 1]
