Author: Avi Shah (2023)
Score Detection Accuracy: 95%
Shot Detection Accuracy: 97%
This project combines the power of Machine Learning and Computer Vision for the purpose of detecting and analyzing basketball shots in real-time! Built upon the latest YOLOv8 (You Only Look Once) machine learning model and the OpenCV library, the program can process video streams from various sources, such as live webcam feed or pre-recorded videos, providing a tool that can be used for an immersive playing experience and enhanced game analytics.
The training process utilizes the ultralytics YOLO implementation and a custom dataset specified in the 'config.yaml' file. The model undergoes a set number of training epochs, with the resulting weights of the best-performing model saved for subsequent usage in shot detection. Although this model worked for my usage, a different dataset or training method might work better for your specific project.
The core of this project is an algorithm that uses the trained YOLOv8 model to detect basketballs and hoops in each frame. It then analyzes the motion and position of the basketball relative to the hoop to determine if a shot has been made.
To enhance the accuracy of the shot detection, the algorithm not only tracks the ball's position over time but also applies data-cleaning techniques to both the ball and hoop positions. The algorithm is designed to filter out inaccurate data points, remove points beyond a certain frame limit and prevent jumping from one object to another to maintain the accuracy of the detection.
A linear regression is used to predict the ball's trajectory based on its positions. If the projected trajectory intersects with the hoop, the algorithm registers it as a successful shot.
Please ensure you have the required Python packages installed, including OpenCV, numpy, and ultralytics' YOLO. Contributions to this project are welcome - submit a pull request. For issues or suggestions, open an issue in this repository.
The model's performance can vary based on factors such as the quality of the video feed, lighting conditions, and the clarity of the basketball and hoop in the video. Furthermore, this program will not work if multiple basketballs and hoops are in frame. For testing, this program had input videos that were shot outdoors from a phone camera on the ground.