Camera Calibration in Sports with Keypoints

Camera calibration is crucial for accurate vision AI systems in sports, as it enables the mapping of movements from video frames to real-world actions on the field, helping track distance, direction, and speed. The technique of homography is used for this purpose, but the dynamic nature of sports events, with cameras frequently changing positions and angles, makes manual calibration challenging. To address this, an Ultralytics YOLOv8 keypoint detection model is trained to automatically identify specific points on a soccer field in video frames, allowing for the establishment of source and target points needed for homography calculations. This process allows for precise mapping of player movements in videos to their actual positions on the field, facilitating comprehensive performance analysis. The project uses data from the DFL - Bundesliga Data Shootout Kaggle competition, with video frames annotated and transformed for training, leading to improved results using YOLOv8 without mosaic augmentations. The trained model's output is refined by filtering for higher confidence keypoints, and a homography matrix is employed to transform perspective, using a ViewTransformer class to map detected player positions to their real-world counterparts on the field. The approach demonstrates the potential of combining computer vision and sports analytics, with further applications and insights available in a dedicated sports repository.