CurlObserver: A framework for curling trajectory detection and real-world mapping from broadcast videos with uncalibrated dynamic perspectives

Observation of trajectories in curling competitions is crucial for competition strategy optimization, athlete technique evaluation, and related scientific research. However, existing methods generally rely on visual sensors with static viewpoints and complicated camera calibration processes, resulting in high costs and insufficient generalizability. To this end, this paper proposes an innovative CurlObserver framework, achieving for the first time accurate detection and reconstruction of complete curling trajectories under real-world court coordinates from dynamic-view curling broadcast videos without camera calibration. The framework utilizes YOLO and the innovative Fusion-YOLOobb model to accurately detect the position and angle information of curling stones; combines the CurlSort algorithm, which integrating NSA Kalman filtering and the multidimensional matching strategy, to achieve stable trajectory tracking; and integrates an innovative homography matrix estimation algorithm based on feature point-line segment matching, effectively addressing the trajectory alignment and coordinate mapping problems under dynamic viewpoints. Moreover, this paper constructs a large-scale curling dataset based on broadcast videos of the Winter Olympics, covering object detection, orientation detection, object tracking, and semantic segmentation, comprehensively validating the performance of the proposed method. Experimental results demonstrate that the framework achieves an RMSE of 3.62 cm and detection accuracy of 98.7 %, significantly outperforming existing techniques. This research not only provides a new technical solution for curling trajectory detection with low cost and high precision but also offers technical references for precise trajectory extraction tasks in other dynamic-view videos, possessing broad application potential.