Mask-EIoU: An EIoU based Feature-guided Association for Person Tracking in Sports Events

• Published in: 2024 International Conference on Consumer Electronics - Taiwan (ICCE-Taiwan) pp. 545 - 546
• Main Authors: Kao, Yu-Yung, Chou, Po-Yung, Lin, Cheng-Hung
• Format: Conference Proceeding
• Language: English
• Published: IEEE 09.07.2024
• Subjects: Cameras; Correlation; ex-pansionIoU; feature-guided association; multiple object tracking; object detection; Prediction algorithms; Real-time systems; Tracking; Trajectory; Visualization
• ISSN: 2575-8284
• DOI: 10.1109/ICCE-Taiwan62264.2024.10674305

Real-time tracking of athlete positions can be used to analyze team tactics in the broadcasting of sports events. However, since the camera adjusts with the position of the ball, the position of the athlete can vary greatly between successive frames. This creates challenges in tracking athlete locations, especially when multiple athletes are moving across the field at the same time. Tracking of multiple objects is typically based on trajectory prediction using Kalman filter, which relies on the assumption of linear motion. However, the movements of professional athletes are usually non-linear. Unlike conventional Kalman filter trajectory prediction, we utilize the Expansion-IoU technique to successfully address the challenges posed by camera movement. Additionally, emphasizing the importance of athlete appearance features in the matching process effectively overcomes the weakness of conventional tracking algorithms over-relying on athlete position. On the SportsMOT test set, the proposed approach demonstrates outstanding performance, achieving a HOTA score of 80.3% and an impressive MOTA score of 96.8%.