Physics based 3D ball tracking for tennis videos

Professional sport is a highly competitive field, and training tools are playing an increasingly important role. Recent improvements in sensor technology allow managers to monitor many new factors during training, but analyzing this data is time consuming, thus the need for specialized signal processing techniques. This paper presents our work on high level data extraction from tennis multi-view video. A full video indexing platform has been developed involving an innovative physics-based 3D ball tracking algorithm. This paper focuses on the tracking algorithm, which makes accurate ball tracking possible despite sparse detections and numerous false alarms, and gives access to very interesting parameters for coaching purposes such as bounce location and approximative ball spin. Our tool has been tested by a professional tennis coaching team resulting in very positive feedback.