Real-time 3D Soccer Ball Tracking from Multiple Cameras

A novel model for 3D soccer ball tracking is proposed, using multiple fixed cameras as input. The main problems are: successfully filtering false alarms, tracking through missing observations and estimating 3D positions from single or multiple camera inputs. The key innovation is to categorise all motion of the ball into four possible phases – rolling, flying, in possession, and out of play, and in different phases the ball trajectory is further modelled as linear or curve segments. Then, triangulations and phase-specific models are employed to estimate 3D ball positions. The system accuracy is evaluated by comparing the estimated ball phases and positions with manual groundtruth, over several minutes of real data sequences. Unlike existing systems using shadows and manual assistance, our approach demonstrates fully automatic ball tracking with the potential for accurate and robust results.