Probabilistic Color-Based Multi-Object Tracking with Application to Team Sports

This paper addresses the problem of tracking multiple non rigid objects --- such as humans --- in videos. Firstly, it aims at showing how tracking can be improved with the help of background analysis. Background color modeling is used to optimise the target discrimination, to detect specific situations such as occlusions or clutter, and to perform selective adaptation of the target model. Background motion due to camera pan, tilt and zoom is estimated and compensated in the tracking procedure. This tracking procedure is based on particle filtering, so that occlusions are also dealt with. Another key feature of the proposed tracking algorithm lies in its extension to multiple objects, possibly with similar appearances. The capacities of these new developments are demonstrated on sequences with both zooms and occlusions, including occlusions between similar objects. Applications to team sports are especially demonstrated