Robust tracking of athletes using multiple features of multiple views

Abstract

This paper presents a robust and reconfigurable object tracker that integrates multiple visual features from multiple views. The tandem modular architecture stepwise refines the estimate of trajectories of the objects in the world coordinates using many plug-ins that observe various features such as texture, color, region and motion in 2D images acquired by the cameras. One of the most important features of our proposed method is that each plug-in innovates the trajectories not only by back-projecting 2D observations of the features, but also by weighting them adaptively to their self-evaluated reliability. In the paper, the architecture of the system and that of the plug-ins are formulated. The behavior and robustness against occlusion are also shown through experiments with football-game sequences.