A general framework for 3d soccer ball estimation and tracking

Abstract

A general framework for automatic 3D soccer ball estimation and tracking from multiple image sequences is proposed. Firstly, the ball trajectory is modelled as planarcurves in consecutive virtual vertical planes. These planes are then determined by two ball positions with accurately estimated height, namely critical points, which are extracted by curvature thresholding and nearest distance of 3D lines from single or multiple views respectively. Finally, unreliable or missing ball observations are recovered using geometric constraints and polynomial interpolation. Experiments on video sequences from different cameras, with over 5000 frames each, have demonstrated a comprehensive solution for accurate and robust 3D ball estimation and tracking, with over 90% ball estimations within 2.5 metres of manually derived ground truth.