Tennis video analysis based on transformed motion vectors

Abstract

Motion Vectors (MV) indicate the motion characteristics between two video frames, and has been widely used in the contentbased sports video analysis. Previous works on sports video analysis have proved the effectiveness and efficiency of the MV-based methods. However, in the tennis video, the MV-based methods are seldom applied because the motion represented by MV is greatly deformed relative to the player's true movement due to the camera's diagonal shooting. In this paper, an algorithm of MV transformation is proposed to revise the deformed MV using a pinhole camera model. With the transformed MVs, we generate the temporal feature curves and employ Hidden Markov Models to classify two types of player's basic actions. Evaluation on four hours live tennis videos shows very encouraging results. © Springer-Verlag 2004.