Automatic rally detection on broadcast tennis videos

Abstract

The aim of this study was to propose a novel method of automatic rally detection on broadcast tennis videos. This method can be useful to predict and estimate the main physical demands in tennis matches. The method developed in Matlab® (Mathworks, MA, USA) was based on histogram extraction, percentile statistical filtering, and keyframe histogram spatiogram similarity. Three sample tennis videos representing three different playing surfaces were used to evaluate the proposed method. Maximal detection ratio was reached in hard-core playing surface (91%) followed by grass (82%) and clay (71%). The average of automatic tennis rallies detection was 81%. The reasons of rally detection errors are related to the camera motion, court illumination changes, smooth frame transitions, and the false high similarity between a non-rally frame and the keyframe histograms. The rally duration obtained with this method was valid when compared with manual annotation. In conclusion, the proposed method has a good performance, and estimation of its rally duration was comparable with the manual detection method (ground true) for broadcast tennis videos.

Keywords:

• sports video
• scene cut detection
• tennis playing surfaces

Acknowledgements

This research was supported by FAPESP (2011/03928-9) and CNPq (473729/2008-3 and 304975/2009-5).