Tennis stroke detection and classification based on boosted activity detectors and particle filtering

Abstract

A new method for tennis stroke detection and classification is proposed. The proposed method consists of two parts: stroke detectors and a classifier. Our stroke detectors are based on a combination of Efros et al.'s motion descriptor and Ke et al.'s event detector. Since the stroke detectors work individually, post-processing is required to merge individual detection results when multiple stroke detectors are applied. The proposed classifier is based on particle filtering and it merges the multiple detection results by introducing the game structure of tennis. The proposed method is applied to off-air tennis video data, and the performance is evaluated by comparison to Efros et al.'s method. The experimental results showed that our proposed method outperformed Efros et al.'s method in terms of the recognition rate.