Summary
International Technical Conference on Circuits/Systems, Computers and Communications
2016
Session Number:P2
Session:
Number:P2-17
Relative Position Feature based Dense Trajectories with Density Adapted Noise Reduction for Tennis Player Action Recognition
Zihan Ma, Shuyi Huang, Masaaki Honda, Takeshi Ikenaga ,
pp.975-978
Publication Date:2016/7/10
Online ISSN:2188-5079
Summary:
Tennis player action recognition plays an important role in tennis sports analysis. The outdoor environment noise and high similarity features of arm motion at different torso sides are two crucial problems in tennis player action recognition. This paper proposes a density adapted noise reduction and a relative position feature to recognize player actions. The density adapted noise reduction adaptively removes sparse noise far away from the player region, according to the variable density of features in each frame. The relative position feature makes features at different sides of the player torso distinguishable to differ backhand receive action from serve and forehand receive actions. Experiments executed on sequences of a practical tennis game achieve accuracy higher than 93.8%, and AUC value larger than 0.94 for all action categories.