Obtaining features for the recognition of human movement style

Abstract

 

Human movement comes in many different styles and this research investigates features that can explain our ability to distinguish between styles. In particular, we studied how training to recognize the style of one individual can influence the ability to recognize movements performed by another individual. Tennis serves were used since they provide complex whole-body motions with the different styles of flat, slice and topspin. The observers used were not tennis players and pretest confirmed that they performed around chance at the identification before training. The service motions were presented using techniques of computer graphics that transformed 3D motion capture recordings of two professional tennis players into computer animations played on a common body model. Thus, removing all pictorial cues to identity. Two groups of 12 observers were trained on one or the other server's motions. After performing 6 training blocks on one server, performance at correctly identifying service type increased to approximately 60%. Observers were then required to identify the service style of the other server. What resulted was a consistent pattern of confusions for identification of the previously unseen server's movements. The pattern being those observers trained on Server A could not recognize the topspin serve of Server B and those trained on Server B could not recognize the slice serve of Server A. We then explored what information in the training phase could have caused this pattern of confusions. These explorations revealed that a linear discriminant classifier based on the first two principal components of the movement set showed a similar pattern of confusions. These data are consistent with the interpretation that a low-dimensional representation of the movement serves as the features used for recognition.