Action Recognition in Tennis Videos using Optical Flow and Conditional Random Fields

Abstract

The aim of Action Recognition is the automated analysis and interpretation of events in video sequences. As result of the applications that can be developed, and the widespread availability and popularization of digital video (security cameras, monitoring, social networks, among many other), this area is currently the focus of a strong and wide research interest in various domains such as video security, humancomputer interaction, patient monitoring and video retrieval, among others. Our long-term goal is to develop automatic action identification in video sequences using Conditional Random Fields (CRFs). In this work we focus, as a case of study, in the identification of a limited set of tennis shots during tennis matches. Three challenges have been addressed: player tracking, player movements representation and action recognition. Video processing techniques are used to generate textual tags in specific frames, and then the CRFs are used as a classifier to recognise the actions performed in those frames. The preliminary results appear to be quite promising.