Recognition of body and facial gestures
Ioannis Patras (in cooperation with M. Pantic and A. Oikonomopoulus)
University of York, Computer Vision and Pattern Recognition Group


Recent years have witnessed an unprecedented interest in the visual analysis of human motion that is stimulated by applications that include human computer interaction, surveillance/security, multimedia retrieval, computer games, animated films and entertainment. In this talk we will present our recent work in human action recognition that proposes a sparse representation of an image sequence as a collection of spatiotemporal events. These events are localized at spatiotemporal 'interesting' points that correspond to peaks of activity variation and are detected by measuring the variation in the information content along the space and time dimensions. In this way redundant information such as static backgrounds and uniform clothing, as well as non-dominant motion due to noise or shadows is filtered out. Subsequently, we introduce an appropriate distance metric between two collections of spatiotemporal interesting points that allows learning and classification. We show results on real image sequences from a small database depicting people performing aerobic exercises and we discuss strengths and limitations of this approach as well as future work. Time permitting, will give a short overview of our work in simultaneous tracking of multiple interacting targets and its application in tracking multiple facial features. The work proposes a computationally efficient way of incorporating static/dynamic shape constraints within the particle filtering tracking framework. For facial feature tracking, the shape constraints model in a statistical way domain specific knowledge about the facial anatomy and facial deformations (e.g. due to facial expressions). We demonstrate experimental results in facial image sequences and briefly discuss applications.