Detection of unnatural movement using epitomic analysis

Epitomic analysis, a recent statistical approach to form a generative model, has been applied to image, video and audio processing applications. We apply the epitomic analysis to motion capture data and define it as a motion epitome, a probabilistic model representing a finite set of primitive movements which retain various lengths of local dynamics. We review the generation, inference and learning procedures of an epitome, adapt them for motion capture data and utilize the epitomic analysis to detect unnatural movements given only positive (natural) training data. We introduce a multi-resolution of motion epitomes as well as a fullbody and an ensemble of epitomes, then present experimental results and compare the performance with other conventional classification methods, including Hidden Markov Models and Switching Linear Dynamic Systems.