Continuous Action Recognition by Action-specific Motion Models

This paper proposes the models of human motion prior with multiple actions for action recognition in videos. A training sequence of each action, such as walking and jogging, is separately recorded by a motion capture system and modeled independently. Unlike existing approaches with similar motion prior models, our method uses the multiple models simultaneously for particle filtering in order to track the pose of a target person without being interfered by ambiguous motion. In addition to robustness to ambiguous motion, the particle transition among multiple motion models allows us to continuously identify the action of the target person frame-by-frame.