A Multilevel Body Motion-Based Human Activity Analysis Methodology

Human body motion analysis is an initial procedure for understanding and perceiving human activities. A multilevel approach is proposed here for automatic human activity and social role identification. Different topics contribute to the development of the proposed approach, such as feature extraction, body motion description, and probabilistic modeling, all combined in a multilevel framework. The approach uses 3-D data extracted from a motion capture device. A Bayesian network technique is used to implement the framework. A mid-level body motion descriptor, using the Laban movement analysis system, is the core of the proposed framework. The mid-level descriptor links low-level features to higher levels of human activities, by providing a set of proper human motion-based features. This paper proposes a general framework which is applied in automatic estimation of human activities and behaviors, by exploring dependencies among different levels of feature spaces of the body movement.