Statistical Recognition of Motion Patterns

Abstract

One key prerequisite for a machine to interact intelligently with people is its ability to recognize humans as interaction partners and to understand their behaviors and the intentions and plans behind them. In this paper we are particularly interested in behaviors which are related to motion, e.g. intentionally moving towards a goal or obstructing somebody’s path. More specifically, we address the problem of detecting and tracking people in natural dynamic environments in real-time and extracting and classifying typical motion patterns. Robust detection and tracking of non-rigid objects in natural changing environments with a moving observer is a challenging problem. Though various sophisticated approaches have been presented, robustness with respect to changing light conditions, non-rigidness of the tracked objects, occlusion and motion of the observer is still a problem which quests for an effective solution. Recognizing and understanding the intention of motion would leverage applications in areas such as smart environments, human-machine interaction and video surveillance. We present an approach for tracking people by integrating range information and visual information and using their complimentary strengths to overcome typical tracking problems. Individual motion trajectories are extracted from these fused sensor data and are condensed in compact statistical spatiotemporal models. These models serve as a basis for the extraction of typical motion patterns by unsupervised learning.