A Unified Framework for Detection of Suspicious and Anomalous Beahvior from Spatio-Temporal Traces

The problem of learning behavior patterns from sensor data arises in many applications including smart environments, video surveillance, network analysis, human-robot interaction, and ambient assisted living. Our focus is on detecting behavior patterns that deviate from regular behaviors and might represent a security risk, health problem, or any other abnormal behavior contingency. In other words, deviant behavior is a data pattern that either does not conform to the expected behavior (anomalous behavior) or matches previously defined unwanted behavior (suspicious behavior). Deviant behavior patterns are also referred to as outliers, exceptions, peculiarities, surprise, misuse, etc. Such patterns occur relatively infrequently; however, when they do occur, their consequences can be quite dramatic, and often negative. We targets a large class of problems with complex, spatio-temporal, sequential data generated by an entity capable of physical motion in environment, regardless of whether the observed entity is human, software agent, or even robot. In such domains, an agent often has an observable spatio-temporal structure, defined by the physical positions relative to static landmarks and other agents in environment. We suggest that this structure, along with temporal dependencies and patterns of sequentially executed actions, can be exploited to perform deviant behavior detection on traces of agent activities over time.