Human home behavior interpretation from video streams

This paper presents a system for human home behavior analysis from video streams using activity triggered state machine combined with contextual information. The transition function of the state machine is embedded with the human activities associated with human behaviors. When a sequence of recognized activities triggers the state machine into a behavior-meaningful state, the system then combines the contextual information for drawing the high level interpretation. In activity recognition, the system first detects and tracks the moving objects, computes the human postures by these features, and then takes the postures for activity recognition. To conquer the problem of frame start point selection associated with activity recognition, this paper also proposes to use posture change as a basis to locate the start frame.