Leveraging temporal, contextual and ordering constraints for recognizing complex activities in video

Abstract

A system (and a method) are disclosed for recognizing and representing activities in a video sequence. The system includes an activity dynamic Bayesian network (ADBN), an object/action dictionary, an activity inference engine and a state output unit. The activity dynamic Bayesian network encodes the prior information of a selected activity domain. The prior information of the selected activity domain describes the ordering, temporal constraints and contextual cues among the expected actions. The object/action dictionary detects activities in each frame of the input video stream, represents the activities hierarchically, and generates an estimated observation probability for each detected action. The activity inference engine estimates a likely activity state for each frame based on the evidence provided by the object/action dictionary and the ADBN. The state output unit outputs the likely activity state generated by the activity inference engine.