Human behavior understanding in image sequences requires to study human interaction with the environment, because human beings behave depending on their location. Therefore, a semantic labeling of the scenario must be performed and provided to the system. In this work we present a method to automatically extract semantic information from a scenario by means of a set of existing agent trajectories and an ontology of the possible semantic regions. Our training algorithm constructs a path database which serves to infer a conceptual model of the scenario and to predict future agent trajectories, even from different camera views.

Peer Reviewed