UC San Diego

Advances in visual computing have reached a point where systems can observe and interpret human activity. Learning to accomplish any human monitoring task needs a careful consideration of the level of observational detail that is required to understand the activity. In the general context of observing human interaction, much information can be obtained by looking at the head. The direction of a person's head gives a strong indication on what person or object they are focusing their attention. Since humans primarily attend to items of interest, estimating head pose can essentially be viewed as applying a saliency filter to determine the level of importance of each object in the field of view. This dissertation presents new computational algorithms and systems for detecting and tracking people using head location, orientation, and movement as the salient cue. Head movement is used to monitor the interaction between people in 3D environment, and experimental evaluations are presented for both monocular and distributed camera setups in intelligent automobiles and meeting spaces