VASC Seminar - Visual surveillance -- movement modeling and event recognition

A very special VASC Seminar this Thursday, 1pm, 3305 NSH.

VASC Seminar Series

Larry Davis
U Maryland

Visual surveillance -- movement modeling and event recognition

This talk will cover two topics central to surveillance under
investigation at the University of Maryland:
1. Movement modeling of simple human hand and leg movements, and
2. Event modeling and recognition of human interactions.

We first consider the problem of recognizing human actions commonly observed during surveillance. The term actions refers to movements in which the subject has a reason and a viable execution plan. The focus is on actions with substantial limb movements and fairly simple structure - examples include reaching out, striking, etc. Computer vision research on recognizing movements has focused on appearance or position-based approaches. Commonly observed movements like reaching, striking, waving, etc., have highly variable target locations - a person can reach above
his/her head, for something on the floor, etc. Predictably, change in target location leads to change in the trajectory followed by the hands and other body parts during the movement. This puts appearance-based techniques at a disadvantage. Either a large number of training examples are needed or specialized models must be trained. There are factors common to reach movements that are independent of the target's location. Psychological studies indicate that one of these factors is the manner in which forces are applied to the hands during these movements. This
manifests itself in the velocity profiles of the hands during the movements. Our approach exploits constraints on the velocity profile to recognize common human movements like reach, strike, etc.

We then consider the problem of representing and recognizing interactions between people, places and objects that can occur over indeterminate periods of time. We describe an approach based on multi-valued default logic that augments purely appearance based analysis of people, vehicles and objects with common sense knowledge about possession, knowledge and closed worlds. Examples from multi-camera surveillance will be provided.