Biometrics Using Stereo Vision
The goal of this project is to extract appearance descriptors for people, based on whole-body appearance. A target application is to observe a person who appears in front of one surveillance camera (say at an entrance to an airport), then to reacquire that person if they reappear later in front of another camera (say at a check-in counter). This type of reacquisition is essential to analyze the movement of people through extended environments such as airports, factories and buildings.
Background & Objective: There is no doubt that faces are a key cue for reliable passive recognition of individuals. But good quality images of faces may not be available in many surveillance situations - the camera resolution may be low, or an elevated camera may be capturing only oblique views of the face if the subject's head is down. There is a range of other cues - including height, body shape, color of clothing - which are weaker than face recognition but which can be reliably captured using typical deployments for surveillance cameras (e.g. cameras above doorways, or in the top corner of a room). The goal of the project is to investigate the measurement and use of these cues, with a focus on 3D measurements.
Technical Discussion: The basic technology uses stereo for computing 3D data for a scene. There are two key components to the work (a) segmenting individuals in a group of people, and (b) characterizing the appearance of an individual. The work on (a) has focused so far on reliable detection of edge chains around an individual's silhouette (by using stereo disparity to link edges with similar depth) as a precursor to recognition of silhouettes. The work on (b) has been on measurements of height and body-shape, with initial work on gait-measurements such as stride length and sideways motion of the head during walking.
Technology Area: Computer Vision
Modification Date: June 13, 2008