Real-time Tracking of Multiple People Using Stereo

David Beyme
Artificial Intelligence Center
SRI International
[Joint work with Kurt Konolige]


A fairly standard approach for object tracking uses correlation with a gray-level template of the object that is recursively updated. Two difficulties with this approach are object initialization and slow template drift off the object. We are investigating how adding stereo input can address these difficulties in a person tracking framework. First, background subtraction on stereo disparities -- which are less sensitive than intensities to changes in lighting -- is used to focus on foreground objects. People are detected in foreground disparity layers by using correlation with head and torso templates. These same templates are used again in the tracking module to re-center the person tracks and avoid template drift. The entire system can track multiple people in crowded scenes and with large changes in scale. It runs at around 10 Hz on standard PC hardware. The system has been evaluated on a number of video clips, some of which have a number of difficult occlusion events.


David Beymer is a computer scientist at the Artificial Intelligence Center of SRI International. He received his Ph.D. in Computer Science from MIT in 1995, worked in the CS Division of UC Berkeley the following year as a postdoctoral researcher, and recently spent two years with the Vision Technology Center of Autodesk. His research interests are in computer vision and learning, especially face and body tracking and face recognition.

Eyal Amir
Last modified: Fri May 14 14:07:27 PDT 1999