Real-Time Maintenance of Figure-Ground Segmentation

Peter Nordlund and Jan-Olof Eklundh

Abstract

An approach to figure-ground segmentation based on a 2-dimensional histogram in feature space is presented. The histogram is then analyzed with a peak-finding algorithm designed with real-time performance in mind. The most significant peaks in the histogram are backprojected to the image to produce an object mask. The method is applied to segmentation based on a 2-D color space and also on a combination of motion and stereo disparities. Experiments with a system grabbing images direct from a CCD-camera with real-time performance having typical frame-rates of about 10 Hz is presented.

Keywords: real-time, image motion, segmentation, cue-integration

Article in postscript, 3.3 Mb
Same article, but with more color images than originally, postscript, 3.9 Mb

Peter Nordlund
Last modified: Wed Jan 13 14:34:32 MET 1999