Segmentation of people in motion

Abstract

A method for segmenting monocular images of people in motion from a cinematic sequence of frames is described. This method is based on image intensities, motion, and an object model-i.e., a model of the image of a person in motion. Though each part of a person may move in different directions at any instant, the time averaged motion of all parts must converge to a global average value over a few seconds. People in an image may be occluded by other people, and usually it is not easy to detect their boundaries. These boundaries can be detected with motion information if they move in different directions, even if there are almost no apparent differences among object intensities or colors. Each image of a person in a scene usually can be divided into several parts, each with distinct intensities or colors. The parts of a person can be merged into a single group by an iterative merging algorithm based on the object model and the motion information because the parts move coherently. This merging is analogous to the property of perceptual grouping in human visual perception of motion. Experiments based on a sequence of complex real scenes produced results that are supportive of the authors approach to the segmentation of people in motion. >