Detecting People in Cluttered Indoor Scenes

Abstract

Motion is an important visual cue for scene analysis. It is particularly useful when the scene is cluttered, such as in typical home or office environments. We present a motion segmentation algorithm that makes use of temporal differencing to detect moving people in cluttered indoor scenes. The algorithm is devised based on a couple of perceptual organization principles. To deal with missing data, noise and outliers, a robust segmentation and grouping technique called tensor voting is employed. The resulting real-time people detector can handle the presence of multiple persons, and varying body sizes and poses. It requires no initialization, uses subjective threshold, which defines the minimum saliency of "significant" motion, and the only two parameters are the scales (sizes) of the local neighborhood for region and contour analysis.