Stereo-based human detection and tracking for crowd monitoring

Abstract

In this paper, novel stereo-based methods for detecting and tracking human objects in crowds are proposed. The method for detecting human heads from a disparity image contains three distinctive steps. In the first step, Object-Oriented Scale-Adaptive Filtering is proposed to extract the evidence of human heads with the most suitable scales. In the second step, a 3D virtual plane parallel and over the ground surface with the average height of human beings is built to filter out spurious evidence of human heads. Finally, a mean-shift algorithm is applied to locate human heads on the evidence map. The detected human heads are tracked by kernel-based feature evaluation, which adaptively fuses motion, color and stereo information. To our knowledge, this research is the first attempt to the tough problem of human individual detection and tracking in crowded scene based on stereo. Good results have been achieved on the test sequences from real scenes.