Real-Time Human Detection Based on Range Image Segmentation

Abstract

In this paper, a human detection method that combines information of segmented range image and human detectors based on image local feature is proposed. The method uses a stereo vision system called Subtraction Stereo, which extracts foreground information and range image. Range image is segmented for every object by Mean Shift Clustering. The proposed method reduces processing time and false detection by limiting the search range to the object. Joint HOG feature is used for human detection and reduces undetected human by integration of detection windows in consideration of occulusion. The proposed method is evaluated by experiments comparing with the method using the conventional Joint HOG feature.