A pedestrian detection system based on binocular stereo

Pedestrian detection in dynamic and outdoor environment is considered to be a challenging task due to complicated scenes and human´s various appearances. To achieve more efficient and precise detection, stereo vision is an attractive approach. This paper introduces a pedestrian detection system which makes use of binocular stereo. A new method of ROIs (Regions of Interest) extraction is proposed and the Latent SVM human classifier is employed to make the final decision. The ROIs extraction method uses two-stage segmentation, together with a minimal object scale acquirement approach based on region growing. And new strategies are adopted in the decision-making process to reduce detection errors. Experimental works show that with our approach, the detection speed increases effectively at the same detection rate, and the false alarm rate is substantially reduced.