Nighttime pedestrian detection with a normal camera using SVM classifier

Abstract

Accidents occurring at night and involving pedestrians represent a significant percentage of the total. This paper presents an approach for pedestrian detection in nighttime with a normal camera using a SVM classifier. Objects in the video are extracted with an adaptive threshold segmentation method at first. In the recognition phase, a preliminary classifier is used to discard most candidates and a SVM classifier is used in detailed shape analyzing. At last, a tracking module is used to verify the classification result. This approach is more cost-efficient than the previous approaches which are based on expensive infrared cameras. Experimental results show that the proposed approach can detect 71.26% pedestrians and run in real-time. © Springer-Verlag Berlin Heidelberg 2005.