Real time pedestrian detection using CENTRIST feature with distance estimation

Abstract

Pedestrian Detection (PD) is an active research area for improving road safety. Most of the existing PD system does not meet the demanded performance. This paper presents a working PD system which improves performance. The system uses CENTRIST feature extractor and the linear Support Vector Machine (SVM) for training and detection of pedestrian. CENTRIST is very easy to compute without any preprocessing and normalization that makes it suitable for on-board system. During the training procedure, we exhaustively searched for negative samples. Detection results on INRIA dataset are more accurate compared to benchmark method HOG. We used monocular camera to estimate pedestrian distance which is fairly accurate. We apply our detector on real-time video without region of interest (ROI) selection and could achieve 7 fps detection speed.