An efficient pedestrian detection method by using coarse-to-fine detection and color histogram similarity

Abstract

Pedestrian detection is an important issue in computer vision. In this paper, we focus on detecting pedestrians in images from automobile blackbox cameras. Our key contributions are from the observation that edge information can be used as coarse human detection by rejecting a large part of the background windows and that color cues are informative for representing humans. We propose a new coarse-to-fine human detection method in order to achieve efficient detection with high accuracy. Color information are represented by using color histogram similarity within each HOG block, which we refer to as CHS feature, then HOG and CHS are weighted and combined into a vector as a feature. Overall, our method yields 60% speedup over HOG-based sliding window approach, and furthermore reduces the error rate by 20%. The results show that our method is robust and accurate for pedestrian detection. © 2012 Springer-Verlag.