Normalized channel features for accurate pedestrian detection

Pedestrian detection is one of the most challenging problems in the field of the computer vision. To improve the detection accuracy several schemes have been proposed by many researchers but sufficient accuracy has not been achieved yet. One of the most accurate schemes is integral channel features proposed by Dollár et al., whose miss rate is about 20% at 10^-1 false positive per image for the INRIA dataset. In this paper, to improve the detection accuracy of the integral channel features, we propose a novel scheme that normalizes extracted features according to characteristics of them. The Experimental result using the INRIA dataset shows that the miss rate by the proposed scheme becomes about 15% at 10^-1 false positive per image that is higher than other leading edge schemes.