An attempt to pedestrian detection in depth images

We investigate pedestrian detection in depth images. Unlike pedestrian detection in intensity images, pedestrian detection in depth images can reduce the effect of complex background and illumination variation. We propose a new feature descriptor, Histogram of Depth Difference(HDD), for this task. The proposed HDD feature descriptor can describe the depth variance in a local region as Histogram of Oriented Gradients(HOG) describes local texture cues. To evaluate pedestrian detection in depth images, we also collected a large dataset, which contains not only depth images but also the synchronized intensity images. There are 4673 pedestrian samples in it. Our experimental results show that detecting pedestrians in depth images is feasible. We also fuse the HDD feature from depth images and HOG from intensity images. The fused feature gives an encouraging detection rate of 99.12% at FPPW=10^-4.