Improving Gradient Histogram Based Descriptors for Pedestrian Detection in Datasets with Large Variations

Gradient histogram based descriptors, that are constructed using the gradient magnitudes as votes to orientation bins, are successfully used for Pedestrian Detection. However, their performance is hampered when presented with datasets having many variations in properties such as appearance, texture, scale, background, and object pose. Such variations can be reduced by smoothing the images. But, the performance of the descriptors, and their classifiers is affected negatively by this, due to the loss of important gradients along with the noisy ones. In this work, we show that the ranks of gradient magnitudes stay resilient to such a smoothing. We show that a combination of image smoothing and the ranks of gradient magnitudes yields good detection performances, especially when the variations in a dataset are large or the number of training samples is less. Experiments on the challenging Caltech and Daimler Pedestrian datasets, and the Inria Person dataset illustrate these findings.