Histograms of Salience for Pedestrian Detection
Yuanyuan Nan, Chao Liang, Wenhua Fang, Xiao Wang, Yuhong Yang, Jun Chen · 2014
Pedestrian detection has been seen huge progress in recent years, much thanks to the Histograms of Oriented Gradients (HOG) features. However, this method (HOG and SVM) has a large number of false detections. To conquer the problem, we provide an affirmative answer by proposing and investigating a salience representation for pedestrian detection, Histograms-Of-Salience (HOS). We extracted saliency map learned from data by using Histogram Based Contrast, and aggregate salient value and oriented gradients to form local HOS. We intentionally keep true to the sliding window framework and only change the underlying features. By learning and using local HOS feature that are much more expressive than HOG, we demonstrate large improvements on the public INRIA dataset.