A New Feature for Night-time Pedestrian Detection

Abstract

The Histograms of Oriented Gradients (HOG) feature is widely used for pedestrian detection because of its high performance in accuracy, but the information it describes still has room for improvement. In order to increase the accuracy of pedestrian detection, an Improved Histograms of Oriented Gradients (IHOG) feature is proposed at the basis of HOG feature. After computation of HOG feature, an improved feature is added to the HOG feature. It applies an 8 8-pixel block with 4-pixel strides to obtain nine images of 15 31 pixels, each image representing one direction range (BIN), in which all blocks’ vectors within the same BIN will be mapped respectively in pixels. By using tri-linear interpolation in these nine images and using linear SVM to train all these feature vectors, an I-HOG feature is obtained. I-HOG contains a larger range of information than HOG and is tested on night-time datasets. The experiment result shows that, by using I-HOG, the detection rate could increase by about 4.26% than when using HOG.