An enhanced Histogram of Oriented Gradient for pedestrian detection

Abstract

Significant researches have been carried out for pedestrian detection in images. The outstanding Histogram-of-Oriented-Gradients (HOG) feature proposed by Dalal and Triggs is the state-of-art for this task, and it is applied with a linear support vector machine (SVM) in a sliding-window framework. The novel method we proposed in this paper is based on this approach in which we add an enhanced feature to contain more feature information. Besides the same gradient information extraction process as HOG's, the enhanced feature extraction contains two steps: firstly, a new way is found to downscale the gradient image to its quarter size without losing much gradient information; secondly, `Circle HOG' features are extracted from those downscaled images. Then we combine the new enhanced features and the original HOG features together as an Enhanced HOG (EHOG) features. Our method is evaluated with a Histogram Intersection Kernel SVM (HIKSVM) on the public “INRIA” pedestrian detection benchmark dataset. The results show that proposed method consistently improves the detection rate by 4.5% in detection accuracy, compared with the original HOG.