Subsampling-based HOG for Multi-scale real-time Pedestrian Detection

Pedestrian Detection is a key problem in computer vision research. And the feature of Histogram of Oriented Gridients(HOG) is commonly used in this field to describe image local texture. In this paper, the method about a background-based model and a multi-scale sliding window subsampling HOG feature( D-HOG) is proposed, which can be used for pedestrian detection. First, generating a foreground image which is regarded as the input image. Then, a plurality of target regions are obtained on the foreground image by using a multi-scale sliding window. Next, extracting the D-HOG features of each obtained target region. Finally, the detection phase is performed by a support vector machine(SVM). The experimental results demonstrate that the D-HOG feature vector with multi-scale sensitivity can maintain the accuracy of HOG and improve real-time performance, which is suitable for multi-scale and different proportions of pedestrian detection.