Pedestrian Detection Using Statistical Structure Gradient Feature

In the past few years, research on pedestrian detection has gradually been hotissues in the field of computer vision. However, due to human factors such as bodyposture, clothing, camera angles and lighting, pedestrian detection has still been aworldwide problem. At present there are two mainstreaming methods: the first is touse global information of pedestrian samples, extract information belongs to theedge of pedestrians, and detect pedestrians by training a classifier or cascadedclassifiers; the second is to use local information of pedestrian samples, extractingvariety of pedestrian local information mostly expressed through machine learningalgorithm, training to be a cascade classifier for testing. Because of employing theIntegral image to speed up, the second way has a speed faster than the first method,but in lower accuracy. This article uses the first method, and a high accuracyalgorithm capable of accelerating has been proposed. Pedestrian detection based onclassifications, mainly consists of three processes: feature extraction and trainingalgorithm for classification, selection of test methods.This article details the predecessors have made major contributions to pedestriandetection algorithm, by Dalal, who in2005proposed the histogram of gradientvectors (HOG, Histogram of Oriented Gradients), which is a high precisiondescription operator for pedestrian. For HOG feature analysis to obtain pedestriansessential characteristics, and HOG gradient extraction to be simple, the paperpresents a natural gradient based on human vision to describe the edge of pedestrians.By Jianxin Wu on2011’s CENTRIST character describing the human body, gettingoutstanding ability as well as the HOG characterization, after an in-depth analysis oftwo kinds of features, this article adopts a method that combines gradient contourand texture characterization to describe pedestrians.On the choice of classifier, this article will use the bootstrap structure with supportvector machines (SVM, Support Vector Machine) as a training algorithms forpedestrian classification, effectively combination of image outline feature andtexture feature which belongs to the pedestrian, and training to obtain classifierswhich have the most distinguishing ability.For detection methods, the paper uses image zooming method, selecting a scalefactor s, downsampling image pyramids, if the classifier repeatedly hit the sametarget, this judgment determines that the region contains pedestrian, with multipledetection window fusion, and the final location of pedestrians be obtained.This article uses VS2005and OpenCV to achieve the part of pedestrian trainingand detection. Pedestrian detection algorithm based on Statistics Structure of Gradient Feature has been proposed. According to the experiment results, a newpedestrian detection algorithm, on the similar accuracy of N.Dalal’s HOG algorithmand calculation speed over HOG, is capable of exactly detecting human with strongrobustness in the complex context.