Gradient And Multiple Classifiers Based Pedestrian Detection Method

As a branch of computer vision, the pedestrian detection technology hasalready attracted more and more attention in recent years. The main purpose ofpedestrian detection is to distinguish pedestrians and background in pictures orvideos, and to determine the location of the pedestrians in the image. The traditionalpedestrian detection method refers to the extraction of size, shape, texture, color orappearance information of pedestrians, and to determine whether there arepedestrians in the image according to some prior knowledge. As the traditionalpedestrian detection method used a single and unstable feature, the discriminantability is not very high. With the introduction of gradient feature and multi-classifierinto pedestrian detection, the classification accuracy and rate have been enhanced.This dissertation mainly focuses on the study of the gradient feature and multiclassifiers based pedestrian detection method. This dissertation includes thefollowing parts:The first part is extraction of gradient feature and multi feature fusion. Thecommonly used features of pedestrian detection are Edgelet, Shapele and Haar.Considering these features are sensitive to the influence of the different postures andthe diversity between pedestrians, this paper adopts gradient features of pedestriansat feature extraction stage. At the same time, considering the different size anddistances, this paper extracts the gradient feature of the Pyramid image. Finally,considering the sidedness of single feature, pyramid histogram of gradient featureand pyramid center symmetric Local Binary Pattern are combined to set up the multifeature.The second part is Bagging based pedestrian detection method. Baggingmethod can significantly improve the classification performance of unstablelearning algorithm, and it is very suitable for parallel operation which can greatlyreduce the training time of multi classifiers. At the decision stage, the basicclassifiers are KMSE and SVM. Considering the speed and scale of classifiersintegration, a pruning method in Bagging is proposed in this paper. Compared withthe original Bagging method, there is a promotion in pruning method.The last part is posterior probability based pedestrian detection method. Thecommonly used two-class classifiers often use a simple threshold to represent thefinal classification result, which may bring great interference in the later multiclassifier fusion stage. By using the posterior probability method, the sample can besent to the other classifiers for further judgment if a classifier cannot decide the source label of a boundary point. At the same time, the classifier should providesome decision information as the classification ground of the multiple classifierfusion, which achieved good classification results.