Research On Pedestrian Detection Method Based On Multiple Feature Cascade

With the developments of artificial intelligence, computer vision, intelligent monitoring technology and intelligent transportation in modern society, Pedestrian Detection technology has been received more and more attention and widely used in the field of intelligent monitoring technology, intelligent transportation, advanced human-computer interaction and so on.This thesis mainly focuses on pedestrian detection method based on HOG features in-depth research, and the corresponding improved method is proposed on the basis of the above method: reducing HOG feature dimension using principal component analysis(PCA) and pedestrian detection based on multi-features cascade which cascades the HOG feature after dimension reduction with the Gabor feature and the color feature. Research results are shown as follows:1. The existing pedestrian detection methods based on histogram of oriented gridients(HOG) have the problems of relatively high dimension and a lot of redundant information which could reduce classification speed, some dimensional information may reduce the accuracy of recognition. This thesis proposed a pedestrian detection method of HOG feature using PCA dimension reduction and 14 experiments were done for HOG feature from the dimension 20 to 3000 and the appropriate dimension of HOG feature was selected for ensuring both the classification accuracy and the time efficiency. Furthermore, comparisons were made for the DET curve and the runing time in our proposed method and original method respectively.2. Although the HOG feature have a good performance in describing pedestrian features, it is insufficient to ensure the detection rate by simply using a single feature. The proposed method for pedestrian detection based on multi-features cascade which cascades the HOG feature after dimension reduction with the Gabor feature and the color feature(RGB, HSI). Detail elaborations for these features, two-dimension Gabor value at 4 scales and 8 directions were selected as the Gabor feature. Among the color feature, only the spatial feature value, mean and standard deviation were used. Finally radial basis kernel function(RBF) of SVM is used to classify.INRIA dataset is used as the classifier to train library and the test library. Experimental results of this proposed method demonstrated that it not only improve the speed of classification, but also improve the detection accuracy. While the proposed method received a better detection in the case of only one pedestrian, but it occurred some miss or false detection for multiple pedestrians or other more complex scenarios.