Research And Implementation Of Pedestrian Detection Under Complex Background

Pedestrian detection is the application of computer vision processing technology in the video sequence to find the existence of pedestrians and to locate them accurately.With the increasing requirement from autonomous vehicles and intelligent monitoring,the demand on pedestrian detection technology,in terms of detecting speed and complex background adaption,is becoming higher.Consequently,pedestrian detection technology has become a hot theoretical research topic.Based on our analysis of related research results,this paper studies the design of pedestrian detection method that exploits Support Vector Machine(SVM)and Depth Convolution Neural Network(CNN).Specifically,we proposed three related pedestrian detection methods.The first method is on the basis of a combination of several SVMs.The second uses SVM and LeNet.The third uses a variant of CNN.To a certain extent,these methods are able to reduce the rate of missed and false detection in complex scene while improve the detection speed.The main contributions of this paper are as follows:(1)A pedestrian detection method based on combined support vector machine(CSVM)is implemented.In view of the problem of performance instability of a single SVM under complex background,a pedestrian detection method that combines HOG features with multiple SVM classifiers is proposed(referred as HOG-CSVM).Experimental results show that compared with HOG-SVM pedestrian detection algorithm,HOG-CSVM has lower false detection rate and missed detection rate,and thus more practical;(2)Implemented the SVM-LeNet pedestrian detection method.The key components of this method include: a)In order to enhance the ability to distinguish pedestrian characteristics,a combination of LUV color feature and HOG feature is used.b)To increase the sensitivity of the classifier to the small differences in pedestrian target characteristics,a method of using the SVM for the first stage classification and then using a three-layer CNN for the second stage classification is proposed.This method can effectively avoid the shortcoming that the deviation of single classification model can notbe corrected in time.c)To handle the problem of incorrect orientation when using the traditional greedy non-maximal suppression algorithm to locate the target,this paper proposes a fine-tuning non-maximum suppression algorithm.Experiments show that compared with NMS,after FTNMS fusion,the rectangle can more accurately filter out the real target;(3)A Deblur-DPDCNN pedestrian detection method is implemented.Firstly,a deblurring process is designed to make the target image under motion to be clearer.And then we use the selection search method(Selective Search)to extract the candidate area,so that to avoid the deficiency that the traditional multi-scale sliding window method showed in dealing with pedestrian attitude changes.Finally,the candidate regions were classified by a five-layer CNN.Experiments show that the Deblur-DPDCNN pedestrian detection algorithm has reached a level equivalent to Faster-RCNN.