Pedestrian Detection Technology Based On Machine Vision

With the development of the smart city,the development of pedestrian detection technology has attracted more and more attention.Pedestrian detection technology is an interdisciplinary research subject of image processing,pattern recognition,machine learning and so on.In addition,pedestrian detection in smart cars,security and other fields also has a broad application.At present,pedestrian detection technology has made great progress,but there are still many obstacles.For example,too much extracted region of interest,no enough features to describe the pedestrians,high feature dimension and low efficiency classifier.In order to solve the above problems,this paper focuses on the research of pedestrian detection technology from three aspects: interest region extraction,feature extraction and classifier.The main contents of this paper are as follows:(1)In the study of the region of interest,when using sliding window method in pedestrian detection,the detection windows are too much.In this paper,we propose a method to extract the region of interest,this method based on the constraint of road surface and image segmentation to reduce the number of windows to be detected.First of all,the FCM algorithm is used to segment the gray image after preprocessing,and then the connected region is labeled.On the other hand,region growing,edge detection and Hough transform are used to extract the road surface in the pedestrian detection image.According to the relationship between the pedestrian and the road surface,the extracted road is used to identify the connected regions.The connected region is the region of interest.In order to reduce the influence of illumination and highlight the pedestrian area,the histogram equalization method is added in the preprocessing.(2)For pedestrian feature extraction,single feature is not enough to describe the pedestrians;the fusion of features has high dimension and other issues.Therefore in order to tackle the problems of feature description and high feature dimensions,this paper presents a multi-scale fusion of sparse features to describe pedestrians.Firstly,the HOG features and LBP features of the pedestrian detection image are extracted.Next,HOG and LBP features are represented respectively by sparse representation,and then the two sparse features are fused by multi-scale.The HOG feature has 3780 dimensions.In the extraction of LBP features,the 64*128 samples are divided into blocks,then the LBP features are extracted,and the features of the blocks are normalized,and then the LBP features of the blocks are composed of the LBP features of the image.The experimental results show that the extracted LBP features are better when the block size is 8*16 pixels.In the process of sparse feature fusion,the pedestrian detection effect is compared under different fusion scales.Finally,this paper realizes the fusion of multi-scale sparse features.(3)In the research of classifier,the KNN algorithm and SVM algorithm are compared.The classification results of RBF kernel function and linear kernel function are compared.In order to further improve the SVM classification results based on RBF kernel function,this paper optimized two RBF parameters.The range of the optimal parameters is determined by the grid search algorithm,on the basis of this range,the particle swarm optimization(PSO)algorithm is used to optimize the parameters.Finally,we compare the detection results of the optimized RBF kernel function and the linear kernel function in SVM classification.