Pedestrian Detection Algorithm With Co-occurrence Relationship And Adaptive HCS-LBP Features

Pedestrian detection means using computer vision algorithm to automatically judge whether there are pedestrian targets in the detected image or video sequence,and return the spatial coordinate information.As a special group,pedestrians have the universality of general goals,and the variety of special features in the class,and there is no fixed shape and space structure.This paper first analyzes the characteristics of the existing pedestrian detection algorithm,then summarizes the difficulties in the process of pedestrian detection algorithm,finally proposed two kinds algorithm of pedestrian detection for improve the utilization of image information,enhance the pedestrian feature description ability,the specific contents are as follows:(1)The development and research status of pedestrian detection algorithms are briefly introduced;The advantages and disadvantages of the algorithm based on video sequence(frame difference,background subtraction,optical flow)and image based algorithm(template matching,statistical learning)are elaborated and analyzed in detail,and the deep neural network models are compared;The important and difficult points in the research of pedestrian detection algorithms are summarized and summarized;The theoretical knowledge related to the detection algorithm is combed,and the theoretical basis is rammed for the follow-up study of this paper.(2)In order to improve the robustness of pedestrian detection algorithms to interference factors such as illumination change,photographing angle of view,overlap of pedestrians,or object occlusion,we proposed a pedestrian detection algorithm based on co-occurrence relationship and matrix-structural cascaded classifier.The algorithm first proposes the Co-occurrence Local Quantization Code(CoLQC)to improve the utilization rate of image local structural information,so as to enhance descriptive power of image feature;then,we improve the Co-occurrence Histogram of Oriented Gradient,and propose an Improved Co-occurrence Histograms of Oriented Gradient(ICoHOG),which bring multi-dimensional information in weights calculate to reduce image information loss;finally,CoLQC and ICoHOG are respectively used to train the front M layer and the latter N layer of the Matrix-Structural Cascaded Classifier to increase the sample complexity and improve the detection performance of the classifier.The experimental results of the ETH pedestrian database and the Daimler pedestrian database shows that The algorithm can enhance the utilization of structural information among neighboring pixels,enhance the description ability of image features.(3)In order to solve the problem when extract Center-Symmetric Local Binary Pattern(CSLBP)in pedestrian detection that the central pixel in the neighborhood is not involved in the calculation,the artificial judgment threshold is subjective,fail to differentiate between different sub blocks,we proposed a pedestrian detection algorithm based on Adaptive HCS-LBP(Haar-like CSLBP)feature.The algorithm first constructs a HCS-LBP feature coding method,uses local central symmetric mode to reduce the encoding length,integral image method is used to quickly calculate the time complexity,and the global adaptive threshold is calculated by introducing the gray level probability into the central pixel,the local adaptive threshold of neighborhood pixels is obtained by using the Gauss matrix to highlight the objective texture information of the image;Then the center pixel is involved in the image coding,and the weight of different sub blocks is determined through the information entropy,so as to enhance the description ability of the image features;Finally,Histogram Intersection Kernel SVM(HIKSVM)is used to train the samples to improve the accuracy of the classifier.