Research On Visible Light And Thermal Infrared Imaging Pedestrian Detection Method For Aided Safe Driving System

Pedestrian detection is a hot research topic in the field of computer vision.Multispectral pedestrian detection is Fascinating problem to researchers.It has a wide range of applications and development prospects in assisted driving,intelligent monitoring,robot and other fields.Multispectral pedestrian detection is combine visible images and thermal infrared images,the purpose is to aggregate the characteristics of visible light and thermal infrared images at the same time,improve the detection accuracy of pedestrian detectors and reduce the missed detection rate.The visible light pedestrian has the problem of poor image imaging in dim light,rain and snow,and the thermal infrared image pedestrian detector has the problems of less texture details,low signal-to-noise ratio,and large influence by background information.Multispectral pedestrian detection In theory can solve these problems.This paper is mainly based on the deep learning target detection algorithm which named Faster-RCNN,combined with the KAIST multispectral pedestrian data set,to study the application of multispectral pedestrian detection in assisted driving.The main work includes:(1)Aiming at the problem of blurred pedestrian contours in infrared images and low contrast between pedestrians and the surrounding environment,experiments have compared a variety of image enhancement and sharpening methods,such as high-pass filtering,low-pass filtering,Sobel operator,Laplacian operator,Scharr operator,etc.The enhancement algorithm enhances the thermal infrared image.(2)Aiming at the problem that the size and position of the detection frame in visible light and thermal infrared image pedestrian detection cannot cover pedestrians well,the optimization problem of the Loss function is studied,and GIOU is used to replace the original IOU to calculate the pedestrian detection frame and marking frame.The overlap rate.(3)Aiming at the problem that the Faster-RCNN deep learning target detection algorithm is not accurate in pedestrian detection,introduce 1*1 convolution,cavity convolution and other structures in the feature extraction network to enhance the expression of the network feature layer,and use the ROIAlign method instead The ROIPooling method maps the candidate boxes to the feature layer.(4)Aiming at the problem of single visible light image and single thermal infrared image feature extraction network extracting pedestrian information,a pedestrian feature extraction network fusing visible light and thermal infrared images is designed,which is divided into pixel fusion,early fusion,mid-term fusion,and late fusion.This paper studies the visible light and thermal infrared images in pedestrian detection for assisted driving from the aspects of fusion feature network design,image enhancement,loss function optimization,and feature extraction network improvement,and designs 6 sets of experiments to verify visible light and thermal infrared images There is information complementarity in pedestrian detection.Among them,the miss rate and detection accuracy of the pixel fusion pedestrian detector are better than other pedestrian detectors.The AP is 8.38% higher than that of the traditional ACF+T+THOG pedestrian detector,and the miss rate is 5.34% lower than that of the visible light pedestrian detector.The multi-spectral pedestrian detector has good detection effect and robustness,and has certain reference significance for the subsequent research of multispectral pedestrian detection.