Research On The Algorithm Of Pedestrian Detection Based On Infrared Camera

At present,traffic accidents in nighttime occurs frequently,from which both the driver and the pedestrian suffer a lot.Compared with daytime,less vehicles travelling at night at higher speed.The effective lighting distance of the low beam headlight is limited,driver's view of the vehicle from the opposite direction can be easily affected by the full beam headlight.Pedestrian-vehicle accident may occur due to driver's carelessness.Infrared ray has long working distance and infrared images are not affected by the background light,and the pedestrian can be captured in the low light or the dark environment at some distance.Pedestrian detection technology based on the infrared camera in the driver-assistance system has broad application prospect.However,compared with the visible image,the infrared image has lower contrast,lower resolution and more noise.The complex and variable driving environment and the diverse pedestrian target make the research based on the infrared camera pedestrian detection algorithm more challenging.This paper combines itself with the National Natural Science Fund Project "Research on the control mechanism and evaluation method of the novel steering-by-wire system based on the driver's characteristics(Number: 51575223)and proposes infrared camera based pedestrian detection algorithm.By analyzing pedestrian detection in literature both at home and abroad,summarizing the problems existing in extraction of regions of interest and pedestrian detection algorithm,aiming at accurate and multi-target extraction of regions of interest and local feature extraction and rapid classification,the algorithm of pedestrian detection based on infrared camera is proposed in this paper.The main contents of this paper is summarized as followed:(1)Considering the defects of infrared imaging,the infrared images collected in front of the vehicle are preliminarily pretreated.The grayscale stretching is adopted to enhance image gray scale,the random noise in the image is filtered out by the bilateral filtering method,the OTSU method is used to segment the threshold,which effectively solves the problem of low contrast and noise of infrared image,and achieves the least degree of misclassification between the targets and the background.(2)By analyzing the advantages and disadvantages of different extraction of regions of interest algorithm in infrared image,the extraction of regions of interest algorithm based on connected domain analysis is used according to the accurate and multi-target requirement.In order to solve the problem that more candidate regions is separated by single connected domain analysis algorithm,the combination of the binary image morphological algorithm and the connected domain analysis algorithm is adopted in this paper.Then by limiting the extraction parameters of the rectangles,the multi-target regions of Interest are extracted.(3)An improved HOG feature is proposed.Because of the HOG cell's repeated calculation and lower non-pedestrian detection rate,the improved HOG feature redefines the computational weight of the gradient direction of the samples,and increase leg details of pedestrian weight coefficient.Based on the OTCBVS dataset and the self-collected infrared image dataset,a pedestrian and non-pedestrian sample database are established.The SVM is used to train and classify pedestrians.In case of missing frames and false check in continuous multi-frame detection,Kalman filter algorithm is used to track pedestrian.(4)The improved HOG features,the polynomial kernel function and the related parameters are set to the optimal selection of the algorithm by off-line simulation.The test system of infrared camera pedestrian detection algorithm is established,and the GUI is designed based on the test system.And the validity of the target tracking algorithm based on Kalman filter is verified.Real vehicle tests show that the improved HOG feature can effectively distinguish the distant pedestrian target.The proposed infrared camera pedestrian detection algorithm has better detection efficiency in the urban road and campus road scene.