Research On Ranging Tracking And Simulating Of Self-driving Vehicle Based On Binocular Vision

As a future mode of transportation,self-driving vehicle attracts the tireless research of universities and scientific research institutions at home and abroad.As a combination of artificial intelligence,visual computing,automatic control and software development,self-driving vehicle can significantly reduce the incidence of traffic accidents,reduce traffic congestion,improve traffic operation efficiency and provide convenience for national travel.Self-driving vehicle involves multidisciplinary knowledge,and mainly studies from four aspects: environment perception,path planning,positioning navigation and decision-making control.As part of the environment information received by self-driving vehicle,environment perception plays an extremely important role.Therefore,how to effectively and correctly identify the surroundings of the vehicle is the most important thing to realize unmanned driving.On the basis of summarizing the existing achievements,this paper uses binocular vision to collect image information,and studies vehicle and pedestrian ranging,tracking and simulation of self-driving vehicle in road environment.The main research includes the following aspects:1)Based on the analysis of the technology status of self-driving vehicle in recent years and the research achievements at home and abroad,this paper explores the research technology of target detection and tracking,points out the disadvantage of monocular vision in target detection and tracking,determines that binocular vision is the main research object,and establishes the key problems to be solved and the technical route of this subject.2)The vehicle detection algorithm with different spatial scales leads to the low efficiency and the monocular vision is difficult to accurately obtain the vehicle distance information,so a vehicle detection method based on improved Faster-RCNN to detect vehicle targets and binocular vision to range the vehicle is proposed.The binocular stereo camera is used to acquire the image information and perform preprocessing.Loading the training data of the deep neural network Faster-RCNN,and training vehicle target images using improved ACF vehicle detector,then,multiple built-in sub networks are introduced for different spatial scales of vehicle,and the output of all sub networks is adaptive combined to detect vehicle.Then SURF feature matching algorithm is used to carry out the stereo matching of the left and right images.Based on matching data and semantics segmentation of images,3D reconstruction and the centroid coordinate of the vehicle are determined to measure the distance between vehicle and binocular camera.3)Aiming at the poor accuracy of pedestrian detection algorithm and monocular vision is difficult to accurately obtain pedestrian distance information,feature fusion is proposed to improve HOG-SVM and using binocular vision to range the distance of pedestrians toachieve method of pedestrian detection and binocular vision ranging.The binocular stereo camera is used to acquire the image information and perform preprocessing.The HOG-SVM pedestrian detection algorithm is improved to increase pedestrian detection accuracy by integrating the feature.Then the SURF feature matching algorithm is used to match the left and right images,and the 3D reconstruction is used to determine the centroid coordinate of the pedestrian according to the matching data and semantics segmentation of images,so the distance between the pedestrian and binocular camera is measured.4)This topic proposes a stereo vision multiple target tracking method based on joint probabilistic data association filter for intelligent vehicle: the method collects the images and videos of vehicles and pedestrians with the help of stereo camera,visual stereo ranging algorithm is used to estimate the displacement of the vehicle in the middle,and the non-conforming objects are regarded as moving objects,and the measured values are sent to the tracking algorithm.Secondly,the state uncertainty representation and motion model are constructed based on the extended Kalman filter in Lie group.Finally,the accurate tracking of multiple target is realized by improved joint probability data association.5)SCANeR as a vehicle dynamics simulation software is used to build simulation experiments.The target vehicle is equipped with a binocular vision camera,and two working conditions,i.e.vehicle detection and tracking of self-driving vehicle and pedestrian detection and tracking of self-driving vehicle,are set up to simulate the operation conditions of the whole vehicle with binocular vision.The validity and rationality of the algorithm are analyzed and verified,and the driving stability and comfort are analyzed.