Research On Navigation And Localization Technology Of Driverless Cars Based On Visual SLAM

Due to the rapid development of the Internet in these years,driverless cars have begun to move into People’s daily life.The prerequisite of driverless cars landing is safety.Therefore,the accuracy of the position estimation of driverless cars is particularly important.VSLAM focuses on reducing the error caused by the pose estimation.In this paper,the VSLAM system was studied on a self-designed drive-by-wire driverless cars,including the calibration of monocular camera,dense map construction of depth camera,front-end improvement of VSLAM system and sample vehicle experiment verification.The following studies were mainly done:(1)Combined with the existing technology and the requirements of this paper for the experimental platform,firstly,the parameters and 3d model of the driverless cars were designed;Then,the control system of driverless cars was developed,including steering system,electric drive system,hybrid braking system,et al;Finally,a complete vehicle platform was built and a series of performance tests were carried out on the vehicle.(2)Aim at the self-designed driverless cars,firstly,the motion model and observation model were used to model the VSLAM system;Then AORB SLAM2 system was designed in this paper,and the on-board monocular camera calibration experiment was carried,at the same time the improved particle swarm optimization algorithm was used to solve the camera internal parameters;Finally,the real-time 3d dense map reconstruction function was realized for the proposed A-ORB-SLAM2 system.(3)The A-ORB feature-based uniform distribution algorithm whose threshold value could adjust by itself was proposed for the problem that the feature points extracted by the conventional ORB algorithm were poorly distributed.First,in order to increase the extraction speed of the feature points,the fixed extraction threshold of the conventional ORB was changed to an adaptive threshold.Second,the image pyramid was used to add the scale information for the extracted feature points.Then characteristic points were extracted in the grid and in a certain order using the method of grid division to reduce computational redundancy.Finally,making use of the modified four-fork tree algorithm to filter out excess feature points.The simulation results showed that the proposed A-ORB algorithm in the paper could make the distribution of the extracted image feature points more uniformity and improved the navigation and positioning accuracy of the whole A-ORB-SLAM2 system.(4)A-ORB-SLAM2 system simulation and real vehicle verification experiment were carried out.Firstly,building communication system between the industrial control machine and the chassis which based on the existing conditions in the laboratory.Then simulation experiments were verified by an open-source image datasets and comparing and analyzing the absolute error during the process of position estimation.Finally,adopting driverless cars to carry on real car verification for the A-ORBSLAM2 system presented in this paper.The simulation and experimental results showed that the A-ORB-SLAM2 system designed in this paper had a better positioning and navigation accuracy.