Research On Vehicle Localization Algorithm Based On Fusion Of Stereo Vision And Inertial Navigation

With the development of artificial intelligence technology,intelligent driving technology has become the focus of current research.The localization of intelligent vehicle is one of the key technologies to realize intelligent driving,and it is a prerequisite for vehicle path planning and control.However,traditional GPS positioning cannot meet the positioning requirements in scenarios such as vehicles passing through tunnels and underground parking lots.Therefore,Simultaneous Localization and Mapping(SLAM)technology has become the focus of vehicle localization.In this paper,for scenes with weak GPS signals such as tunnels and underground parking lots,low-cost cameras and inertial measurement units are used as sensors to build a SLAM system that integrates vision and inertial navigation data to achieve vehicle positioning.The main research contents of this paper are as follows:1.In this paper,the vehicle motion is described mathematically,and the vehicle location problem is transformed into a state estimation problem.Compare the advantages and disadvantages of various cameras,select the stereo camera suitable for the application scene of this paper,and analyze the stereo camera model.On the premise that the stereo camera can calculate the depth information,the principle of the visual odometry method is studied,and the advantages and disadvantages of each algorithm in different scenarios are analyzed.2.In view of the error in the sensor measurement of the SLAM system,the sensor parameters were calibrated,the imaging principle of the pinhole camera was studied,and the geometric model of the process of mapping the coordinate points in the three-dimensional space to the pixels in the image through the camera was obtained.Because the distortion may lead to the change of imaging position,this paper analyzes the camera distortion,and uses the nonlinear optimization method to calibrate the distortion coefficient of binocular camera.The error of the IMU was explored,the Gaussian white noise and random walk of the IMU at rest were calibrated,and the external parameters between the vision and the IMU were jointly calibrated.Research the advantages and disadvantages of vision and IMU sensors and their data fusion solutions,and select the optimal tight fusion solution for the system as a whole.3.In this paper,three methods of visual odometer are experimentally studied.The optical flow method which is most suitable for the application scene in this paper is used as the visual front end,and the Gaussian Newton method is used to realize GFTT corner plus multi-layer reverse optical flow tracking,which has achieved good tracking effect.Aiming at the problem that IMU integration needs to be carried out again for each pose optimization update,this paper adopts the method of IMU pre integration to avoid repeated calculation.Based on visual residuals,IMU residuals and marginalized prior residuals,the state optimization problem of residual function is constructed to solve the vehicle pose.In addition,for loop closure detection,this paper selects relevant data sets to extract uniform feature points based on the bag-of-words model,and trains a dictionary suitable for the application scenarios of this paper,which improves the recall rate of loop closure detection.4.In order to evaluate the effectiveness of the positioning algorithm in this paper,a software platform and a hardware platform are built,and indoor experiments and real vehicle experiments in tunnels are carried out to verify the feasibility of the algorithm.Finally,the automatic driving data set Kitti is used to test the positioning algorithm in this paper.The experiments show that the positioning algorithm in this paper has high accuracy.