Vehicle Model Constraint Based Visual-inertial Localization Algorithm For Autonomous Driving In Closed Park

In recent years,due to the rapid development of automobile industry and artificial intelligence,autonomous driving technology has become a research hotspot in the industry,and various scientific research institutions,automobile manufacturers and Internet companies have stepped in and carried out technical landing and commercial application in specific park scenes.The key technologies of autonomous driving mainly include environment perception,navigation and positioning,decision planning and motion control,etc.Accurate positioning is an important link to realize autonomous driving of vehicles,providing accurate and real-time vehicle posture information for decision planning.At present,most self-driving vehicles use GNSS/INS to achieve real-time and high-precision positioning.However,due to the occlusion of buildings and trees in the environment,GNSS signals are unstable in some areas,resulting in positioning deviation,which can not meet the positioning requirements of self-driving vehicles in the park.In recent years,the development of SLAM technology provides a new idea to solve the above problems,and it is expected to realize accurate and realtime vehicle positioning in occluded environment through existing sensors in the autonomous driving system.However,when the existing visual SLAM technology is applied to autonomous driving positioning in closed parks,there are still some problems,such as lost feature tracking and large accumulated drift,and the specific application scenarios and actual motion characteristics of vehicles are not fully considered,which affects the positioning accuracy and robustness of vehicles.Therefore,this paper takes the visual inertial positioning system of autonomous driving in closed park as the research object,adopts the visual front-end based on characteristic optical flow tracking and adaptive depth estimation,constructs a kinematic model that conforms to the actual motion constraints of vehicles,realizes the joint pose optimization of visual inertia under the constraints of vehicle model based on the sliding window tight coupling optimization,and designs a loop detection algorithm considering the geometric constraints of image feature space,aiming at improving the accuracy and robustness of the visual inertial positioning system of autonomous driving in closed park.The main research contents of this paper are as follows:(1)Framework design of visual inertial positioning system based on vehicle model constraints.Environmental image information is obtained by camera,visual reprojection error is constructed based on feature correspondence,vehicle motion state information is obtained by IMU,IMU measurement residual is constructed based on IMU pre-integration,vehicle pose change constraint is constructed by vehicle kinematics model,and positioning system cost function is jointly constructed based on visual measurement residual,IMU measurement residual and vehicle kinematics model constraint to realize tight coupling optimization of vehicle pose.The attribute map matching test considering geometric constraints of feature space is added to improve the accuracy of loop detection and reduce the cumulative drift of vehicle pose estimation.Global pose map optimization is used to ensure the global consistency of pose,and the function of saving and loading pose map is added to realize map reuse.(2)Visual front end based on feature optical flow tracking and adaptive depth estimation.The front-end processing combining corner extraction and optical flow tracking is used to extract corner features from each frame image,and the sparse optical flow algorithm is used to track it and estimate its motion.Optical flow tracking can avoid the time-consuming of computing and matching descriptors,directly obtain the corresponding relationship of feature points,and effectively improve the tracking loss.A feature depth estimation method combining binocular matching and moving perspective stereo vision is adopted.Parallax is calculated according to binocular image matching,and far and near feature points are distinguished.The near feature points directly obtain depth information from binocular images,and the far feature points are estimated by moving perspective stereo vision of multi-frame images.The combination of the two methods can not only make full use of the effective information of binocular matching,but also make up for the limitation of depth estimation and realize adaptive feature depth estimation.(3)Visual inertia tightly coupled pose optimization based on vehicle kinematics model constraints.Combined with specific application scenarios and actual motion characteristics of vehicles,considering the movement of vehicles on the same plane at adjacent measurement times,the local motion is described according to Ackerman steering model,and the kinematics model of self-driving vehicles is constructed based on bicycle model principle without introducing additional sensors.The prediction expression of vehicle posture change at adjacent measurement times is derived,and the constraints of vehicle kinematics model are constructed,thus realizing the visual inertia tight coupling posture optimization constrained by vehicle kinematics model,reducing the optimization space of vehicle posture and improving the posture estimation results.(4)Design of loop detection algorithm considering geometric constraints of image feature space.The attribute map model of scene image is established,and the attribute map matching test considering the position relationship of feature space is added to the traditional word bag model,so that the loop verification based on attribute map matching is realized,the accuracy of loop detection is improved,and the cumulative drift of vehicle pose estimation is reduced.The position of the vehicle passing by is identified by loop detection,the feature level connection between the current frame and the loop frame is established,and the corresponding relationship is integrated into the sliding window of local visual inertia joint optimization,so that the pose estimation of the low drift vehicle can be realized with low computational cost.(5)Experimental verification based on public data sets and closed campus scene data.Based on KITTI public data set and closed park data collected by intelligent networked vehicle platform,the algorithm is tested and tested,and compared with the current mainstream visual inertial positioning algorithm ORB-SLAM3,which fully verifies the effectiveness of the algorithm.Experimental results show that the visual inertial positioning system based on vehicle kinematics model constraints can effectively improve the positioning accuracy and robustness of self-driving vehicles in closed parks.