| Since the 13 th Five-Year Plan,Our Country has always implemented the strategy of building a strong country through science and technology.The information and intelligent industries have developed rapidly.Autonomous driving and robotics,as the palm pearls in artificial intelligence-related fields,are also developing rapidly from the rise to the landing.How to make autonomous vehicles in an unknown environment The realization of positioning and surrounding environment map construction is a necessary prerequisite for the realization of subsequent environmental perception,path planning and decision-making in the entire automatic driving system process,and is a key technology for the stable and safe operation of automatic driving.In this paper,aiming at the L4-level autonomous driving scene on urban roads,using lidar and monocular camera as the main sensors,combined inertial navigation and other auxiliary equipment,to study the real-time positioning of urban dynamic road scenes and the construction of environmental maps.The research contents of this paper mainly include:(1)Multi-sensor fusion spatial parameter calibration and data preprocessing:Aiming at the point cloud distortion problem caused by the insufficient frame rate of the lidar in high-speed motion,analyze the cause of the distortion,and use the IMU sensor information to analyze the laser point cloud for each frame Motion distortion is compensated.In order to take advantage of the respective data advantages of the monocular camera and the lidar sensor,the internal parameters of the monocular camera are calibrated.In order to make the 3D points in each frame of data correspond to the 2D plane pixels of the image one-to-one,a plane target calibration model and related tools are used to The camera and lidar are calibrated jointly.Finally,in order to compensate for the blind spots and sparseness of the 32-line lidar point cloud,16-line lidar was added on both sides of the vehicle to remove the blind spots and perform multiple lidar external parameter calibrations.(2)Road dynamic target removal: In view of the problem that the sparseness of the lidar point cloud information may lead to a decrease in perception accuracy,with the help of the rich planar information of the two-dimensional image information,a two-dimensional image detection network is added to the original point cloud perception algorithm and constructed The fusion algorithm realizes the data fusion of the two sensors.(3)Point cloud map construction: In order to improve the robustness and accuracy of the odometer in the laser SLAM system in this environment,the laser SLAM framework is improved in order to improve the stability of the traditional mapping algorithm in the urban dynamic road environment.It includes adding a multi-modal perception algorithm to detect dynamic objects before the point cloud frame matching,and removing the dynamic objects in each frame of the point cloud based on the detected three-dimensional position information.At the same time,it uses the loop detection algorithm based on the lidar point cloud frame loop detection,and uses the observation constraint and loop constraint to correct the odometer pose and local point cloud map,which effectively reduces the odometer and mapping errors of the autonomous vehicle.(4)Data set and real vehicle algorithm verification: Using multiple sequences in the KITTI data set to compare the algorithm and the LOAM method in the three aspects of relative trajectory error,absolute trajectory error and mapping effect,confirming that the method in this paper is superior to LOAM.At the same time,data are collected on actual urban roads to verify the adaptability of the algorithm.In a dynamic road environment,the five aspects of key frame extraction,loop detection effectiveness,positioning error analysis,global and local mapping effects,and algorithm calculation time are analyzed.The test results show that the method in this paper can meet the corresponding indicators. |
