Research On Offline Construction Method Of High-Definition Map Based On Intelligent Vehicle Perception System

The field of autonomous driving has become a hot field in the automotive industry.The safety of functions such as autonomous driving is inseparable from the assistance of highprecision maps.Nowadays,most vehicles need the assistance of maps when traveling.Highprecision maps provide a new technical route for the intelligent automobile industry,and also promote the rapid development of the autonomous driving technology industry.At present,the smart cars launched by major automobile manufacturers all have selfdriving functions such as adaptive cruise and path planning.Most of the vehicle navigation modules in the vehicle machine system are provided by special map service providers,such as Gaode Map,Baidu Map,etc.,and a small part is a map module developed by automobile manufacturers themselves.However,the above map navigation software is generally civil,which requires the driver to pay attention to and operate in real time.The automatic driving system cannot interact with the map software on its own,and the map navigation software is not suitable for intelligent vehicles.Nowadays,many indoor or limited-site automated guided vehicles use SLAM technology to achieve regional mapping and autonomous positioning.It has low cost,but it has few sensors and limited mapping scale and ability.It is generally only used in indoor venues and only supports relatively low speed driving.Smart cars generally carry a large number of sensors with high precision.This paper explores a kind of sensor based on intelligent vehicle itself,and applies SLAM technology to the mapping and positioning of intelligent vehicle,so as to construct a high-precision map suitable for large outdoor scenes and more information.In view of the above problems,the following research work has been carried out in this paper:Firstly,this paper compares the general point cloud registration methods,and selects the NDT registration method with better overall performance as the point cloud registration method in this experiment.Then,a filter is used to downsample the point cloud and remove outliers,so as to improve the registration efficiency and map accuracy.Then,through the ground point cloud filtering algorithm,non-ground objects,such as vehicles,buildings,plants,etc.,are segmented to obtain non-ground point cloud data suitable for efficient registration and positioning.Secondly,a three-dimensional mapping scheme based on graph optimization is explored.GNSS constraints and closed-loop detection constraints are added to the original odometeronly pose map to control the positioning error within a reasonable range,and effectively reduce the error and improve the consistency of the map.Through the time synchronization of each sensor,the data is more accurate at the time level.Then,the internal and external parameters of the lidar are calibrated,and the conversion relationship between the sensors and the vehicle is established to realize the spatial synchronization of the data.By fusing IMU data,the motion distortion caused by vehicle movement is compensated,and the point cloud motion distortion analysis and correction are achieved,which accelerates the processing speed of the algorithm.Finally,the intelligent vehicle experimental platform used in the experiment is built,and the software environment and hardware environment of the experimental system are developed.Through the experiment of integrating IMU data and 3D mapping algorithm based on graph optimization,the 3D map of campus scene is obtained.The experimental results of this paper are compared with the results of traditional algorithms and the results are analyzed.The experimental results show that the three-dimensional map obtained by the map construction method proposed in this paper has higher accuracy and better consistency,which solves the problems of sparse map point cloud and low accuracy caused by large error,and paves the way for the safe travel of intelligent vehicles.The method is effective.