Research On Lidar SLAM Algorithm For Unmanned Vehicles In Dynamic Scene

Lidar Simultaneous Localization And Mapping(SLAM)is a method for unmanned vehicles to achieve high-precision positioning in satellite signals rejected environment,and it is of great value in both military and civilian fields.By comparing multi-frame lidar point clouds,the motion of the unmanned vehicle is calculated and the position of the unmanned vehicle is determined.However,dynamic objects in the environment will lead to incorrect matching of feature points,and motion distortion will occur in the lidar point cloud during the acquisition process,which will lead to a decrease in the accuracy of the calculation results.This paper studies a lidar SLAM algorithm for unmanned vehicles in dynamic scenes,mainly involving point cloud distortion elimination and ground recognition,removal of dynamic objects,feature extraction and odometry considering ground point clouds.The main contents are as follows:(1)For the problem of motion distortion,the point cloud data is preprocessed,and the inertial navigation data is coupled to eliminate nonlinear distortion.The point cloud comparison results show that the distortion has been eliminated.In order to mark the ground point cloud,compared with the existing ground recognition algorithms,and considering the problem that the point cloud at the junction of the ground is dense and difficult to segment,an improved method based on multi-level map grid is proposed.After comparing the recognition results,it is proved that the improved algorithm has better recognition efficiency and faster running speed compared with the existing algorithm.(2)For dynamic objects in the scene,use point cloud combined with range image for recognition.Project unstructured point cloud data into a range image based on horizontal angle and number of lines.Optimize the range image pixel access method to simulate the circular access characteristics of the sub-bundle.Calculates the multi-step gradient of the range of each point and the adjacent points and marks the arc segments in a bundle in pixel order.Using the segmental principal component analysis results to judge the similarity of each segment and cluster dynamic objects,complete the marking and elimination of dynamic point clouds.Finally,the repaired point cloud is grown from both ends of the eliminated point using the principle of invariance of the second derivative to complete the repair of the point cloud.According to the analysis results of the geometric error between the repaired point cloud and the real object,it can be seen that the elimination and repair of dynamic objects are completed well.(3)In the part of odometry and mapping,a variety of feature point extraction methods are compared,and considering the real-time performance of the algorithm,the curvature operator is selected to extract features.According to the ground recognition results,the feature points are divided into four categories: ground edge points,non-ground edge points,ground plane points,and non-ground plane points,and the feature points are matched according to the categories to improve the matching efficiency.The motion state estimated by the inertial navigation is used as the initial value of the motion estimation of the L-M method to reduce the number of iterations.After the odometry is completed,the adjacent sub-maps are extracted for integration and optimization to further improve the positioning accuracy.Finally,the point cloud is down-sampled to reduce the occupation of resources and build a point cloud map.Finally,experiments are carried out on the point cloud data in the actual dynamic scene and the widely recognized open source KITTI lidar data set to evaluate the positioning accuracy of the algorithm and the running speed of each module of the algorithm.The results show that,compared with the current mainstream lidar algorithm A-LOAM,this algorithm greatly improves the positioning accuracy and slightly increases the running speed;compared with the popular algorithm combining with semantic information S-ALOAM,this algorithm keeps approximate positioning accuracy and greatly increases running speed.