Research On Algorithm Of 3D High-precision Map Generation Based On Lidar And IMU

With the rapid development of Internet technology and data science technology,self-driving cars are becoming smarter and smarter.Driverless vehicles have positive applications in the fields of environmental protection,road safety,and resource utilization.However,modern urban road conditions are becoming increasingly complex,and simple positioning methods can hardly solve the complex terrain positioning problems faced by driverless vehicles.The 3D high-precision map can effectively solve the above problems and can meet the requirements of real-time and high-precision positioning of driverless vehicles at the same time.Simultaneous Localization And Mapping(SLAM)is the mainstream method for generating high precision point cloud maps.SLAM is divided into visual SLAM and laser SLAM.3D LIDAR measures the distance between the environment and itself by emitting a laser beam,which can better reflect the geometric features of the surrounding environment.In the field of autonomous driving where safety is paramount,laser SLAM is more robust than visual SLAM methods.The IMU can not only de-motionally distort the original point cloud data,but also the IMU pre-integration can provide an initial positional transformation to the point cloud alignment.The 3D LIDAR and IMU have high complementarity and reliability,and are the hot spots for multi-sensor fusion research.Therefore,this thesis uses SLAM algorithms that fuse IMU and LIDAR sensors to generate 3D high-precision maps.The frequency of key frame insertion in SLAM algorithm is very fast,which leads to the rapid increase of redundant information,but these redundant information has almost no effect on the accuracy of the system,instead,more computational resources are lost.The loop closure detection algorithm that is widely used and has a good loop closure effect is Scan Context,but in scenarios where LIDAR passes through the same place in opposite directions,or LIDAR passes through different directions from the intersection direction,Scan Context is less effective in loop closure detection because it does not have translation invariance.When building a map with SLAM in a dynamic environment,it is inevitable to encounter the interference of dynamic objects,which eventually causes the point cloud map generated by SLAM to contain dynamic points.These dynamic points formed by dynamic objects in the point cloud map are harmful to the subsequent localization or navigation tasks.To address the above problems,this thesis mainly investigates the following.I.To solve the problem that the key frame insertion frequency in SLAM algorithm is too fast,which leads to the increase of redundant point clouds involved in the operation,this thesis improves the conditions for selecting key frames on the basis of LIO-SAM.Without losing accuracy and robustness,filtering redundant key frames reduces the time of back-end factor map optimization,decreases computer resource loss,and ensures the smooth operation of the system.II.To solve the problem of poor Scan Context loop closure detection due to different directions of LIDAR passing through the same location,the Lidar Iris loop closure detection algorithm is integrated into the LIO-SAM framework in this thesis.The algorithm in this thesis has improved the accuracy and recall rate compared with the original algorithm.After the loop closure is detected by the Lidar Iris algorithm,the rotation matrix obtained by aligning the loop closure frame with the current frame is added to the factor graph as a loop closure detection factor,so that the factor graph can be optimized to reduce the cumulative drift error.III.In order to solve the problem that the final point cloud map built by SLAM contains dynamic points,this thesis proposes an improved dynamic point filtering algorithm based on the ERASOR algorithm.In this thesis,dynamic points are successfully filtered out from the original point cloud map containing dynamic points to generate the final 3D high-precision map without dynamic points,and the preservation rate and rejection rate are improved compared with the original algorithm.Finally,this paper was tested by KITTI public dataset,and the experimental results verified the effectiveness of the algorithm in this paper,which laid a solid foundation for further development.