NNDT-SLAM:Normal And Normal Distributions Transform Odometry And Mapping

Simultaneous Localization and Mapping(SLAM),as the key technique to mobile robot,has attracted the attention of many researchers in recent years.As a stable and reliable real-time location and mapping scheme,Lidar SLAM technology has become an important technology for intelligent mobile robot in the industry.In recent years,the rapid development of driverless makes the mass production of Lidar possible,which makes it possible for the industry of laser slam positioning scheme,which was originally expensive,to land.Compared with visual SLAM,Lidar has stronger data stability and environmental anti-jamming ability.Therefore,the research of Lidar SLAM is of great significance to realize the new era of information,digital and intelligent.Lidar can perceive the surrounding scene by acquiring the three-dimensional structure information of the environment.Due to the incomplete data acquisition of the three-dimensional scene and the complexity of the actual scene,Lidar SLAM becomes more challenging.Lidar SLAM based on normal distribution transformation is one of the most effective Lidar SLAM schemes.It can compress the information of 3D scene by gridding 3D space and using normal distribution to describe the distribution of point cloud in grid,so as to improve the efficiency of storage and matching.However,the choice of grid scale and guarantee of matching robustness has been difficult.To solve these problems,a Lidar SLAM algorithm based on the fusion of normal and normal distribution transform(NNDT)is proposed.In order to improve the robustness of the Lidar odometry,the NNDT-SLAM front-end uses the Lidar odometry based on the uniform filtering algorithm of normal feature points.The complexity of the environment causes the non-uniform distribution of feature points,which makes the distribution of matching constraints uneven in space,thus leading to the failure of matching algorithm.In order to get a more robust matching result,firstly,the point cloud is segmented in range image.The normal vectors are estimated based on the segmented point cloud.the normal vectors are distributed in the sphere space as evenly as possible,thus constraining the distribution of the feature points in the space.The Lidar odometry based on frame-frame registration algorithm is not accurate enough.This paper describes the frame-map NDT matching algorithm,and proposes an improved OMNDT registration algorithm for the point cloud registration in dynamic scene.Finally,in order to effectively resist the degradation of covariance matrix in NDT grid,this paper proposes a covariance matrix estimation algorithm based on normal vector correction.Finally,this paper briefly describes the loop detection algorithm based on appearance and position,designs NNDT-SLAM system based on multi-line Lidar,and builds a hardware experiment platform to verify the designed system.The results show that the NNDT-SLAM algorithm has excellent positioning accuracy and mapping results in both the actual scene and the KITTI data set.