Research On Motion Estimation Based On Multi-beam Laser Scanning Point Clouds

Multi-beam LiDAR can quickly provide high-precision point cloud data in a large range.It provides a reliable data guarantee for autonomous vehicles’ perception of the environment.The correct perception of the surrounding environment and its motion state is the prerequisite for an autonomous vehicle’s decision-making.The ego-motion perception is mainly to estimate the position and posture of the autonomous vehicle itself in the current environment,and it provides a basis of determination for vehicle path planning and motion control.Laser scanner-based motion estimation makes autonomous vehicles be independent of satellite navigation systems and inertial navigation systems,thereby reducing interference from weather or environment,and improving the reliability and accuracy of vehicle pose estimation in complex environments.Therefore,it has become a hot research topic and is widely employed in tasks such as autonomous robot navigation and 3D scene reconstruction.This article focuses on the key scientific problem of motion estimation based on multi-beam laser scanning point clouds.We analyzed the advantages of motion estimation based on point cloud matching and the problems existing in prior methods.We summarized three critical issues as follows:the existing point cloud matching and motion estimation methods can not extract robust feature descriptions to establish the effective matching relationship between point clouds;it is difficult to establish the reliable matching relationship between point clouds using only geometric information;there are interference elements in the environment(such as dynamic objects)that are not conducive to estimating the ego-motion of LiDAR.To address these issues,we researched three aspects:deep feature description of point clouds for 3D motion estimation,LiDAR point cloud matching with semantic information,and LiDAR ego-motion estimation combined with mask prediction.Meanwhile,we correspondingly proposed three novel algorithms,i.e.,FeatFlow,DeepSIR and LO-Net.We conducted experimental analysis on multiple public LiDAR benchmark datasets and verified the effectiveness of our algorithms.The results of comparative experiments with related methods show the superiority of our algorithms.The excellent performance of the algorithm allows it to be used in various applications that rely on laser scanning point clouds for motion estimation.