Research On Algorithm Of LiDAR SLAM Based On DTORB Feature Loop Closure Detection

With the rapid advancement of artificial intelligence and related technologies,autonomous driving has become a highly scrutinized area of development within the automotive industry.It can improve driving safety,reduce accident rates and casualties.In the realm of autonomous driving technology,the Simultaneous Localization and Mapping(SLAM)technique plays a critical role.Through the use of on-board sensors such as cameras and Light Detection and Ranging(LiDAR),data is continually gathered to estimate the vehicle’s pose over a period of time to achieve the purpose of positioning and mapping.LiDAR sensors,due to their high resolution,strong anti-interference ability,and insensitivity to lighting changes,have been widely used in SLAM.Traditional LiDAR SLAM only relies on frame-to-frame matching of front-end odometry for pose estimation,which can lead to cumulative errors and cause map ghosting and drifting.To address this issue,loop closure detection has been introduced to SLAM,improving system performance and stability.Loop closure detection technology recognizes repeatedly visited areas,builds constraints between current and historical frames,and performs position optimization to reduce cumulative errors.However,existing loop closure detection solutions mainly rely on single-point cloud information to construct descriptors,which lacks scene discrimination and is sensitive to rotational transformation,leading to false positives or negatives.Therefore,in order to enhance the advancement of self-driving technology,further improvements to the loop closure detection module are necessary to improve the robustness and accuracy of SLAM systems.To address the aforementioned issues,this paper proposes a LiDAR SLAM algorithm based on Dynamic Threshold ORB(DTORB)feature for loop closure detection.The algorithm fully utilizes LiDAR data and introduces rotation-invariant features to enhance its anti-interference ability.By transforming the point cloud similarity problem into an image feature similarity problem,the effectiveness of loop closure detection accuracy is significantly enhanced,and drifting and ghosting in the map are reduced.The primary novelties of this paper are as follows.(1)This paper puts forward an approach for improving the accuracy of point cloud registration through point cloud feature extraction and matching.The method is based on gradient information,which is used to identify salient and sub-salient features.The relative pose transformation between adjacent frames is estimated through distance constraints from points to lines and points to planes.(2)This paper puts forward an approach for constructing multi-channel descriptors to address the issue of point cloud descriptors lacking scene discrimination ability.The method utilizes the distance,height,and intensity information of the point cloud,encoding them as three channels of the projected sub-regions.This generates a twodimensional global descriptor,which is used for subsequent loop closure detection.(3)This paper presents a novel DTORB feature extraction method to address the subjectivity problem of fixed threshold ORB feature extraction algorithms.Based on the objective global and local distribution of point clouds,this method dynamically adjusts the feature extraction threshold and then extracts ORB features according to the dynamic threshold.(4)This paper proposes a similarity measurement scheme based on DTORB features to solve the problem of sensitivity to rotation in existing similarity measurement methods.This scheme uses FLANN algorithm and LMEDS algorithm to quickly match DTORB features,and then determines whether a loop is formed based on the distance between the matched DTORB features.