LiDAR-based 3D Mapping And Update Leveraging Heterogeneous Point Clouds

During the endeavor of bridging the physical realm with the digital realm,the threedimensional(3D)point cloud has emerged as the most direct manifestation of digitization.Point cloud maps with multi-dimensional properties can better meet the needs of the 3D real scene or digital twin than other data.Li DAR-based 3D mapping has emerged as the predominant and highly efficient approach for acquiring highprecision point cloud maps.On one hand,conventional mobile mapping equipment utilizing Push-Broom Li DAR(PBL)facilitates the acquisition of high-accuracy observations by leveraging precise inertial navigation systems.This enables the construction of large-scale 3D point cloud maps of scenes with centimeter-level accuracy.However,the aforementioned acquisition equipment entails substantial costs,and map updates necessitate significant human resources.Conversely,Multi-Beam Li DAR(MBL)presents distinct advantages in integrated urban mapping encompassing both indoor and outdoor settings,exploration of unknown underground environments,and intelligent driving.This is owing to its lightweight,miniaturized form,costeffectiveness,high timeliness,and all-weather data collection capabilities.Whereas,at the same time,MBL inherits the nature of sparse data,obvious difference in point cloud distribution,high acquisition frequency,internal motion distortion error of data,and limited vertical field of view.This brings great challenges to the rapid and accurate construction and update of large-scale 3D point cloud maps using MBL.Therefore,the fusion of the above two kinds of point cloud data can not only improve the accuracy of MBL-based 3D mapping,but also reduce the cost of timely update of the geographic information basic database.However,due to the obvious difference between the acquisition method of MBL and PBL data,it is called heterogeneous point clouds,as they differ in point density,point accuracy as well as Fo V.Specifically,the problems in Li DAR-based 3D mapping and update leveraging heterogeneous point clouds include reliable pose estimation of Li DAR odometry(LO),difficulty in constructing highquality and low-redundancy 3D point cloud map as well as in robust registration and change detection between heterogeneous point clouds.It is of great significance to study efficient,robust and high-precision 3D mapping and update algorithms for Li DAR to promote the application of MBL scanning systems in various fields of intelligent unmanned mapping and digital twin.Therefore,with the aim of overcoming the technical bottlenecks mentioned above,this doctoral dissertation conducts in-depth research from theoretical methods and key technologies.The specific research contents focus on:(1)The dissertation first introduces the important value of 3D point cloud map in the field of autonomous driving and new basic surveying and mapping.The important advantages of Li DAR-based 3D mapping are pointed out,clarifying the background and significance of this research.The research objectives and contents are determined in view of the shortcomings and difficulties of the existing large-scale point cloud acquisition methods.By comprehensively reviewing the related research works referring to LO,3D mapping,prior-map-based localization,life-long SLAM as well as3 D change detection,the technical route of this dissertation is put forward.(2)Aiming at the degradation problem of point cloud registration caused by low robustness of feature extraction,insufficient features or uneven feature distribution in LO,the structural feature extraction method based on the distribution chracteristics of the laser scanning line is proposed in this dissertation.To realize the robust and reliable pose estimation of LO,a novel robust feature association method assisted by error vector ellipsoid presented as well,which lays a foundation for the construction and update of high-quality point cloud maps.(3)Aiming at the problems of motion distortion in MBL point cloud data,a 3D continuous spatio-temporal consistency mapping method(3D-CSTM)is proposed in this dissertation.Based on the continuous-time spline motion model,the error accumulation and data redundancy in point cloud map are well handled.Through the novel point cloud fusion strategy considering the redundancy of time and space as well as the highly unified pose graph structure,the construction of a globally consistent point cloud map is realized.The trajectory closure difference can reach 0.78‰,which improves the performance of 3D point cloud map construction.Compared with other traditional methods,the distance residuals reduce by 20%,which provide complete data support for timely update of environmental 3D map.(4)To address the challenges of robust point cloud fusion and insufficient accuracy in 3D change detection arising from notable differences in distribution characteristics and field of view among heterogeneous point clouds,this dissertation proposes a novel joint optimization and map update method(CAOM)that integrates heterogeneous point cloud data from both PBL and MBL.Through this approach,the construction of high-precision point cloud maps and efficient 3D change detection is achieved.Therein,the generated virtual range and distance images are combined to detect local disparities,and the global factor graph is at last used to optimize the multiperspective observation probability propagation and update model.The precision of estimated trajectory improves with a maximum decrease of 1.5m for ATE,0.25 m for RMSE,0.12 m for STD.The overall accuracy of 3D points with residuals below 10 cm in the point cloud map is improved by up to 16%,and the overall 3D change detection accuracy surpasses 95% which can provide strong technical support for timely update of point cloud database.