An Offline Precision Improvement Algorithm For 3D Laser SLAM Point Cloud

With the digitization and informationization of various industries,traditional twodimensional maps can not meet the needs of people's spatial geographic information.Therefore,3D spatial information capturing technology has become an important research direction in the field of surveying and mapping.At present,when the GNSS(Global Navigation Satellite System)signal is good,the vehicle-borne and airborne laser scanning technology can already achieve the three-dimensional information collection work well,but in the case that the GNSS signal is severely occluded or missing,these method can not work correctly,the traditional static laser measurement methods are difficult to reach the requirements of fast acquisition.The Simultaneous Localization and Mapping(SLAM)technology can perform self-positioning and mapping of the surrounding environment synchronously by the sensors mounted on the measurement platform.The laser SLAM technology combined with laser scanning technology and SLAM technology can achieve fast and efficient acquisition of spatial three-dimensional information.However,due to the characteristics,solving in real-time,of SLAM technology,the laser SLAM technology can not use of information in the point cloud sufficiently,and the acquired spatial threedimensional information has limited precision,which limits its wide application in various fields.In the field of surveying and mapping,solving point cloud in real-time is not necessary,and using post-processing algorithm to improving the precision of point cloud becomes a viable option.So in order to make full use of the information in the 3D laser SLAM point cloud to solve the problem that the improving of laser SLAM point cloud accuracy,this paper proposes a hierarchical laser SLAM point cloud global registration algorithm based on feature matching and closed loop constraint.And the data captured by a self-made laser SLAM system is taken as an example to verify the effectiveness of the proposed algorithm.The main research contents and conclusions of this paper are as follows:(1)Segmented laser SLAM point cloud sequential registration.The laser SLAM point cloud is segmented by time,and the point cloud in the segment is registered to eliminate the accumulation error inside the segment.We using the average entropy and the average plane variance as the metric of the registration result,and compare the advantages and disadvantages of different point cloud registration algorithms.The results show that the point to plane ICP algorithm achieves the best registration results in the experimental data.(2)Pose graph construction based on point cloud similarity.First,we calculate the FPFH(Fast Point Feature Histograms)feature of each point in the cloud,and calculate the VLAD(Vector of Locally Aggregated Descriptors)global features of each point cloud according to the obtained FPFH features.Then,we calculate the Similarity between different point clouds segment.Running the Feature based registration of point clouds with high similarity,and refining of feature based registration result.Finally,according to the obtained registration result,the constraint of relative position and rotation between local point clouds is formed to construct pose graph.(3)Consistency detection of pose graph based on closed loop constraint.Using the closed-loop search algorithm of the undirected graph,the closed loop in the pose graph is searched and validated,the edges formed by the false matching result are removed,and the pose is calculated by the nonlinear optimization algorithm to obtain the final optimization result.(4)Experiment and evaluation of precision.We comparing the point cloud before and after optimization with the reference point cloud obtained by using high-precision static terrestrial laser scanning.The experiment results show that the accuracy of the three experiment point cloud data has been improved significantly.The corresponding point RMSE of data 1 reduced by 41.3%,the corresponding point RMSE of data 2 reduced by 62.7%.,and the corresponding point RMSE of data 3 reduced by 33.8%,which verifies the effectiveness of the proposed algorithm.The hierarchical 3D laser SLAM point cloud global registration algorithm based on feature matching and closed loop constraint proposed in this paper can effectively improve the accuracy of obtaining point cloud and help overcome the shortcomings of 3D laser SLAM point cloud precision.It can promote the applications of 3D laser SLAM point cloud in various fields.