The Research On Fast Registration Method Of Point Cloud Data Based On Voxel Grid

In recent years,with the increasing maturity of terrestrial laser scanning(TLS)technology,3D digital scene reconstruction based on laser point cloud data has gradually become a research hotspot in the field of computer vision.TLS can quickly and accurately capture geospatial information and has become an important technical means for applications such as terrain surveying and urban reconstruction.However,a single TLS has limitations such as instrument range and line of sight obstruction,and3 D scene reconstruction often needs to be achieved by acquiring multi-station laser scanning point cloud data.One of the key steps in multi-site cloud data processing is to merge them into the same coordinate system,which is the point cloud registration problem.The TLS point cloud data has characteristics such as large data volume and uneven density distribution,which have a greater impact on the efficiency of automatic registration.At the same time,there are a large number of repetitive and symmetrical structures in the artificial environment,such as buildings,streets,etc.,which are prone to misalignment and other registration failures during registration.The above problems have become the bottleneck of current laser scanning in engineering applications.In view of the above-mentioned difficulties,this research makes the following research:Aiming at the problem that a large amount of TLS point cloud data and the uneven density affect the registration efficiency,this paper proposes a Harris point cloud keypoint extraction algorithm based on a voxel grid.Create a voxel grid for the original point cloud to homogenize the point cloud density;at the same time,calculate the keypoints of the point cloud to replace the original point cloud,thereby reducing the amount of calculation and improving the registration efficiency.In this study,a voxel grid was established for the original point cloud data,and a density value was assigned to the voxel grid according to the distribution of points in each voxel grid.The voxel grid was used as the calculation point to make the calculation data evenly distributed Solve the problem of noise and outliers in the original point cloud.Based on calculating points with a density value grid,this paper is based on the idea of a two-dimensional Harris corner detection algorithm and improves the new corner response function for point cloud data to obtain threedimensional keypoints.Since the keypoints sought are discrete,there are problems such as insufficient location and poor quality.This study uses Taylor’s formula and the Gauss-Newton method to optimize the keypoint positions and improve the accuracy of the keypoint positions.Through the verification of experimental data,the keypoints generated by the keypoint extraction method in this paper are evenly distributed,and the time cost is 48%～81% of the SIFT3 D algorithm and 28%～52% of the Harris3 D algorithm.Aiming at the situation that a large number of repetitions and symmetrical structures in the scene lead to registration failure,this study uses voxel grids to improve the classic Super4 PCS algorithm.Use the voxel grid index to filter the four-element basis in the Super4 PCS algorithm to improve the registration success rate.Use the voxel grid index to query the decimal points on the four-element base edge,and obtain a ten-dimensional code from the query result to determine whether this edge exists on an object table plane.It is judged by multiple edges on the quaternary base whether the base is on the surface of an object,and if the quaternary base is on the surface of an object,it is discarded.By restricting the screening of the quaternary base,the quaternary bases on the surface of the same object(such as wall,ground,etc.)can be filtered out,which can reduce the number of candidate sets,improve the quality of candidate sets,and effectively reduce misalignment and other registration failures.Improve registration success rate and registration efficiency.In addition,the voxel grid index is used to replace the original KD tree index to improve the efficiency of the Largest Common Pointset(LCP)query in the Super4 PCS algorithm.The time complexity of querying random point cloud through voxel grid index is(1).Compared with the traditional KD tree index query with the time complexity of(),the time consumption is reduced by 50%～74%.Validated on different indoor and outdoor data sets,and compared with other four-point consistent set methods,the algorithm in this paper has improved time efficiency by more than 70% and registration success rate by more than 8% compared with the classic Super4 PCS algorithm.