CN-122023109-A - Ultra-large scale multi-time phase point cloud layered registration evaluation method and system

CN122023109ACN 122023109 ACN122023109 ACN 122023109ACN-122023109-A

Abstract

The invention provides a hierarchical registration evaluation method and system for ultra-large-scale multi-time-phase point clouds, which comprises the steps of acquiring first time-phase point clouds of source data and second time-phase point clouds of target data, analyzing multi-resolution level depth information and total points of the first time-phase point clouds and the second time-phase point clouds, performing coarse registration level selection according to the multi-resolution level depth information and the total points, executing a manual point selection registration process or an automatic feature matching process, outputting a global coarse registration transformation matrix, determining iteration times or deepest levels by taking the coarse registration transformation matrix as an initial transformation matrix, performing hierarchical progressive fine registration, outputting a fine registration transformation matrix, and judging whether to enable a multi-scale model to be compared with a model cloud based on the fine registration transformation matrix The algorithm carries out precision evaluation, if yes, the algorithm outputs And (5) evaluating the result of the precision, otherwise, executing a standard evaluation flow and outputting the result of the standard precision evaluation.

Inventors

GUI ZHIPENG
WANG YAN
PENG DEHUA

Assignees

武汉大学

Dates

Publication Date: 20260512
Application Date: 20260121

Claims (10)

1. A hierarchical registration evaluation method for ultra-large-scale multi-phase point clouds is characterized by comprising the following steps: Acquiring first time phase point cloud data of source data and second time phase point cloud data of target data, and analyzing multi-resolution level depth information and total points of the first time phase point cloud data and the second time phase point cloud data; Coarse registration level selection is carried out according to the multi-resolution level depth information and the total points, a manual point selection registration process or an automatic feature matching process is executed, and a global coarse registration transformation matrix is output; the coarse registration transformation matrix is used as an initial transformation matrix, iteration times or the deepest level is determined, hierarchical progressive fine registration is carried out, and a fine registration transformation matrix is output; determining whether to enable multi-scale model-to-model cloud comparison based on a fine registration transformation matrix The algorithm carries out precision evaluation, if yes, the algorithm outputs And (5) evaluating the result of the precision, otherwise, executing a standard evaluation flow and outputting the result of the standard precision evaluation.
2. The method of claim 1, wherein obtaining first phase point cloud data of source data and second phase point cloud data of target data comprises: and reading metadata of the pre-constructed octree index file in an out-of-core mode to obtain first time phase point cloud data of the source data and second time phase point cloud data of the target data.
3. The hierarchical registration evaluation method of a very large scale multi-phase point cloud according to claim 1, wherein the coarse registration level selection is performed according to multi-resolution level depth information and total points, a manual point selection registration process or an automatic feature matching process is performed, and a global coarse registration transformation matrix is output, comprising: Calculating to obtain target points according to a preset sampling percentage based on the total points, traversing multiple resolution levels, counting the total points of each level, and determining the level closest to the target points as a rough resolution level; if the manual point selection registration process is confirmed to be executed, loading first time phase point cloud data and second time phase point cloud data of the coarse resolution level, calculating fast point characteristic histogram features of the first time phase point cloud data and the second time phase point cloud data, executing a fast global registration algorithm, and obtaining the coarse registration transformation matrix through matching in a fast point characteristic histogram feature space; If the automatic feature matching process is determined to be executed, loading the first time phase point cloud data and the second time phase point cloud data of the coarse resolution level, visualizing, reading a plurality of pairs of homonymous points picked up by a user on the first time phase point cloud data and the second time phase point cloud data through a graphical interface, and resolving the plurality of pairs of homonymous points through point-to-point estimation to obtain the coarse registration transformation matrix.
4. The hierarchical registration evaluation method of a super-large-scale multi-phase point cloud according to claim 1, wherein the method comprises determining the iteration number or the deepest level by taking a coarse registration transformation matrix as an initial transformation matrix, performing hierarchical progressive fine registration, and outputting a fine registration transformation matrix, and comprises: in any iteration, determining a fine registration level, and calculating the physical resolution corresponding to the fine registration level; Based on the initial transformation matrix, dynamically inquiring in indexes of the first time phase point cloud data and the second time phase point cloud data to obtain spatial overlapping data nodes on a level closest to the physical resolution corresponding to the fine registration level; loading the spatial overlapping data nodes which are subjected to query and spatial downsampling in the spatial overlapping data nodes; Estimating normal vector of the down-sampled spatial overlapping data node, executing a point-to-plane iterative nearest point algorithm, and calculating to obtain incremental transformation; Updating the initial transformation matrix based on the incremental transformation to obtain an updated initial transformation matrix; If the iteration times or the deepest level is reached, the iteration converges, the updated initial transformation matrix is output as the fine registration transformation matrix, otherwise, the next fine registration level and the corresponding physical resolution are continuously selected, and the iteration is continuously carried out.
5. The hierarchical registration evaluation method of a very large scale multi-phase point cloud according to claim 4, wherein dynamically querying in an index of first phase point cloud data and second phase point cloud data based on the initial transformation matrix, obtaining a spatial overlapping data node on a level closest to a physical resolution corresponding to the fine registration level, comprises: Acquiring three-dimensional bounding boxes of all nodes of the current fine registration level of the first phase point cloud data corresponding to the physical resolution; performing coordinate transformation on the three-dimensional bounding box by using the initial transformation matrix to obtain a transformed bounding box; and in the multi-resolution index of the second time-phase point cloud data, searching to obtain nodes with all the physical resolutions closest to the physical resolution and intersected with the transformed bounding box as the spatial overlapping data nodes.
6. The method for hierarchical registration evaluation of a very large scale multi-phase point cloud as claimed in claim 5, wherein loading the queried and spatially downsampled spatial overlapping data nodes of the spatial overlapping data nodes comprises: The list of the spatial overlapping data nodes is sampled spatially uniformly, so that the number of nodes loaded to the memory does not exceed a preset memory budget threshold; And loading the sampled node data with the preset memory budget threshold number from outside the disk core to serve as a down-sampled space overlapping data node.
7. The method of claim 6, wherein estimating normal vectors of downsampled spatially overlapping data nodes, performing a point-to-plane iterative closest point algorithm, and solving for incremental transformations comprises: Estimating a normal vector for the down-sampled spatial overlapping data nodes of the second phase point cloud data; Establishing an optimization objective function from point to plane based on the corresponding point pairs of the first time phase point cloud data and the second time phase point cloud data and the estimated normal vector; And solving the optimized objective function through a least square method to obtain incremental transformation.
8. The method for hierarchical registration assessment of a very large scale multi-temporal point cloud of claim 1, wherein the steps of The algorithm comprises the following steps: sampling reference points of the first time phase point cloud data and the second time phase point cloud data to obtain core points; Estimating a normal vector for the second phase point cloud data; Aiming at the core point of each first time phase point cloud data, searching a neighborhood point of the second time phase point cloud data in a preset radius cylinder along the corresponding normal direction in the second time phase point cloud data, and calculating the average distance of the neighborhood points; Repeating the steps of estimating normal vectors as second time phase point cloud data and calculating average distances of the neighborhood points on at least two different scales, wherein the two different scales correspond to different neighborhood radiuses sampled when estimating the normal vectors, and carrying out weighted average on calculation results under different scales to obtain comprehensive average distances; Calculating a local standard deviation of the comprehensive average distance; Generating a containment The errors of the distance and confidence attributes distribute the point cloud.
9. The method for hierarchical registration evaluation of a very large scale multi-phase point cloud as set forth in claim 1, wherein the standard evaluation procedure comprises: Setting a preset distance threshold on the space overlapping data node of the deepest hierarchy; Calculating the point proportion of the distance between the source data node and the target data node, which is smaller than the preset distance threshold value after the fine registration transformation matrix transformation, as the overlapping degree; And calculating root mean square error for the nodes in the overlapping degree range.
10. A super-large multi-phase point cloud hierarchical registration evaluation system, comprising: The acquisition module is used for acquiring first time phase point cloud data of the source data and second time phase point cloud data of the target data, and analyzing multi-resolution level depth information and total points of the first time phase point cloud data and the second time phase point cloud data; the coarse registration module is used for carrying out coarse registration level selection according to the multi-resolution level depth information and the total points, executing a manual point selection registration process or an automatic feature matching process and outputting a global coarse registration transformation matrix; The fine registration module is used for taking the coarse registration transformation matrix as an initial transformation matrix, determining iteration times or the deepest level, carrying out hierarchical progressive fine registration and outputting a fine registration transformation matrix; an evaluation module for judging whether to enable based on the fine registration transformation matrix The algorithm carries out precision evaluation, if yes, the algorithm outputs And (5) evaluating the result of the precision, otherwise, executing a standard evaluation flow and outputting the result of the standard precision evaluation.

Description

Ultra-large scale multi-time phase point cloud layered registration evaluation method and system Technical Field The invention relates to the technical field of three-dimensional data processing, in particular to a hierarchical registration evaluation method and system for ultra-large-scale multi-time-phase point clouds. Background Point pattern analysis (Point PATTERN ANALYSIS, PPA) is an important component of spatial statistics, aimed at studying the distribution pattern of Point data over a range of space. With the dramatic development of sensor technology (e.g., three-dimensional laser scanning), analysis of point data has gradually expanded from a spatial dimension to a spatio-temporal dimension. In structural health monitoring (Structural Health Monitoring, SHM) applications of large infrastructures, periodically acquiring multi-temporal three-dimensional point cloud data has become a key technical means. By comparing the point cloud data of different time phases, the settlement, displacement and deformation of the structure can be analyzed with high precision. However, before the above comparative analysis is implemented, a critical and challenging pre-step is to register (i.e., "align") the point cloud data at these different phases, under respective independent scanning coordinate systems, into the same unified coordinate system with high accuracy. The point cloud registration method in the related art generally faces serious memory bottleneck problems when applied to the ultra-large-scale multi-time phase point cloud data. The point cloud data volume of a large structure is very short to billions, and the original file size can reach tens of GB and even TB levels. Most classical registration algorithms in the related art use an "in-core" processing mode, i.e. assuming that all data can be loaded into the Random Access Memory (RAM) of a computer at once. For massive data, this is not feasible at all on conventional workstations, which would directly result in Out-of-Memory errors. Meanwhile, the related art also faces the problem of low calculation efficiency. Even on a server cluster with massive memory, performing global computation (e.g., computing FPFH features) or iterative matching (e.g., nearest neighbor search of ICP) on several billions of points, its temporal complexity (e.g.Or (b)) The time consumption of calculation is astronomical number, and the engineering practicability is not realized. In addition, the related art faces challenges of registration accuracy and robustness. Large industrial facility scenes have a large number of geometrically repeatable structures (e.g., pipes, trusses) that result in very low discrimination of local geometric features (e.g., FPFH) such that feature-based global registration algorithms (e.g., FGR, RANSAC) are very prone to produce a large number of mismatches that fail to find the correct initial alignment pose, while iterative algorithms (e.g., ICP) are very prone to fall into a locally optimal solution (Local Minima) in the absence of a good initial pose or in the face of a repeated structure. Finally, the related art also has limitations in evaluating the system. The most common indicator, such as Root Mean Square Error (RMSE), relies on a subjectively set "interior point" distance threshold, which is chosen with randomness. More seriously, the index cannot distinguish between "registration errors" and "true topology changes" (e.g., inAt the moment remove oneEquipment present at a time) resulting in distortion of the evaluation results, masking the high precision alignment facts of the stable structure. In view of the foregoing, there is a lack of a point cloud registration method in the related art that systematically addresses the above-mentioned challenges, which is scalable, efficient, robust, and objective in the evaluation system. Disclosure of Invention The invention provides a hierarchical registration evaluation method and system for ultra-large-scale multi-time-phase point clouds, which are used for solving the defects that in the prior art, when ultra-large-scale (billions of points) multi-time-phase point clouds are processed, the data cannot be completely loaded, so that the bottleneck of an 'out-of-core' computing memory is caused, the registration efficiency is low and is easy to fall into a local optimal solution due to huge data quantity and strong structural repeatability, and the registration error cannot be objectively distinguished from the real topology change by the traditional evaluation method. In a first aspect, the present invention provides a hierarchical registration evaluation method for a super-large-scale multi-phase point cloud, including: Acquiring first time phase point cloud data of source data and second time phase point cloud data of target data, and analyzing multi-resolution level depth information and total points of the first time phase point cloud data and the second time phase point cloud data; Coarse re