Search

CN-122023488-A - Forest multi-site point cloud registration method and related equipment

CN122023488ACN 122023488 ACN122023488 ACN 122023488ACN-122023488-A

Abstract

The invention discloses a forest multi-site point cloud registration method and related equipment, which are used for constructing a feature grid graph by fusing multiple types of semantic tags, avoiding registration failure caused by single feature deficiency or poor quality, improving suitability of the forest multi-site point cloud registration method in multi-season multi-shielding scenes, selecting a high-stability registration area by means of stability scoring, realizing self-adaptive adjustment of registration constraint by combining a semantic weighting mechanism, reducing dependence of coarse registration on initial pose and artificial experience, and reducing error matching interference. Through semantic weighted ICP fine registration and global optimization, geometrical stability differences of different semantic objects are distinguished, unstable structure interference is reduced, and registration accuracy is improved. Automatic stable registration can be realized without manual targets, so that the field operation cost and workload are reduced, and the wide-range forest scene popularization is facilitated.

Inventors

  • WANG DI
  • HE JIAQI

Assignees

  • 西安交通大学

Dates

Publication Date
20260512
Application Date
20260203

Claims (10)

  1. 1. A method for registering a forest multi-site point cloud, comprising: acquiring forest multi-site point cloud data and preprocessing; Carrying out semantic segmentation on the point cloud data preprocessed by each site to obtain multi-class semantic tags; Establishing a regular grid on a horizontal plane, and establishing a multi-semantic feature grid graph based on multi-type semantic tags in each grid unit; calculating stability scores of all grid units, and selecting a high-stability registration area according to the stability scores; performing coarse registration based on the multi-semantic feature grid graph, and estimating initial rigid body transformation between stations; performing fine registration based on a semantic weighted ICP algorithm on the basis of coarse registration; And determining a site pair set with effective overlapping, taking the pose of each site as a variable, converting the fine registration of the site pair into constraint, weighting the constraint by combining with the statistical information of the high-stability semantic region, constructing a global error function, and optimizing and solving to obtain a global consistent multi-site cloud registration result.
  2. 2. A method of forest multi-site point cloud registration as claimed in claim 1 wherein obtaining forest multi-site point cloud data and preprocessing comprises: Acquiring forest multi-site point cloud data through a scanner; Reserving points with the distance d being more than or equal to 1m and less than or equal to 50-150 m in the measuring range of the scanner, and eliminating points outside the range; deleting isolated noise points through a neighborhood point threshold value or statistical filtering; Compressing data by using a voxel grid with a side length of 0.02-0.1, balance density; Optionally, one station is used as a reference station or an external coordinate system is utilized, and each station cloud is used for aligning the z-axis with the gravity direction through simple plane fitting or using built-in inclination sensor information of an instrument.
  3. 3. The method for registering forest multi-site point clouds according to claim 1, wherein the step of establishing a regular grid on a horizontal plane and establishing a multi-semantic feature grid map based on multi-type semantic tags in each grid unit comprises the following steps: Constructing a regular grid for each station on a plane of a unified coordinate system, and projecting a point cloud to a horizontal plane for grid statistics; setting the grid size for each site, and dividing a regular grid by taking the grid size as a step length along the x and y directions; collecting all points falling into the area range of each grid unit, counting the points of each semantic category, and counting the point occupation ratio, the height statistical characteristics and the geometric characteristics of each semantic category; according to the points of each semantic category, the point proportion of each semantic category, the high-statistics features and the geometric features form multi-dimensional semantic feature vectors of grid units, and a multi-semantic feature grid graph is formed based on the multi-dimensional semantic feature vectors of each grid unit.
  4. 4. The method for registering a plurality of points in a forest according to claim 1, wherein the step of calculating a stability score of each grid unit and selecting a high-stability registration area according to the stability score is specifically as follows: Combining the trunk ratio, the ground ratio, the shrub ratio and the height fluctuation, constructing a scoring formula: Wherein, the Is the ratio of trunk class points in the grid, Is the duty ratio of the ground class points in the grid, For the bush-like point duty cycle, A1, a2, a3 and a4 are weight parameters for the standard deviation of the height in the grid; after scoring is calculated on all grids of each site, one mode is to set a stability threshold, only grids larger than the stability threshold are selected to serve as high-stability semantic grids, and the other mode is to sort all grids according to the stability score from large to small, and select the grids according to proportion to serve as the high-stability semantic grids, wherein the proportion of the grids is set to be 20% -50% according to an actual scene.
  5. 5. The method for registering forest multi-site point clouds according to claim 1, wherein the step of estimating an initial rigid body transformation between sites is specifically: Selecting one of two sites with an overlapping area as a processing object, firstly selecting the other site as a reference site and the other site as a target site, then searching candidate rotation angles by taking 1 DEG as a step length in a range of [ -180 DEG, 180 DEG ] around a z-axis for the multi-semantic feature grating graph of the target site, firstly rotating the target site grating graph around the z-axis by the angle for each candidate rotation angle, then estimating horizontal translation vectors of the target site grating graph relative to the reference site on an x-y plane, then aligning the rotated and translated target site grating graph with the reference site grating graph, screening out overlapping grating pairs with the same index, obtaining multi-dimensional semantic feature vectors and stability scores of each overlapping grating for each overlapping grating, calculating similarity between the feature vectors, weighting and accumulating the similarity according to the stability scores of the two gratings, obtaining overall matching scores of the rotation angle and translation vector combination, traversing all candidate and translation vector combination, then selecting the parameter with the highest matching score as a plane initial transformation parameter, finally utilizing the height median or average value of ground class points in the two sites, estimating translation quantity in the vertical direction, and finally forming a complete rigid body transformation including the three-dimensional rigid body transformation around the z-axis.
  6. 6. The forest multi-site point cloud registration method according to claim 1, wherein the step of performing fine registration based on a semantically weighted ICP algorithm on the basis of coarse registration is specifically as follows: Based on an initial alignment result obtained by rough registration, firstly, an overlapping area point set of two site clouds is intercepted, points involved in registration are screened by combining semantic labels, trunk class and ground class points are reserved as constraint objects, coarse branch class points are reserved as auxiliary constraints, bush class and fine branch class points are downsampled or directly removed, registration weights are distributed to the points of different semantic classes, wherein the trunk class point weight is the largest, the ground class point weight is the next smaller, the bush class point weight is the smallest or is set to be 0, the weight is set to be a fixed constant or is adjusted by referencing a stability score of a corresponding grid, in an ICP framework, differential error measures are adopted for the trunk class points, a cylindrical model or a dry axis is fitted in a local neighborhood, errors are defined as shortest distances from points to cylindrical surfaces or axes, ground class points are defined as distances from points to planes in a local fitting ground plane, standard point-to-point Euclidean distances are used as errors, and other points which are not specially processed are then, an objective function containing ICP weights is constructed, and iterative operations are carried out: Searching the closest point or the corresponding geometrical entity in the target point cloud for each source point under the current transformation, updating the rotation matrix and the translation vector according to the minimized target function of the current corresponding relation, repeating the process until the change of the target function is smaller than a preset threshold value or the maximum iteration number is reached, and finally obtaining the fine registration transformation between the sites.
  7. 7. A forest multi-site cloud registration system, characterized in that it is based on a forest multi-site point cloud registration method as claimed in any one of claims 1-6, comprising: the data acquisition module is used for acquiring forest multi-site point cloud data and preprocessing the data; the label module is used for carrying out semantic segmentation on the point cloud data preprocessed by each site to obtain multiple types of semantic labels; the grid diagram establishing module is used for establishing a regular grid on a horizontal plane and establishing a multi-semantic characteristic grid diagram based on multi-type semantic tags in each grid unit; The scoring module is used for calculating the stability score of each grid unit and selecting a high-stability registration area according to the stability score; The coarse registration module is used for performing coarse registration based on the multi-semantic feature grid graph and estimating initial rigid body transformation between stations; the fine registration module is used for carrying out fine registration based on a semantic weighting ICP algorithm on the basis of coarse registration; The registration module is used for determining a site pair set with effective overlapping, taking the pose of each site as a variable, converting the fine registration of the site pair into constraint, weighting the constraint by combining with the statistical information of the high-stability semantic region, constructing a global error function, and optimizing and solving to obtain a global consistent multi-site cloud registration result.
  8. 8. A computer device comprising a memory, a processor and a computer program stored on the memory, characterized in that the processor executes the computer program to implement the steps of a forest multi-site point cloud registration method as claimed in any of claims 1-6.
  9. 9. A computer readable storage medium, on which a computer program is stored, characterized in that the computer program, when being executed by a processor, implements the steps of a forest multi-site point cloud registration method as claimed in any of claims 1-6.
  10. 10. A computer program product comprising a computer program, characterized in that the computer program when executed by a processor implements the steps of a forest multi-site point cloud registration method as claimed in any of claims 1-6.

Description

Forest multi-site point cloud registration method and related equipment Technical Field The invention belongs to the technical field of forest multi-site registration, and particularly relates to a forest multi-site point cloud registration method and related equipment. Background With the development of three-dimensional laser scanning technology, ground laser scanning is increasingly widely applied to the services of forest resource checking, forest stand structure analysis, biomass estimation, ecological monitoring and the like. Because of the serious shielding caused by the factors of trees, bushes, topography fluctuation and the like in the forest, the point cloud collected by a single site often has the problems of limited visual angle, serious shielding, partial missing and the like, and the spatial structure of the forest stand is difficult to be completely reflected. Therefore, in actual operation, a plurality of scanning stations are usually required to be arranged on the same forest land, multi-site cloud data are acquired from different directions, and then the multi-site clouds are registered under a unified coordinate system to form a continuous and complete three-dimensional point cloud model, so that the works such as single wood extraction, tree height and chest diameter estimation, forest stand structure statistics and the like can be carried out later. For forest multi-site cloud data processing, multi-site cloud registration is a key link, and aims to automatically calculate the relative pose among all scanning stations under the condition of no or few artificial targets, so that the structures such as trunks, crowns and ground observed by different stations are aligned as accurately as possible under a unified coordinate system. At present, the common technical routes for forest multi-site cloud registration are mainly characterized in that one method takes trunks as main features, the positions and the shapes of trunks are extracted from site clouds through geometric rules or model fitting, then spatial geometric relations among trunks are utilized for feature matching, initial transformation among sites is estimated, image optimization and an ICP algorithm are assisted to complete multi-site registration, another method is used for attempting to take ground as a main registration object, the ground is divided into a plurality of areas through filtering and fitting to extract ground points, local geometric features are extracted, overlapping ground areas are searched among the multi-site clouds, then fine registration is performed through an ICP algorithm from point to face, and the other methods directly follow point cloud registration thought under a general scene, extract local geometric features or three-dimensional descriptors on the forest point clouds, obtain coarse registration through feature matching and RANSAC, and finish fine registration through standard ICP. However, the existing forest multi-site cloud registration method still has obvious defects in the actual complex forest environment. Firstly, the existing method often depends on a single type of registration primitive, for example, only a trunk or only the ground is used, the effect is better in a scene that the trunk is obvious and the ground is exposed well, but under the conditions that the trunk or the ground is heavy and blocked in various forests and different seasons and in the forests, the number, the integrity and the visibility of trunk points or ground points can be obviously reduced, so that the feature extraction is unstable and the available features are insufficient, thereby the registration result is unreliable and even the registration cannot be completed. Secondly, the existing method generally does not perform unified modeling and fusion utilization on multiple types of semantic objects (such as trunks, floors, bushes, thick branches, manual facilities and the like) in the point cloud, only one type or a few types of semantic objects are usually selected as registration objects, and a mechanism for automatically selecting and weighting registration areas according to different semantic structure stability and importance is lacked. Once the preselected features are of poor quality in certain standing or certain stand conditions, the registration process can significantly degrade and lack adaptive adjustment capabilities. Again, in the coarse registration stage, many methods rely on matching of local geometry or local feature descriptors to estimate initial relative pose, when the inter-site view angle difference is large, the horizontal rotation angle is unknown, the forest structure has certain repeatability, the local features are easy to be confused, the erroneous matching is difficult to be completely eliminated, and the coarse registration has strong dependence on the initial pose and the artificial experience, and has poor stability on different data sets. In addition, in the fine registratio