Forest multi-site point cloud registration method and related equipment

Abstract

The invention relates to the technical field of forest multi-site registration, and discloses a forest multi-site point cloud registration method and related equipment, and aims to construct a feature grid map by fusing multiple classes of semantic tags, avoid registration failure caused by single feature deletion or poor quality, improve the adaptability in multi-forest, multi-season and multi-shielding scenes, and improve the registration accuracy. A high-stability registration area is selected by means of stability scores, self-adaptive adjustment of registration constraints is achieved in combination with a semantic weighting mechanism, the dependence of coarse registration on an initial pose and artificial experience is reduced, and wrong matching interference is reduced. Through semantic weighting ICP fine registration and global optimization, geometric stability differences of different semantic objects are distinguished, unstable structure interference is reduced, and registration precision is improved. Automatic stable registration can be achieved without manual target, the field operation cost and workload are reduced, and large-scale forest scene popularization is facilitated.

Description

Technical Field

[0001] This invention belongs to the field of forest multi-site registration technology, specifically a forest multi-site point cloud registration method and related equipment. Background Technology

[0002] With the development of 3D laser scanning technology, terrestrial laser scanning is increasingly widely used in forest resource inventory, stand structure analysis, biomass estimation, and ecological monitoring. However, due to severe shading caused by trees, shrubs, and topographical undulations in forests, point clouds collected from a single site often suffer from limited viewing angles, severe shading, and localized gaps, making it difficult to fully reflect the spatial structure of the forest stand. Therefore, in practice, it is usually necessary to deploy multiple scanning stations in the same forest area to acquire multi-site point cloud data from different directions. These multi-site points are then registered to a unified coordinate system to form a continuous and complete 3D point cloud model, which is essential for subsequent tasks such as individual tree extraction, tree height and diameter at breast height estimation, and stand structure statistics.

[0003] In terms of forest multi-site cloud data processing, multi-site cloud registration is a key step. Its goal is to automatically obtain the relative poses between each scanning station without or with little human target, so that the tree trunks, canopies and ground structures observed at different stations are aligned as accurately as possible in a unified coordinate system. Currently, the commonly used technical approaches for multi-site cloud registration in forests can be roughly categorized as follows: One approach uses tree trunks as the primary feature, extracting the position and shape of tree trunks from each site cloud through geometric rules or model fitting, then using the spatial geometric relationships between tree trunks for feature matching, estimating the initial transformation between sites, and supplementing with graph optimization and ICP algorithms to complete multi-site registration; Another approach attempts to use the ground as the primary registration object, extracting ground points through filtering and fitting, dividing the ground into several regions and extracting local geometric features, searching for overlapping ground regions between multi-site clouds, and then using the point-to-surface ICP algorithm for fine registration; Some other methods directly follow the point cloud registration approach in general scenarios, extracting local geometric features or 3D descriptors from forest point clouds, using feature matching and RANSAC to obtain coarse registration, and then using standard ICP to complete fine alignment.

[0004] However, the existing multi-site cloud registration methods for forests still have significant shortcomings in real-world complex forest environments. First, existing methods often rely on a single type of registration primitive, such as using only tree trunks or only ground. While these methods work well in scenarios where tree trunks are clearly visible and the ground is well exposed, they can significantly reduce the number, completeness, and visibility of tree trunk or ground points in various forest types, different seasons, or when there is dense undergrowth and severe ground obstruction. This leads to unstable feature extraction and insufficient usable features, resulting in unreliable registration results or even failure to complete registration.

[0005] Secondly, existing methods generally fail to uniformly model and integrate multiple semantic objects (such as tree trunks, ground, shrubs, branches, and artificial structures) in point clouds. They typically select only one or a few types as registration objects, lacking a mechanism for automatically selecting and weighting registration regions based on the stability and importance of different semantic structures. If the pre-selected features are of poor quality at certain stations or in certain forest stands, the registration process will significantly degrade, lacking adaptive adjustment capabilities.

[0006] Furthermore, in the coarse registration stage, many methods rely on the matching of local geometric structures or local feature descriptors to estimate the initial relative pose. When there are large differences in viewpoints between sites, the horizontal rotation angle is unknown, and the forest structure has a certain degree of repetition, local features are easily confused, and incorrect matches are difficult to completely eliminate. This results in coarse registration being highly dependent on the initial pose and human experience, and having poor stability on different datasets.

[0007] Furthermore, in the fine registration stage, traditional ICP and its variants typically treat all point clouds involved in registration equally, applying the same error metric and weight to all points. This fails to differentiate between semantic objects such as tree trunks, ground, shrubs, and twigs in terms of geometric stability and temporal variation, making them susceptible to interference from irregular and changeable structures like shrubs and twigs, thus affecting the final registration accuracy. To improve registration reliability, some methods require the deployment of artificial reflective targets or obvious markers on-site, which increases the workload and cost of fieldwork, hindering their widespread adoption in large-scale forest scenarios.

[0008] In summary, existing forest multi-site cloud registration technologies suffer from poor adaptability, strong dependence on single features, sensitivity of coarse registration to initial pose, and susceptibility to unstable structural interference in complex environments with multiple forest types, seasons, and shading. As a result, it is difficult to stably and accurately complete forest multi-site cloud registration under conditions that do not require manual targets and have a high degree of automation. Summary of the Invention

[0009] This invention provides a method and related equipment for forest multi-site point cloud registration, which solves the problem that existing forest multi-site cloud registration technologies are difficult to stably and accurately complete forest multi-site cloud registration under conditions that do not require manual targets and have a high degree of automation.

[0010] To achieve the above objectives, the present invention provides the following technical solution: A method for multi-site point cloud registration in forests includes: Acquire point cloud data from multiple forest sites and perform preprocessing; Semantic segmentation is performed on the preprocessed point cloud data of each site to obtain multiple semantic labels; A regular grid is established on the horizontal plane, and a multi-semantic feature grid map is built based on the multi-class semantic labels within each grid cell. Calculate the stability score of each grid cell, and select highly stable registration regions based on the stability score; Coarse registration is performed based on multi-semantic feature raster maps to estimate the initial rigid body transformation between stations; Based on coarse registration, fine registration is performed using the semantically weighted ICP algorithm; The set of site pairs with effective overlap is determined. The pose of each site is used as a variable, and the fine registration transformation of the site pairs is used as a constraint. The constraint is weighted by combining statistical information of highly stable semantic regions. A global error function is constructed and optimized to obtain a globally consistent multi-site cloud registration result.

[0011] Preferably, acquiring multi-site point cloud data of the forest and performing preprocessing includes: Acquire multi-site point cloud data of the forest using a scanner; Retain points within the scanner's measurement range of 1m≤d≤50~150m, and discard points outside this range; Isolated noise points are removed by using a neighborhood point count threshold or statistical filtering. Compress the data using a voxel grid with a side length of 0.02~0.1m to balance the density; Choose any station as a reference station or use an external coordinate system to align the z-axis with the direction of gravity by simple plane fitting or using the information from the instrument’s built-in tilt sensor.

[0012] Preferably, the steps of establishing a regular grid on a horizontal plane and building a multi-semantic feature grid map based on multiple semantic labels within each grid cell are as follows: A regular grid is constructed for each station on a plane with a unified coordinate system, and the point cloud is projected onto the horizontal plane for grid statistics. Set the grid size for each station, and divide the grid into regular grids along the x and y directions with the grid size as the step size; identify each grid cell; For each grid cell, collect all points that fall within the area of ​​that grid cell, and count the number of points in each semantic category, the percentage of points in each semantic category, and the height statistical and geometric features; Based on the number of points in each semantic category, the proportion of points in each semantic category, high statistical features, and geometric features, a multidimensional semantic feature vector of the raster unit is formed, and a multi-semantic feature raster map is formed based on the multidimensional semantic feature vector of each raster unit.

[0013] Preferably, the step of calculating the stability score of each grid cell and selecting a highly stable registration region based on the stability score is as follows: A scoring formula is constructed by combining the proportion of tree trunks, ground surface area, shrubland area, and height variation:

[0014] in, The percentage of tree-like points within a grid. The percentage of ground-type points within a grid. The percentage of shrub-type sites, denoted as the standard deviation of height within the grid, and a1, a2, a3, and a4 are weighting parameters; After calculating the score for all rasters at each site, one approach is to set a stability threshold and select only rasters with scores greater than the stability threshold as highly stable semantic rasters. Another approach is to sort all rasters from largest to smallest stability score and select rasters as highly stable semantic rasters proportionally, where the raster proportion is set to 20% to 50% depending on the actual scenario.

[0015] Preferably, the steps for coarse registration based on multi-semantic feature raster maps to estimate the initial rigid body transformation between stations are as follows: Taking two sites with overlapping areas as the processing objects, one site is first selected as the reference site and the other as the target site. Then, for the multi-semantic feature raster image of the target site, candidate rotation angles are searched around the z-axis in the range [-180°, 180°] with a step size of 1°. For each candidate rotation angle, the target site raster image is first rotated around the z-axis by that angle, and then its horizontal translation vector relative to the reference site is estimated in the xy plane. Subsequently, the rotated and translated target site raster image is aligned with the reference site raster image, and overlapping raster pairs with the same index are selected. For each pair of overlapping graticles, the multidimensional semantic feature vector and stability score of each pair of overlapping graticles are obtained. The similarity between the feature vectors is calculated, and the similarity is weighted and accumulated according to the stability scores of the two graticles to obtain the overall matching score of the combination of rotation angle and translation vector. After traversing all candidate rotation angle and translation vector combinations, the parameter with the highest matching score is selected as the initial transformation parameter of the plane. Finally, the vertical translation is estimated by using the median or average height of ground-type points in the two stations, and finally a complete coarse registration rigid body transformation including the rotation angle around the z-axis and the three-dimensional translation vector is formed.

[0016] Preferably, the steps for fine registration based on semantically weighted ICP algorithm on the basis of coarse registration are as follows: Based on the initial alignment results obtained from coarse registration, the overlapping area point set of the two site clouds is first extracted. Points participating in registration are then selected based on semantic labels. Tree trunk and ground-type points are retained as constraint objects, and thick branch points are retained as auxiliary constraints. Shrub and fine branch points are downsampled or directly removed. Registration weights are assigned to points of different semantic categories, with tree trunk points having the highest weight, followed by ground-type points, then thick branch points, and shrub points having the lowest weight or set to 0. Weights are set as fixed constants or adjusted based on the stability score of the corresponding raster. Within the ICP framework, differentiated error measures are used for different semantic categories. For tree trunk points, a cylindrical model or trunk axis is fitted within the local neighborhood, and the error is defined as the shortest distance from the point to the cylinder or axis. For ground-type points, a ground plane is fitted locally, and the error is defined as the distance from the point to the plane. For thick branch points and other untreated points, the standard point-to-point Euclidean distance is used as the error. Subsequently, an ICP objective function containing semantic weights is constructed, and the following operations are performed iteratively: Under the current transformation, search for the nearest point or corresponding geometric entity in the target point cloud for each source point, minimize the objective function according to the current correspondence, update the rotation matrix and translation vector, repeat the process until the change in the objective function is less than the preset threshold or the maximum number of iterations is reached, and finally obtain the fine registration transformation between sites.

[0017] A forest multi-site cloud registration system includes: Data acquisition module: Used to acquire point cloud data from multiple forest sites and perform preprocessing; Tag module: Used to perform semantic segmentation on the preprocessed point cloud data of each site to obtain multiple semantic tags; Raster image creation module: used to create regular graticles on a horizontal plane, and to create multi-semantic feature raster images based on multiple semantic labels within each raster cell; Scoring module: Used to calculate the stability score of each grid cell and select highly stable registration regions based on the stability score; Coarse registration module: used for coarse registration based on multi-semantic feature raster maps, and to estimate the initial rigid body transformation between stations; Fine registration module: Used to perform fine registration based on semantically weighted ICP algorithm on the basis of coarse registration; The registration module is used to determine the set of site pairs with effective overlap. It uses the pose of each site as a variable and the fine registration transformation of the site pairs as a constraint. It combines statistical information of highly stable semantic regions to weight the constraints, constructs a global error function and optimizes the solution to obtain a globally consistent multi-site cloud registration result.

[0018] A computer device includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of a forest multi-site point cloud registration method.

[0019] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a forest multi-site point cloud registration method.

[0020] A computer program product includes a computer program that, when executed by a processor, implements the steps of a forest multi-site point cloud registration method.

[0021] Compared with existing technologies, this invention has the following advantages: This invention provides a forest multi-site point cloud registration method. By fusing multiple semantic labels to construct a feature raster map, it avoids registration failures caused by missing or poor-quality single features, improving adaptability in multi-forest, multi-season, and multi-occlusion scenarios. It selects highly stable registration regions using stability scoring and achieves adaptive adjustment of registration constraints through a semantic weighting mechanism, reducing the dependence of coarse registration on initial pose and human experience, and minimizing erroneous matching interference. Through semantically weighted ICP fine registration and global optimization, it distinguishes the geometric stability differences of different semantic objects, reduces interference from unstable structures, and improves registration accuracy. Automated and stable registration can be achieved without manual targets, reducing field operation costs and workload, and facilitating large-scale promotion in forest scenarios. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart of a forest multi-site point cloud registration method according to an embodiment of the present invention; Figure 2 This is a block diagram of a forest multi-site point cloud registration system according to an embodiment of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0025] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0026] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0027] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.

[0028] like Figure 1 As shown, this invention provides a forest multi-site point cloud registration method, including the following steps: S1: Acquire point cloud data from multiple forest sites and perform preprocessing; S2: Perform semantic segmentation on the preprocessed point cloud data of each site to obtain multiple semantic labels; S3: Establish a regular grid on the horizontal plane, and build a multi-semantic feature grid map based on the multi-class semantic labels in each grid cell; S4: Calculate the stability score of each grid cell and select a highly stable registration region based on the stability score; S5: Coarse registration is performed based on multi-semantic feature raster maps to estimate the initial rigid body transformation between stations; S6: Based on coarse registration, fine registration is performed using the semantically weighted ICP algorithm; S7: Determine the set of site pairs with effective overlap, use the pose of each site as a variable, the fine registration transformation of the site pairs as a constraint, combine the statistical information of highly stable semantic regions to weight the constraints, construct a global error function and optimize the solution to obtain a globally consistent multi-site cloud registration result.

[0029] By fusing multiple semantic labels to construct a feature raster map, registration failures caused by missing or poor-quality single features are avoided, improving adaptability in multi-forest, multi-season, and multi-occlusion scenarios. Highly stable registration regions are selected using stability scoring, and a semantic weighting mechanism is combined to adaptively adjust registration constraints, reducing the dependence of coarse registration on initial pose and human experience, and minimizing erroneous matching interference. Semantic weighted ICP fine registration and global optimization differentiate the geometric stability differences of different semantic objects, reducing interference from unstable structures and improving registration accuracy. Automated stable registration can be achieved without manual targets, reducing field operation costs and workload, and facilitating large-scale promotion in forest scenarios.

[0030] The detailed steps are as follows: S1: Acquire point cloud data from multiple forest sites and perform preprocessing.

[0031] Specifically, in this embodiment, several ground-based laser scanning stations are deployed within the target forest area to scan the same forest area from different directions, acquiring 3D point cloud data for each station. The original point cloud of each station contains at least the 3D coordinate information of the points. It can also include attributes such as reflection intensity and scanning angle.

[0032] After acquiring the multi-site cloud, the following preprocessing operations are performed on each site cloud in sequence: 1) Distance cropping: Based on the scanner's effective range and the forest area, remove points that are too close or too far away. For example, only retain points from the distance scanner. exist points within, You can take 1m. The setting can be 50-150m depending on the instrument's performance and the size of the forest.

[0033] 2) Noise Reduction Filtering: For each point, count the number of its neighbors within a certain radius (e.g., 0.1–0.3 m); when the number of neighbors is less than a preset threshold, the point is considered an isolated point or a noise point and is deleted. Alternatively, a statistical filtering method can be used to remove noise based on whether the distance between the point and the mean of its neighborhood exceeds a certain number of standard deviations.

[0034] 3) Voxel downsampling: Construct a regular voxel grid in three-dimensional space. The voxel side length can be set to 0.02~0.1m. Replace multiple points falling in the same voxel with a representative point to reduce the amount of data and balance the point density.

[0035] 4) Coordinate standardization: Select any station as a reference station, or use an external coordinate system to align the z-axis of each station cloud with the gravity direction as much as possible through simple plane fitting or by using the information from the built-in tilt sensor of the instrument, so as to ensure the consistency of all station clouds in the vertical direction and reduce the degree of freedom in subsequent registration.

[0036] S2: Perform semantic segmentation on the preprocessed point cloud data of each site to obtain multiple semantic labels.

[0037] Specifically, in this embodiment, the point cloud is divided into semantic categories, including tree trunks, ground, shrubs or understory vegetation, thick branches, and artificial structures, based on the geometric and height features of the forest point cloud.

[0038] One feasible implementation method is: 1) Ground Class Recognition: For each site cloud, a ground filtering method is used to identify ground points. Specifically, the point cloud can be divided into several height layers along the z-axis, and a local plane is fitted layer by layer. Points higher than a certain height threshold of the plane are removed, and finally, the near-horizontal continuous points on the lower side are retained as ground class points.

[0039] 2) Tree Trunk Candidate Point Identification: Calculate the local normal vector for non-ground points and compare it with the angle between the normal vector and the vertical direction (0,0,1). When the angle is less than a preset threshold (e.g., 15°) and the point's height is within a certain range (e.g., 0.5–10m), it is marked as a tree trunk candidate point. Further, cylindrical fitting or Hough transform can be performed on the tree trunk candidate points in the horizontal plane to filter out noise points that do not conform to the tree trunk shape.

[0040] 3) Identification of shrub or understory vegetation: For non-ground points below a certain height (e.g., 2m) and with a relatively messy distribution of normal vectors and poor local planarity, they are marked as shrub or understory vegetation points.

[0041] 4) Identification of coarse branches and artificial structures: Points with high height and whose normal vectors are neither close to vertical nor horizontal and are distributed at an angle can be marked as coarse branch candidate points; local structures with regular spatial distribution, smooth surface and obvious differences from the shape of natural trees can be marked as artificial structure points in combination with pre-set geometric templates.

[0042] In another implementation, the preprocessed point cloud can be input into a pre-trained point cloud semantic segmentation deep learning model (such as PointNet++, KPConv, etc.), and the corresponding semantic category label can be directly output for each point. The rule-based segmentation method and the deep learning-based semantic segmentation method described above can be used alone or in combination, that is, the rule-based method can be used for coarse classification first, and then the deep learning model can be used for refinement.

[0043] S3: Establish a regular grid on the horizontal plane, and build a multi-semantic feature grid map based on the multi-class semantic labels in each grid cell.

[0044] Specifically, in this embodiment, a regular grid is constructed for each station on the x–y plane of a unified coordinate system, and the point cloud is projected onto the horizontal plane for grid statistics.

[0045] 1) Grid division: Set the grid size for each site. (e.g., 0.5m or 1.0m), along the x and y directions. Divide the grid into regular grids according to the step size. Each grid cell uses... A unique identifier, corresponding to the index in the x and y directions.

[0046] 2) Semantic and height statistics: For each raster cell, collect all points falling within the area of ​​that raster cell and perform statistics: The number of points for each semantic category, such as the number of points for the trunk category. Ground type points shrub type points coarse-branch type points wait; The percentage of points in each semantic category, such as the percentage of tree trunks. Ground area shrubland proportion ,in This represents the total number of points within the grid. Height statistics, such as the mean height of all points within a grid. Maximum height Standard deviation of height ; Geometric features, such as local planarity indices (which can be calculated from the eigenvalues ​​of the local covariance matrix) and roughness.

[0047] 3) Feature Vector Construction: The above statistics are combined to form a multidimensional semantic feature vector for this raster unit. For example, this embodiment can take the following form:

[0048] In other embodiments, it may be extended to include Features such as planarity and roughness are considered. Through the above steps, a multi-semantic feature raster image can be obtained for each site, with each raster cell corresponding to a multi-dimensional semantic feature vector.

[0049] This invention performs semantic segmentation on multi-station TLS point clouds in forests and performs unified statistical modeling of various semantic objects such as tree trunks, ground, and shrubs at the horizontal raster scale to construct a multi-dimensional semantic feature raster map. Simultaneously, it introduces a semantic stability score and automatically selects highly stable raster areas as registration regions. This approach overcomes the limitations of existing methods that rely solely on single features like tree trunks or ground. When the quantity or quality of a certain feature is insufficient under certain forest types or collection conditions, effective constraints can still be obtained from other semantic structures. Furthermore, by actively excluding unstable regions such as shrubs and drastically undulating areas through stability scoring, registration prioritizes raster areas with abundant tree trunks and clear ground. Therefore, even under conditions of multiple forest types, multiple seasons, and complex understory vegetation, it can automatically select reliable registration regions, significantly improving the applicability and overall robustness of the method.

[0050] S4: Calculate the stability score of each grid cell and select a highly stable registration region based on the stability score.

[0051] Specifically, in this embodiment, a semantic stability score is introduced to measure the reliability of each grid cell in the registration process. Calculate a stability score for each grid cell: 1) Semantic stability scoring model: Combining the proportion of tree trunks, ground surface area, shrubland area, and height fluctuation, a scoring formula is constructed, for example:

[0052] in, The percentage of tree-like points within a grid. The percentage of ground-type points within a grid. The percentage of shrub-type sites, The standard deviation of height within the grid is denoted as a, and a1, a2, a3, and a4 are weighting parameters. In this embodiment, they can be taken as positive numbers, and a1 and a2 are greater than a3 and a4, which are used to highlight the positive contribution of the tree trunk and the ground to stability.

[0053] 2) High stability region selection: Calculate for all grid cells at each site. Alternatively, one approach is to set a stability threshold. Select only One approach is to use a raster as a highly stable semantic raster; another approach is to classify all rasters according to... Sort the rasters from largest to smallest and select the top p% as highly stable semantic rasters, where p can be set to 20%–50% depending on the actual scenario. These highly stable semantic rasters constitute the highly stable registration region of the site and will be given higher weight in subsequent coarse and fine registration.

[0054] S5: Coarse registration is performed based on multi-semantic feature raster maps to estimate the initial rigid body transformation between stations.

[0055] Specifically, this embodiment takes any pair of stations A and B with overlapping areas as an example to illustrate the coarse registration process: 1) Reference and target station determination: Select station A as the reference station and station B as the target station. Use steps S1 to S4 to obtain the multi-semantic feature raster images of stations A and B and the stability score of each raster.

[0056] 2) Rotation Angle Search: Considering the characteristic of the ground laser scanner rotating around the vertical direction to collect data, this embodiment mainly searches for the rotation angle around the z-axis. The yaw angle search range is set as follows: ,For example 180°, step size A value of 1° forms the candidate angle set. .

[0057] For each candidate corner : Rotate the multi-semantic feature raster image of site B around the z-axis. The rotated raster image B is obtained. ); For B( on the x–y plane) Translation vector relative to station A Perform sampling or quickly estimate possible translational shifts through cross-correlation; For each candidate translation , will B ( After translating and aligning with A, find overlapping raster pairs with the same raster index. ), obtain and corresponding feature vector and and stability score and ; calculate and Feature similarity Similarity can be measured using methods such as cosine similarity or negative Euclidean distance; The similarity of each overlapping grid pair is weighted and summed according to the stability score to obtain the result. Overall matching score of the combination ,For example:

[0058] 3) Determining coarse registration parameters: For all candidate parameters... calculate Then, select Largest parameter combination This serves as the initial planar transformation of station B relative to station A.

[0059] 4) Vertical Translation Estimation: After obtaining the initial planar transformation, the vertical translation can be estimated using the height information of ground-like points in the cloud at both stations. Specifically, station B is aligned to the coordinate system of station A under the initial planar transformation, and the median or average height of the ground-like points at both stations is calculated. The difference between the two can be used as the vertical translation. Thus, a complete coarse registration rigid body transformation is obtained. It includes the rotation angle θ about the z-axis and the three-dimensional translation vector. .

[0060] A joint search for rotation angles and translations is performed on a multi-semantic feature raster map, using feature similarity between highly stable semantic rasters as the scoring criterion to achieve coarse registration between stations. Because a global search for candidate yaw angles and horizontal translations is conducted at the raster level, rather than relying on matching a few local trunk triangles or local geometric features, the optimal coarse registration transformation can be automatically found through the scoring function even if there are significant viewpoint differences or unknown rotation angles between stations. This effectively overcomes the shortcomings of traditional methods in the coarse registration stage, which rely heavily on initial poses and human experience and are prone to getting trapped in local optima. It reduces the need for human intervention and improves the success rate and stability of the coarse registration stage.

[0061] S6: Based on coarse registration, fine registration is performed using the semantically weighted ICP algorithm.

[0062] Specifically, after coarse registration and alignment of site A and site B, this embodiment further performs fine registration at the point level to improve alignment accuracy.

[0063] 1) Registration point set selection: Based on the spatial location after coarse registration, the overlapping area point set of the two site clouds is extracted, and points participating in fine registration are selected based on the semantic labels of S2: Tree trunk points and ground points are the primary constraint objects to be retained. You can choose to retain coarse-branch points as auxiliary constraints; Significantly downsample or directly remove shrub and branch types to reduce interference from unstable structures.

[0064] 2) Semantic weight allocation: Define registration weights for points of different semantic categories. ,For example: Tree trunk type point weights maximum; Ground point weights Next; coarse-branch point weights Smaller; Shrub class point weights Minimum or set to 0.

[0065] The weights can be fixed constants or fine-tuned by referring to the stability score of the corresponding raster.

[0066] 3) Differentiated Error Measurement: In the ICP framework, different error forms are used for different semantic categories: For tree trunk-like points, a cylindrical model of the tree trunk or the trunk axis can be fitted within a local neighborhood, and the error can be defined as the shortest distance from the point to the cylinder or the point to the axis. For ground-like points, a ground plane can be locally fitted, and the error can be defined as the distance from the point to the plane; For coarse branches and other points that have not undergone special treatment, the standard point-to-point Euclidean distance can be used as the error.

[0067] 4) Semantically Weighted ICP Iteration: Construct an ICP objective function that includes semantic weights, for example:

[0068] Where R and t are the rotation matrix and translation vector to be solved, respectively. Let be the error of the i-th point or point pair under the current transformation. The following steps are performed iteratively: Under the current transformation, search for the nearest point or corresponding geometric entity (cylinder, plane, etc.) in the target point cloud for each source point. Based on the current correspondence, minimize ,renew and ; Repeat the above process until... The change is less than the preset threshold or the maximum number of iterations is reached.

[0069] Finally, the fine registration transformation of station B relative to station A is obtained. .

[0070] In the fine registration stage, a semantically weighted iterative nearest-point model is introduced, assigning different registration weights to point clouds of different semantic categories, such as tree trunks, ground, thick branches, and shrubs. Different error measures are used for tree trunks, such as the distance from a point to a cylinder or axis, and for the ground, the distance from a point to a plane. This design ensures that geometrically stable and regularly shaped objects like tree trunks and ground contribute more constraint during registration, while irregularly shaped and temporally variable objects like shrubs and thin branches are weakened or eliminated. Since the registration results are mainly influenced by stable structures with minimal variation and good repeatability across different sites, the overall registration result is insensitive to noise and local deformation. This fundamentally alleviates the problem of traditional ICP treating all points equally and being easily "distorted" by unstable vegetation structures, significantly improving the accuracy and robustness of fine registration.

[0071] S7: Determine the set of site pairs with effective overlap, use the pose of each site as a variable, the fine registration transformation of the site pairs as a constraint, combine the statistical information of highly stable semantic regions to weight the constraints, construct a global error function and optimize the solution to obtain a globally consistent multi-site cloud registration result.

[0072] Specifically, in the case of multiple scanning stations (e.g., N stations), this embodiment eliminates error accumulation through global optimization: 1) Site Pair Set Determination: Based on the site deployment and point cloud overlap information, determine which site pairs have effective overlapping areas and construct a site pair set. For each Performing steps S5 and S6 yields a relatively fine registration transformation. The site also compiled statistics on the number of medium-to-high stability semantic rasters and their average stability scores.

[0073] 2) Global pose parameterization: Select one station as the reference station and fix its pose as a unit transformation, and then parameterize the poses of the other stations. As variables to be solved, each This is a rigid body transformation.

[0074] 3) Construction of global error function: For each station pair , will be and Calculated relative transformation The result of fine registration By comparing and defining constraint errors, and combining the semantic stability statistics of the site pair, constraint weights are given. For example, construct the global error function in least squares form:

[0075] Where ||·|| is a function that measures the difference between two rigid body transformations, which can map rigid body transformations to rotation and translation parameter vectors and calculate the Euclidean distance.

[0076] 4) Global Optimization Solution: Numerical optimization methods such as nonlinear least squares are employed to solve for all unknowns. Perform joint optimization until The coordinates converge, thus obtaining the global pose of each site. After transforming all site clouds to a unified coordinate system based on their global poses, a consistent multi-site cloud registration result is obtained.

[0077] A global optimization model is established using the pose of each scanning station as a variable. The fine registration transformation of each station pair is used as a constraint, and different weights are assigned to each constraint based on the number of highly stable semantic gratings and the average stability score in the station pair. In this way, station pairs with richer semantic information and more stable geometric structures have a greater impact on global optimization, while station pairs with weaker semantic constraints or poor overlap have a smaller impact on the results. This makes the global solution mainly dominated by the reliable local registration results. Compared with simple serial multi-station registration or pose graph optimization that treats all constraints equally, this semantic stability-weighted global optimization strategy can effectively suppress the accumulation and propagation of errors among multiple stations, avoiding situations where local areas are aligned but the overall model is distorted or drifted. This ensures the consistency and global accuracy of multi-station cloud registration results across the entire scope.

[0078] Furthermore, the entire registration process of this invention relies entirely on various semantic features naturally present in the scanning scene, such as tree trunks, ground, and shrubs, eliminating the need for artificial targets like reflector balls or target boards. Through the synergistic effect of semantic segmentation, multi-semantic raster modeling, semantic stability scoring, grid-level coarse registration, semantically weighted fine registration, and multi-station global optimization, automatic registration of forest multi-site cloud data is achieved. Compared to traditional methods that require extensive manual target placement or intervention, this invention significantly reduces fieldwork and on-site deployment costs, while improving the automation and efficiency of data processing. It is more suitable for widespread application in large-scale, multi-batch forest resource surveys and monitoring tasks.

[0079] like Figure 2 As shown, this embodiment of the invention also provides a forest multi-site cloud registration system, including: Data acquisition module: Used to acquire point cloud data from multiple forest sites and perform preprocessing; Tag module: Used to perform semantic segmentation on the preprocessed point cloud data of each site to obtain multiple semantic tags; Raster image creation module: used to create regular graticles on a horizontal plane, and to create multi-semantic feature raster images based on multiple semantic labels within each raster cell; Scoring module: Used to calculate the stability score of each grid cell and select highly stable registration regions based on the stability score; Coarse registration module: used for coarse registration based on multi-semantic feature raster maps, and to estimate the initial rigid body transformation between stations; Fine registration module: Used to perform fine registration based on semantically weighted ICP algorithm on the basis of coarse registration; The registration module is used to determine the set of site pairs with effective overlap. It uses the pose of each site as a variable and the fine registration transformation of the site pairs as a constraint. It combines statistical information of highly stable semantic regions to weight the constraints, constructs a global error function and optimizes the solution to obtain a globally consistent multi-site cloud registration result.

[0080] Each module of the system can be implemented as a software functional module or as a combination of hardware and software. The modules are connected through a data interface and are called sequentially according to the above method steps to complete the automatic registration process of the forest multi-site cloud, enabling those skilled in the art to implement the present invention based on the contents of this specification.

[0081] A computer device is provided according to an embodiment of the present invention. This computer device includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps in the various method embodiments described above. Alternatively, when the processor executes the computer program, it implements the functions of each module / unit in the various device embodiments described above.

[0082] The computer program can be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention.

[0083] The computer device may be a desktop computer, laptop, handheld computer, or cloud server, etc. The computer device may include, but is not limited to, a processor and memory.

[0084] The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0085] The memory can be used to store the computer program and / or module, and the processor implements various functions of the computer device by running or executing the computer program and / or module stored in the memory, and by calling the data stored in the memory.

[0086] If the modules / units integrated into the computer device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory, random access memory, electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0087] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the scope of the invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0088] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can be appropriately combined to form other embodiments that can be understood by those skilled in the art. The above content is only for illustrating the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made based on the technical concept proposed in this invention shall fall within the scope of protection of the claims of this invention.

Claims

1. A method for registering point clouds across multiple forest sites, characterized in that, include: Acquire point cloud data from multiple forest sites and perform preprocessing; Semantic segmentation is performed on the preprocessed point cloud data of each site to obtain multiple semantic labels; A regular grid is established on the horizontal plane, and a multi-semantic feature grid map is built based on the multi-class semantic labels within each grid cell. Calculate the stability score of each grid cell, and select highly stable registration regions based on the stability score; Coarse registration is performed based on multi-semantic feature raster maps to estimate the initial rigid body transformation between stations; Based on coarse registration, fine registration is performed using the semantically weighted ICP algorithm; The set of site pairs with effective overlap is determined. The pose of each site is used as a variable, and the fine registration transformation of the site pairs is used as a constraint. The constraint is weighted by combining statistical information of highly stable semantic regions. A global error function is constructed and optimized to obtain a globally consistent multi-site cloud registration result.

2. The forest multi-site point cloud registration method according to claim 1, characterized in that, Acquire point cloud data from multiple forest sites and perform preprocessing, including: Acquire multi-site point cloud data of the forest using a scanner; Retain points within the scanner's measurement range of 1m≤d≤50~150m, and discard points outside this range; Isolated noise points are removed by using a neighborhood point count threshold or statistical filtering. Compress the data using a voxel grid with a side length of 0.02~0.1m to balance the density; Choose any station as a reference station or use an external coordinate system to align the z-axis with the direction of gravity by simple plane fitting or using the information from the instrument’s built-in tilt sensor.

3. The forest multi-site point cloud registration method according to claim 1, characterized in that, The specific steps for establishing a regular grid on a horizontal plane and creating a multi-semantic feature grid map based on multiple semantic labels within each grid cell are as follows: A regular grid is constructed for each station on a plane with a unified coordinate system, and the point cloud is projected onto the horizontal plane for grid statistics. Set the grid size for each station, and divide the grid into regular grids along the x and y directions with the grid size as the step size; identify each grid cell; For each grid cell, collect all points that fall within the area of ​​that grid cell, and count the number of points in each semantic category, the percentage of points in each semantic category, and the height statistical and geometric features; Based on the number of points in each semantic category, the proportion of points in each semantic category, high statistical features, and geometric features, a multidimensional semantic feature vector of the raster unit is formed, and a multi-semantic feature raster map is formed based on the multidimensional semantic feature vector of each raster unit.

4. The forest multi-site point cloud registration method according to claim 1, characterized in that, The specific steps for calculating the stability score of each grid cell and selecting highly stable registration regions based on the stability score are as follows: A scoring formula is constructed by combining the proportion of tree trunks, ground surface area, shrubland area, and height variation: in, The percentage of tree-like points within a grid. The percentage of ground-type points within a grid. The percentage of shrub-type sites, denoted as the standard deviation of height within the grid, and a1, a2, a3, and a4 are weighting parameters; After calculating the score for all rasters at each site, one approach is to set a stability threshold and select only rasters with scores greater than the stability threshold as highly stable semantic rasters. Another approach is to sort all rasters from largest to smallest stability score and select rasters as highly stable semantic rasters proportionally, where the raster proportion is set to 20% to 50% depending on the actual scenario.

5. The forest multi-site point cloud registration method according to claim 1, characterized in that, The specific steps for coarse registration based on multi-semantic feature raster maps, and estimation of initial rigid body transformations between sites, are as follows: Taking two sites with overlapping areas as the processing objects, one site is first selected as the reference site and the other as the target site. Then, for the multi-semantic feature raster image of the target site, candidate rotation angles are searched around the z-axis in the range [-180°, 180°] with a step size of 1°. For each candidate rotation angle, the target site raster image is first rotated around the z-axis by that angle, and then its horizontal translation vector relative to the reference site is estimated in the xy plane. Subsequently, the rotated and translated target site raster image is aligned with the reference site raster image, and overlapping raster pairs with the same index are selected. For each pair of overlapping graticles, the multidimensional semantic feature vector and stability score of each pair of overlapping graticles are obtained. The similarity between the feature vectors is calculated, and the similarity is weighted and accumulated according to the stability scores of the two graticles to obtain the overall matching score of the combination of rotation angle and translation vector. After traversing all candidate rotation angle and translation vector combinations, the parameter with the highest matching score is selected as the initial transformation parameter of the plane. Finally, the vertical translation is estimated by using the median or average height of ground-type points in the two stations, and finally a complete coarse registration rigid body transformation including the rotation angle around the z-axis and the three-dimensional translation vector is formed.

6. The forest multi-site point cloud registration method according to claim 1, characterized in that, Based on coarse registration, the specific steps for fine registration using the semantically weighted ICP algorithm are as follows: Based on the initial alignment results obtained from coarse registration, the overlapping area point set of the two site clouds is first extracted. Points participating in registration are then selected based on semantic labels. Tree trunk and ground-type points are retained as constraint objects, and thick branch points are retained as auxiliary constraints. Shrub and fine branch points are downsampled or directly removed. Registration weights are assigned to points of different semantic categories, with tree trunk points having the highest weight, followed by ground-type points, then thick branch points, and shrub points having the lowest weight or set to 0. Weights are set as fixed constants or adjusted based on the stability score of the corresponding raster. Within the ICP framework, differentiated error measures are used for different semantic categories. For tree trunk points, a cylindrical model or trunk axis is fitted within the local neighborhood, and the error is defined as the shortest distance from the point to the cylinder or axis. For ground-type points, a ground plane is fitted locally, and the error is defined as the distance from the point to the plane. For thick branch points and other untreated points, the standard point-to-point Euclidean distance is used as the error. Subsequently, an ICP objective function containing semantic weights is constructed, and the following operations are performed iteratively: Under the current transformation, search for the nearest point or corresponding geometric entity in the target point cloud for each source point, minimize the objective function according to the current correspondence, update the rotation matrix and translation vector, repeat the process until the change in the objective function is less than the preset threshold or the maximum number of iterations is reached, and finally obtain the fine registration transformation between sites.

7. A forest multi-site cloud registration system, characterized in that, A forest multi-site point cloud registration method according to any one of claims 1-6 includes: Data acquisition module: Used to acquire point cloud data from multiple forest sites and perform preprocessing; Tag module: Used to perform semantic segmentation on the preprocessed point cloud data of each site to obtain multiple semantic tags; Raster image creation module: used to create regular graticles on a horizontal plane, and to create multi-semantic feature raster images based on multiple semantic labels within each raster cell; Scoring module: Used to calculate the stability score of each grid cell and select highly stable registration regions based on the stability score; Coarse registration module: used for coarse registration based on multi-semantic feature raster maps, and to estimate the initial rigid body transformation between stations; Fine registration module: Used to perform fine registration based on semantically weighted ICP algorithm on the basis of coarse registration; The registration module is used to determine the set of site pairs with effective overlap. It uses the pose of each site as a variable and the fine registration transformation of the site pairs as a constraint. It combines statistical information of highly stable semantic regions to weight the constraints, constructs a global error function and optimizes the solution to obtain a globally consistent multi-site cloud registration result.

8. A computer device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the forest multi-site point cloud registration method according to any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the forest multi-site point cloud registration method according to any one of claims 1-6.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the forest multi-site point cloud registration method according to any one of claims 1-6.

Images