Multi-station ground-based laser scanning point cloud registration method and system suitable for forest environment

Abstract

The application belongs to the technical field of laser scanning, and discloses a multi-station ground-based laser scanning point cloud registration method and system suitable for forest environment. Firstly, the application converts original three-dimensional point clouds to two-dimensional distance images, extracts tree trunks based on continuity constraints, and finds tree trunk root point positions. Secondly, a triangular descriptor is constructed, the spatial relationship between each tree trunk is scanned, and is stored in a hash table. The corresponding triangular pairs are found by searching the hash table, the similarity between scans is calculated, and three-dimensional transformation parameters are extracted. The scan station pose to be solved is taken as a node, and the relative transformation is taken as a factor constraint, so that the overall consistency optimal scan station pose parameters are obtained through factor graph optimization. Finally, the conversion parameters after factor graph optimization are taken as initial values, the ICP method is used for optimization and adjustment, the poses of each scan station in the unified coordinate system are derived through a depth-first strategy, the local point clouds are converted to the unified coordinate system, and global point cloud registration is realized.

Description

Technical Field

[0001] This invention belongs to the field of laser scanning technology, and in particular relates to a multi-station ground-based laser scanning point cloud registration method and system suitable for forest environments. Background Technology

[0002] LiDAR (Light Detection and Ranging) possesses inherent advantages such as speed, high precision, and penetration, enabling it to accurately characterize the three-dimensional spatial structure surrounding the sensor and has been widely used in forestry surveys. The three-dimensional point clouds it acquires can accurately digitize trees and calculate a series of attribute parameters such as biomass, trunk curves, and canopy structure, which is difficult to achieve in traditional plantation forestry surveys.

[0003] Currently, Terrestrial Laser Scanning (TLS) systems include two common scanning methods: monostation scanning and multistation scanning. Multistation scanning involves setting up stations at different locations within a forest plot to acquire point clouds from different perspectives, reducing the occlusion caused by dense foliage and shrubs and covering the entire measurement area. However, the point cloud data in multistation scanning reside in different local coordinate systems, requiring stitching the data from multiple stations into a unified coordinate system for subsequent practical applications. The most common existing point cloud stitching method involves deploying multiple artificial targets within the plot, ensuring that at least three targets are scanned during the operation, and then manually extracting the corresponding targets and calculating the 3D transformation parameters during post-processing. This method is sufficiently accurate and reliable, but it is time-consuming and labor-intensive in both field operations and data processing, resulting in low operational efficiency.

[0004] Based on the different registration primitives, existing TLS point cloud registration methods can be roughly divided into three categories: point-based, feature-based, and object-based.

[0005] Point-based registration methods directly establish corresponding point pairs in the original data, minimizing the distance between all point pairs to obtain the optimal transformation parameters. These methods include typical Iterative Closest Point (ICP), generalized ICP, and globally optimal ICP. These methods can perform fine-tuning to obtain accurate transformation parameters under good initial conditions; however, they are sensitive to initial values, and inappropriate initial transformations can easily lead point-based methods into local optima. Forest scenes have more noisy and isolated points compared to urban buildings. This presents new challenges for point-based registration methods, making it difficult to establish correct point pair relationships. Although advanced methods such as TEASER, GORE, and GROR aim to remove the influence of erroneous point pairs by incorporating a series of robust gross error removal methods, they are still difficult to apply to forest scenes. Furthermore, the massive point cloud scanned by TLS in forest scenes also imposes a huge computational burden on these methods.

[0006] The performance of feature-based registration methods relies on two key steps: feature extraction and feature description. First, representative keypoints are extracted from the original point cloud. Then, feature descriptors are used to describe the local neighborhoods of the keypoints, and assumed associations are established based on the similarity between descriptors. Various feature extraction and description algorithms have been proposed, including hand-designed feature points such as LSP, ISS, and Harris 3D, and popular feature descriptors such as FPFH and SHOT. These hand-designed methods require expert knowledge of the scene and have poor robustness, making them difficult to generalize to ordinary users. Another category is deep learning-based methods, such as PPFNet, D3Feat, and SpinNet, which utilize deep learning networks to mine knowledge about the scene and effectively extract feature points. Although these feature extraction methods can better handle the influence of noisy points, changes in scanning viewpoint, point density, and occlusion effects in forest point clouds can significantly cause hand-designed features to fail. Furthermore, deep learning methods suffer from the problems of requiring large amounts of labeled point cloud data for training, high computational demands, and poor applicability to different scenarios.

[0007] Object-based registration methods use fragments of point clouds as registration primitives to establish corresponding associations. These methods effectively address the influence of noise points, avoid complex feature design, and preserve original geometric information. Tree trunks in forests, as stable and highly representative components, are widely used in various TLS point cloud registration methods. However, large-scale plots with high tree trunk density significantly increase the computational burden of existing methods.

[0008] Based on the above analysis, the problems and shortcomings of the existing technology are as follows:

[0009] The most common point cloud stitching method currently used involves deploying multiple artificial targets within the sample plot, ensuring that at least three targets are scanned during the operation, and then manually extracting the corresponding targets and calculating the 3D transformation parameters during post-processing. This method is time-consuming, labor-intensive, and inefficient in both field operations and data processing. Summary of the Invention

[0010] To address the problems existing in the prior art, this invention provides a multi-station ground-based laser scanning point cloud registration method and system suitable for forest environments.

[0011] This invention is implemented as follows: A multi-station ground-based laser scanning point cloud registration method and system suitable for forest environments includes:

[0012] Step 1: Generate distance image;

[0013] Step 2: Continuity constraint of tree trunk points;

[0014] Apply distance continuity constraints along the column direction to extract the stable portion of the point cloud;

[0015] Step 3: Constrain the orientation of the tree trunk point;

[0016] Even with continuity constraints, some dense branches and scattered ground fragments remain in the extracted tree trunk points. Point clouds are clustered, and then the geometric properties of each cluster are used to remove branches, leaves, and ground points, while retaining the tree trunk points.

[0017] Step 4: Construct the triangular descriptor;

[0018] Triangles are used to describe the relative relationships between discrete tree trunk locations; multiple triangles are combined to form a triangular network to describe the overall distribution of tree trunks in the scan; the number of triangles depends on the size of the scanned plot and the tree trunk density.

[0019] Step 5: Coarse registration of 3D point clouds;

[0020] The algorithm uses efficient hash table search to find triangle matching pairs, calculates the similarity between matching pairs of scanning stations, and outputs the coarse registration parameters of the scan point cloud pairs if the similarity meets the threshold.

[0021] Step 6: Factor plot optimization;

[0022] A factor graph optimization method is introduced to model the inconsistencies between relative transformations at multiple stations, and globally consistent transformation parameters are obtained after overall optimization.

[0023] Step 7: Point cloud fine registration and pose estimation;

[0024] The relative pose transformation parameters optimized by factor graphs are used as initial values ​​and optimized by point-to-point ICP method. The pose of each scanning station in a unified coordinate system is derived by the optimized transformation parameters, and all scanning station point clouds are transformed to a unified coordinate system to form a consistent global point cloud.

[0025] Furthermore, the generation of distance images:

[0026] (1) Ground-based lidar data acquisition;

[0027] TLS scanning stations are deployed according to the size of the forest plot. For a 20m×20m plot, a scanning method of one station in the middle and one station at each of the four corners can be adopted to obtain complete plot scanning data to the greatest extent.

[0028] (2) Converting 3D point clouds into distance images;

[0029] Processing single-site point cloud data, let the three-dimensional coordinates of any point i in the point cloud be (x, y, z). i ,y i ,z i It can also be represented as polar coordinates (α). i ,β i ,γ i The form of ) where α i ,β i These are the horizontal angle and the vertical angle, γ. i Let i be the distance from the scan point i to the scan center; then, the index of the scan point i in the distance image can be defined by formula (1);

[0030]

[0031] In the formula, row i ,col i Here, i represents the row and column indices of the scan point i in the distance image, respectively, and θ represents the angular resolution of the distance image.

[0032] Distance to each pixel in the image using γ i For padding, if multiple scan points correspond to the same index position after conversion, the pixel retains the information of the scan point closest to the center.

[0033] Furthermore, the continuity constraint of the trunk points:

[0034] (1) Define the search step size (Step) and the pixel difference threshold (R). diff ), Group pixel count threshold (Seg th The three parameters are set to 15 pixels, 0.03m, and 100 pixels respectively;

[0035] (2) For each column in the distance image, starting from the first point, compare the pixel value difference between the search point and the current point within the search step size Step. If it is less than the preset threshold R, diff If the search point is in the same group as the current point, then the search point and the current point are grouped together; otherwise, a new group is created.

[0036] (3) Repeat the above operation until all pixels in the distance image have been accessed;

[0037] (4) Count the number of pixels in each group. If it exceeds the preset threshold Seg th If a pixel in a group is not found, it is retained and marked as 1; otherwise, it is discarded and marked as 0.

[0038] (5) Traverse all pixels. If the pixel is marked as 1, keep the corresponding scan point. Otherwise, if the pixel is marked as 0, discard the scan point.

[0039] Furthermore, the tree trunk point direction constraint is as follows:

[0040] (1) Using Euclidean distance clustering, the point cloud data extracted in step 2 is processed and divided into different point cloud clusters;

[0041] (2) Perform principal component analysis (PCA) on each point cloud cluster and calculate the linear L. λ Planarity P λ Dispersion S λ Three parameters describe the geometric distribution of point cloud clusters in space, defined by formula (2);

[0042]

[0043] In the formula, λ1>λ2>λ3 are the eigenvalues ​​of the point cloud cluster, arranged in descending order, which correspond to the degree of change in the direction of each eigenvector;

[0044] (3) The tree trunk points exhibit a linear distribution in space, that is, they have a large linearity L. λ The ground point has a large planarity P λ The branch and leaf points have a large degree of dispersion. λ Therefore, this characteristic can be used to determine the type of extracted point cloud clusters. If a point cloud cluster satisfies condition L... λ If the value is greater than 0.95, then this point cloud cluster is considered to be a tree trunk; the distribution of point cloud clusters before and after applying geometric constraints;

[0045] (4) Using the extracted tree trunk point cloud cluster, extract the corresponding tree trunk position; its planar position (x,y) is the center of the point cloud cluster, and the z component is selected as the root point of the tree trunk, that is, the intersection of the tree trunk and the ground. The black dots at the bottom are the extracted tree trunk positions.

[0046] Furthermore, the construction of the triangular descriptor:

[0047] (1) Filter the triangles;

[0048] 1) For any given tree trunk location, search for 10 nearest neighbor points centered on it;

[0049] 2) Randomly select two nearest neighbor points and combine them with the current point to form a triangle;

[0050] 3) If the constructed triangle is an isosceles / equilateral triangle, remove it; otherwise, if the three sides are different, keep it.

[0051] 4) Repeat the above operation to traverse all tree trunk positions;

[0052] (2) Hash table storage

[0053] 1) Construct a hash function using the "division remainder method", defined by formula (3);

[0054] h(l1,l2,l3)=mod[mod(l1*m,N)+l2*m,N]+l3 (3)

[0055] In the formula, (l1, l2, l3) are the side lengths of any triangle, and are arranged in descending order (l1 > l2 > l3); m is a large prime number, and N represents the size of the hash table;

[0056] 2) Perform the above operation on each triangle in the triangulation so that each triangle can be projected onto a unique index in the hash table;

[0057] 3) All TLS scans in the sample plot are projected to the corresponding positions in the hash table; if a hash collision is encountered, the chaining method is used to handle it.

[0058] Furthermore, the coarse registration of the 3D point cloud:

[0059] (1) Search for triangle matching pairs

[0060] 1) For any triangle in the scanning site to be searched, use a hash function to directly map the triangle descriptor to the corresponding index in the hash table;

[0061] 2) Find the triangle descriptor contained at the current index position, calculate the direction vector difference of the corresponding vertex using formula (4), and if it is less than the threshold of 0.1, keep the tree trunk matching pair; otherwise, discard it.

[0062]

[0063] In the formula, and These represent the direction vectors of the query trunk and the candidate trunk, respectively.

[0064] (2) Calculate the coarse parameters of the three-dimensional transformation

[0065] 1) Count the scanning station IDs corresponding to each tree trunk, vote, calculate the similarity, and define it by formula (5);

[0066]

[0067] In the formula, S q With S c These represent query scan and candidate scan, respectively. c N represents the number of votes for the candidate scan. q The number of tree trunks for candidate scans;

[0068] 2) If the similarity between the query scan and the candidate scan is greater than the preset threshold, it is assumed that there is an adjacency relationship between the two stations, and the three-dimensional transformation parameters are solved using the paired trunk positions.

[0069] The factor graph optimization:

[0070] (1) Factor plot modeling

[0071] For a specific forest plot, a multi-station TLS scan was performed. Subsequently, a factor graph model was established using the unknown pose of each station in a unified coordinate system as a node and the coarse transformation between adjacent stations as a factor constraint.

[0072] (2) Global optimization

[0073] By comprehensively considering all nodes and constraints in the factor graph, the problem is transformed into a least squares problem within a Bayesian estimation framework; the inconsistencies between relative transformations are reduced, and the optimal scanning station pose and relative transformation are solved.

[0074] The point cloud fine registration and pose estimation are as follows:

[0075] (1) Point cloud precision registration;

[0076] Using the relative pose transformation parameters optimized by the factor map as initial values, the transformation parameters between each effective paired scanning station are optimized and adjusted using a point-to-point ICP fine registration method.

[0077] (2) Global pose derivation;

[0078] Based on the optimized relative transformation parameters, the pose of each scanning station in a unified coordinate system is derived using a depth-first strategy.

[0079] (3) Unify the point cloud coordinate system;

[0080] After the above steps, the poses of each scanning station in a unified coordinate system are known and applied to each scanning point cloud, transforming it from an independent local coordinate system into a global point cloud under a unified coordinate system.

[0081] Another objective of this invention is to provide a multi-station ground-based laser scanning point cloud registration system suitable for forest environments, comprising:

[0082] Distance image generation module, used to generate distance images;

[0083] The continuity constraint module is used for the continuity constraint of the trunk points; it performs distance continuity constraints from the column direction and extracts the stable part of the point cloud.

[0084] The orientation constraint module is used to constrain the orientation of tree trunk points; it removes the influence of still existing dense branches and leaves as well as scattered ground fragments.

[0085] The descriptor building module is used to construct triangular descriptors; triangles are used to describe the relative relationships between discrete tree trunk locations; multiple triangles are combined to form a triangular mesh, which is used to describe the overall distribution of tree trunks in the scan; the number of triangles depends on the size of the scanned plot and the tree trunk density;

[0086] The registration module is used for coarse registration of 3D point clouds. It uses efficient hash table search to find triangle matching pairs, calculates the similarity between matching pairs of scanning stations, and outputs the coarse registration parameters of the scanned point cloud pairs if the similarity meets the threshold.

[0087] The optimization module is used for factor graph optimization. It introduces the factor graph optimization method to model the inconsistencies between relative transformations of multiple stations and obtains globally consistent transformation parameters after overall optimization.

[0088] The pose estimation module is used for fine registration and pose estimation of point clouds. It uses the relative pose transformation parameters optimized by factor graph as initial values ​​and optimizes them by point-to-point ICP method. It derives the pose of each scanning station in a unified coordinate system through the optimized transformation parameters and transforms the point clouds of all scanning stations to a unified coordinate system to form a consistent global point cloud.

[0089] Another object of the present invention is to provide a computer device comprising a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform the steps of the multi-station ground-based laser scanning point cloud registration method applicable to forest environments.

[0090] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the multi-station ground-based laser scanning point cloud registration method applicable to forest environments.

[0091] Another objective of this invention is to provide an information data processing terminal for implementing the multi-station ground-based laser scanning point cloud registration system suitable for forest environments.

[0092] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows:

[0093] First, this invention proposes a multi-station ground-based laser scanning point cloud registration method and system suitable for forest environments. This method requires no prior information and achieves automated conversion of multi-station point clouds to a unified coordinate system. It solves the problem of time-consuming and labor-intensive manual target deployment during multi-station TLS scanning, significantly improving operational efficiency. This method bridges the gap between raw TLS scanning data and practical applications, providing a novel solution for forest surveys. This invention proposes a multi-station ground-based laser scanning point cloud registration method and system suitable for forest environments, solving the problems of significant noise point influence, low computational efficiency, and poor applicability. This method can convert multi-station local scans in forest plots to a unified coordinate system without requiring any prior information, such as initial transformations or scanning station adjacency relationships, providing data support for subsequent applications. First, this invention converts the original 3D point cloud into a 2D distance image, extracts tree trunks based on continuity constraints, and locates the root points of the trunks. Second, it constructs triangle descriptors to describe the spatial relationships between tree trunks during scanning and stores them in a hash table. The hash table search operation is used to find corresponding triangle pairs, calculate the similarity between scans, and simultaneously extract 3D transformation parameters. Because the 3D transformations calculated in the above steps are relatively coarse and inconsistent, it is difficult to accurately derive the poses of each TLS scanning station. This invention introduces factor graph modeling to address the inconsistency in coarse transformations between multi-site point clouds. The desired scanning station pose is treated as a node, and the relative transformation is used as a factor constraint to obtain the optimal scanning station pose parameters with overall consistency. Finally, using the factor graph-optimized transformation parameters as initial values, the ICP method is used to further optimize and adjust them. A depth-first strategy is employed to derive the poses of each scanning station in a unified coordinate system, transforming the local point cloud to the unified coordinate system and achieving global point cloud registration without any prior information.

[0094] Secondly, as supplementary evidence of the inventive step of the claims of this invention, it is also reflected in the following important aspects:

[0095] (1) The expected benefits and commercial value of the technical solution of this invention after transformation are as follows:

[0096] This invention proposes a multi-station ground-based laser scanning point cloud registration method and system suitable for forest environments. Using tree trunk locations as registration units, it achieves multi-station point cloud registration without the need for artificial targets in the field. It bridges the gap between field data collection and practical production applications, providing data support for the automated extraction of individual tree attributes and forest surveys. Therefore, this invention can ensure the achievement of "carbon peaking" and "carbon neutrality," and provides a new evaluation scheme for the carbon trading market, possessing significant commercial and economic value.

[0097] (2) The technical solution of this invention fills a technical gap in the industry both domestically and internationally:

[0098] The technical solution of this invention innovatively proposes three aspects: 1) Using the tree trunk position as the registration primitive, a triangular descriptor (the simplest polygon) is constructed to represent the relative position between the tree trunks. 2) A hash table is used to efficiently store the triangular descriptors from multi-station scans. Paired triangles are found using hash key values ​​and directional constraints of the tree trunk point cloud clusters. The similarity between the point clouds from two stations is calculated, and the coarse transformation parameters are solved. 3) Using the pose to be determined at each scanning station in a unified coordinate system as nodes and the relative transformation as a factor constraint, a factor graph is constructed to obtain the globally consistent and optimal pose parameters. Finally, the optimized relative pose is used as the initial transformation, and fine-tuning is performed through a point-to-point iterative nearest neighbor registration algorithm to obtain high-precision pose parameters. All point cloud data are transformed to a unified coordinate system.

[0099] The above three innovations fill the gap in reliable multi-site cloud registration technology for forest environments both domestically and internationally, enabling efficient, accurate, and automatic registration of multi-site TLS scanning in complex forest environments without the need for any prior information, manual operation, or special markings.

[0100] (3) The technical solution of the present invention solves a technical problem that people have long wanted to solve but have never been able to solve successfully:

[0101] Traditional multi-station laser scanning methods require on-site target deployment and manual selection of corresponding points during data processing, which is time-consuming and labor-intensive, significantly reducing operational efficiency. Furthermore, forest scenes often exhibit irregular geometric features and numerous noisy points, making existing point cloud registration methods insufficient for practical applications. Therefore, a robust, accurate, and automated method is urgently needed. The method and system proposed in this invention can solve the technical challenges of multi-station TLS scanning point cloud registration in complex forest environments, improve the efficiency of forest survey operations, and promote the application of TLS in forestry.

[0102] Third, the multi-station ground-based laser scanning point cloud registration system and method of this invention addresses the registration requirements of multi-station laser scanning data in forest environments, significantly solving the problem of existing technologies struggling to obtain high-precision global point clouds in forest environments. Its key parameter selection, algorithm design, and mathematical model have all achieved technological breakthroughs in accuracy and efficiency, specifically in the following aspects:

[0103] 1. Efficient point cloud filtering and feature extraction

[0104] This invention utilizes multiple filtering methods, including distance imagery and tree trunk point continuity constraints, to effectively remove dense branches and ground interference fragments from point clouds. By setting search step size, pixel difference threshold, and group pixel number threshold, point cloud data is processed efficiently while retaining stable tree trunk data, which is difficult to achieve in traditional forest environment laser scanning point cloud processing. Fast Euclidean distance clustering combined with Principal Component Analysis (PCA) classifies point cloud clusters based on geometric characteristics, ensuring purer extracted tree trunk point data and significantly improving registration accuracy.

[0105] 2. Accurate triangle descriptor construction and matching

[0106] The triangular descriptors proposed in this invention are constructed based on the relative positional relationships of tree trunks, effectively describing the overall distribution characteristics of tree trunks in a forest environment. By storing the constructed triangles in a hash table and using the "division remainder method" to construct the hash function, fast and accurate triangle matching is achieved, reducing redundant data calculations in traditional point cloud registration. Using these descriptors, the system can efficiently find similar tree trunk matching pairs between scanning stations, laying the foundation for subsequent registration parameter calculations. This method has higher matching efficiency compared to traditional direct point-to-point registration methods.

[0107] 3. Factor graph optimization and global pose derivation

[0108] To address the common problem of accumulated errors in multi-site point cloud registration, this invention introduces factor graph optimization, modeling the coarse transformation result as a least-squares problem within a Bayesian estimation framework. By modeling the poses of multiple scanning stations using factor graphs, and comprehensively considering the relative transformations between stations, inconsistencies between relative transformations are reduced, thereby optimizing the global pose. Factor graph optimization not only enhances the system's robustness but also significantly improves global consistency, ensuring the accuracy and reliability of the final output global point cloud.

[0109] 4. Point cloud fine registration and global pose estimation using a depth-first strategy

[0110] Based on the factor graph optimization results, this invention uses the ICP algorithm for fine registration. Using the factor graph-optimized relative pose transformation parameters as initial values, the transformation parameters between paired scanning stations are progressively optimized to reduce pose errors. Subsequently, a depth-first strategy is used to derive the pose of all scanning stations in a unified coordinate system. Compared to traditional methods, this approach can reduce computation time while maintaining registration accuracy, enabling rapid transformation of the point clouds of each scanning station to a unified coordinate system, forming a complete global point cloud, and providing a solid data foundation for subsequent processing of large-scale forest data.

[0111] In summary, this invention achieves efficient and accurate multi-station laser scanning point cloud registration in forest environments. Compared to existing technologies, the innovations of this invention mainly lie in the refinement of data filtering, the efficiency of feature matching, and the consistency of global pose optimization. This technology can significantly improve the efficiency and accuracy of data acquisition in forest ecological research and resource management, and has broad application prospects, especially playing an important role in environmental monitoring and forest resource surveys.

[0112] Fourth, the mathematical model of this invention solves the following prior art problems:

[0113] 1. The problem of insufficient point cloud registration accuracy.

[0114] In complex forest environments, point cloud data contains a large number of branch and leaf points and ground noise, making it difficult for existing technologies to accurately extract tree trunk points, resulting in large registration errors.

[0115] Point cloud matching relies on simple features (such as nearest neighbor search), which can easily lead to matching failures due to environmental complexity and occlusion.

[0116] 2. The problem of poor global consistency

[0117] In multi-site cloud registration methods, common local feature matching (such as the ICP method) is difficult to guarantee global consistency, which can easily lead to accumulated errors and ultimately cause overall point cloud distortion.

[0118] There is a lack of systematic modeling methods to describe the global relationships between multi-station scanning stations.

[0119] 3. There is a lack of effective optimization methods to handle pose inconsistencies between multiple stations.

[0120] Existing technologies cannot effectively eliminate local pose errors between multiple stations when performing multi-station point cloud registration, resulting in significant mismatch in point clouds under a unified coordinate system.

[0121] Significant technological advancements in the mathematical model of this invention:

[0122] 1. Distance image and trunk point extraction model improve point cloud processing accuracy

[0123] Distance image generation formula: Convert the point cloud from a three-dimensional coordinate system to a two-dimensional distance image, and solve the problem of large computational load in traditional three-dimensional point cloud processing by mapping pixel values ​​(distance from a point to the scan center).

[0124] Mathematical formula (1) ensures that the scan points are accurately mapped to the two-dimensional pixel index, which facilitates subsequent operations.

[0125] Continuity constraint model: Accurately separates trunk points from branches and ground points through dual threshold constraints of pixel difference and group pixel number.

[0126] Geometric direction constraint formula (2): Based on the calculation of geometric characteristics by principal component analysis (PCA), it significantly improves the classification ability of trunk points, branch and leaf points and ground points.

[0127] 2. Triangular descriptors and hash storage models improve feature matching efficiency.

[0128] Triangle descriptor model: It uses the side lengths and angles of triangles to describe the geometric distribution between trunk points. By constructing a sparse but stable triangular network, it solves the problem of unstable point cloud features.

[0129] Hash table storage formula (3): Project the triangular descriptor to the unique index position through the hash function, and combine the chaining method to handle collisions, so as to efficiently realize fast matching between multiple cloud sites.

[0130] 3. Achieving global consistency through factor graph optimization model

[0131] Factor graph modeling: By using the pose of the scanning station as nodes and the coarse transformation between adjacent stations as factor constraints, a global optimization model is established, which makes up for the global consistency defect of the traditional ICP method.

[0132] Least squares optimization under the Bayesian estimation framework: By minimizing the constraint inconsistency between nodes, the registered point cloud is ensured to have global consistency.

[0133] 4. Multi-site cloud-based precise registration and pose inference model improve global accuracy

[0134] ICP optimization model: Using the parameters optimized by the factor graph as initial values, the pose transformation between multiple stations is further precisely adjusted through the point-to-point ICP method.

[0135] Depth-first pose derivation: Based on the optimized transformation parameters, the global pose in a unified coordinate system is derived through depth-first search, achieving uniformity and accuracy of the global point cloud.

[0136] 5. Multi-level modeling and distributed processing address adaptability issues in complex scenarios.

[0137] This invention establishes a complete point cloud processing system from multiple levels (distance image, trunk points, triangular mesh, global optimization), and combined with a distributed computing framework, it can efficiently process large-scale, multi-station scan data and adapt to forest scenes of different densities and complexities.

[0138] This invention addresses the problems of insufficient registration accuracy, poor global consistency, and inconsistent poses among multiple stations in existing technologies through a mathematical model that incorporates distance image generation, continuity constraints, geometric orientation constraints, triangle descriptors, factor graph optimization, and ICP fine registration. The significant technological advancements are reflected in the following aspects:

[0139] - Improved accuracy: Through multi-level modeling and optimization, the registration accuracy of multi-site clouds in forest environments has been significantly improved.

[0140] - Efficiency Improvement: By utilizing hash tables and sparse triangular descriptors, the computational cost of feature matching and pose optimization is significantly reduced.

[0141] - Enhanced Adaptability: The model demonstrates excellent processing performance on point cloud data with varying forest densities, shading conditions, and complex terrain, providing strong technical support for industrial applications such as forest resource surveys and ecological monitoring. Attached Figure Description

[0142] Figure 1 This is a flowchart of a multi-station ground-based laser scanning point cloud registration method suitable for forest environments, provided by an embodiment of the present invention.

[0143] Figure 2 This is a structural block diagram of a multi-station ground-based laser scanning point cloud registration system suitable for forest environments, provided by an embodiment of the present invention.

[0144] Figure 3 This invention provides a method for converting 3D point clouds into distance images. (a) TLS scanning principle. (b) Original 3D point cloud. (c) Converted distance image.

[0145] Figure 4 These are the continuity constraints in the distance image provided in the embodiments of the present invention. (a) Distance image of point cloud projection. (b) Continuous segments obtained from continuity constraints.

[0146] Figure 5 These are the geometric constraints of the tree trunk point cloud provided in this embodiment of the invention. (a) All point cloud clusters generated after Euclidean clustering. (b) Tree trunk point cloud clusters after geometric constraints.

[0147] Figure 6 This refers to the extracted tree trunk location provided in this embodiment of the invention.

[0148] Figure 7This is a triangle descriptor provided in an embodiment of the present invention. (a) Filtering triangle shapes. (b) Triangle rotation invariance.

[0149] Figure 8 The hash table provided in this embodiment of the invention stores triangle descriptors. (a) Single triangle storage. (b) Multi-scan station data storage.

[0150] Figure 9 These are the search triangle matching pairs provided in the embodiments of the present invention. (a) Hash search. (b) Trunk direction constraint.

[0151] Figure 10 This is the algorithm flowchart factor graph optimization provided in the embodiments of the present invention. (a) Multi-station TLS scanning of forest plots. (b) Factor graph modeling.

[0152] Figure 11 This is an algorithm flowchart provided in an embodiment of the present invention.

[0153] Figure 12 This is a schematic diagram of tree trunk point extraction provided in an embodiment of the present invention. Different tree trunks have different gray levels, and black dots represent the location of the tree trunk. (a) Shennongjia data example. (b) ETH data example. (c) FGI data example. (d) Tongji University data example.

[0154] Figure 13 This is a bar chart for evaluating the coarse registration accuracy of point clouds provided in this embodiment of the invention. The shaded areas in the bar chart represent whether factor map optimization has been performed. (a) Rotation error. (b) Translation error.

[0155] Figure 14 This is the point cloud registration error assessment provided in the embodiments of the present invention. (a) Rotation error. (b) Translation error.

[0156] Figure 15 These are schematic diagrams of the overall point cloud after registration, provided in embodiments of the present invention. (a) Shennongjia dataset. (b) ETH dataset. (c) FGI dataset. (d) Tongji University dataset. Detailed Implementation

[0157] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0158] The technical solution of this invention solves the following key problems in the prior art in industrial applications and achieves significant technological progress:

[0159] Technical issues:

[0160] 1. The difficulty of point cloud registration in complex forest environments

[0161] When registering multi-station ground-based laser scanning point clouds in complex forest environments, existing technologies struggle to effectively separate and match points from tree trunks, branches, leaves, and ground surfaces due to the mixture of these points. Traditional methods rely on manual selection or a single algorithm, resulting in low efficiency and insufficient accuracy.

[0162] 2. Insufficient global consistency in the registration process.

[0163] Existing registration methods mostly rely on local feature matching, which makes it difficult to guarantee the consistency and robustness of the global point cloud. Especially in multi-station scan data, cumulative errors are prone to occur, leading to overall point cloud deformation or distortion.

[0164] 3. Challenges in Pose Estimation in Dynamic Forest Environments

[0165] In dynamic forest environments, due to uneven tree distribution and severe local occlusion, existing methods struggle to accurately estimate the global pose of scanning stations and generate global point clouds in a unified coordinate system.

[0166] Significant technological progress

[0167] 1. Precise extraction of tree trunk points improves registration efficiency.

[0168] This invention utilizes distance image generation, trunk point continuity constraints, and geometric orientation constraints to determine trunk points using pixel difference thresholds, Euclidean distance clustering, and geometric characteristics, accurately separating trunk points from branches, leaves, and ground points. This improvement significantly enhances the efficiency and accuracy of point cloud processing, laying a solid foundation for subsequent registration.

[0169] 2. Triangular descriptors improve feature matching accuracy

[0170] By constructing triangular descriptors, this invention describes the geometric relationships between discrete trunk points and employs hash tables for efficient storage and searching. This method can quickly and accurately match trunk points in multi-station scan point clouds, avoiding redundancy or mismatch problems that are prone to occur in traditional feature matching methods.

[0171] 3. Factor graph optimization achieves global consistency

[0172] A factor graph optimization method is introduced to model the poses and relative transformations of multi-station scanning stations as factor graphs. By using least squares optimization under the Bayesian estimation framework, the problem of local error accumulation is solved, which significantly improves the global consistency and robustness of point cloud registration.

[0173] 4. The organic combination of precise registration and pose derivation

[0174] Based on factor graph optimization, a point-to-point ICP method is used for fine registration, and a depth-first strategy is combined to derive the global pose of each scanning station. This process not only improves registration accuracy but also successfully transforms the station clouds from a local coordinate system to a global coordinate system.

[0175] 5. Application adaptability in dynamic forest environments

[0176] By deeply modeling and optimizing the distribution of tree trunk points in the forest environment, this invention can adapt to forest plot scanning tasks with different densities and structures, and maintain high registration accuracy and efficiency even in dynamic environments with severe shading and complex structures.

[0177] 6. Industrial Application Value

[0178] This invention significantly improves the registration accuracy and efficiency of multi-station ground-based laser scanning technology in forest environments, and can be widely applied in fields such as forest resource surveys, carbon sink measurement, and vegetation growth analysis. Furthermore, its versatility design allows it to be extended to other complex environments (such as urban or industrial scenarios), providing an innovative solution for the industrialization of lidar point cloud technology.

[0179] Example 1: Forest Resource Survey and Carbon Sequestration Measurement

[0180] Forest resource surveys are an important means of assessing forest health, tree density, tree species distribution, and carbon sequestration capacity. Traditional methods rely on manual measurement or drone remote sensing, which are insufficient to achieve the high precision required for detailed measurements.

[0181] 1. Data Collection

[0182] Multi-station ground-based laser scanning (TLS) equipment is deployed in designated forest plots (e.g., 20m×20m areas), using a central and corner station layout based on the plot shape and vegetation density.

[0183] Each station scan yields high-precision point cloud data, covering tree trunks, branches, leaves, and ground information.

[0184] 2. Data Processing

[0185] The point cloud registration method of this invention is used for trunk point extraction, geometric orientation constraint, and point cloud matching:

[0186] Accurate extraction of tree trunk points removes branch and leaf noise and ground noise while preserving tree trunk information;

[0187] Triangular descriptors and factor graph optimization achieve global consistency registration for multi-site cloud environments.

[0188] 3. Results Output

[0189] Obtain high-precision global point cloud data, including the spatial location, diameter, density, and distribution of tree trunks.

[0190] Extracting total biomass and carbon storage from forest plots provides support for assessing forest carbon sequestration capacity.

[0191] Value embodiment

[0192] - It enables high-precision quantitative assessment of forest resources, providing a reliable data foundation for forestry management departments.

[0193] - Improve the accuracy of carbon sink capacity measurement to support the development of the carbon emissions trading market.

[0194] Example 2: Monitoring of Forest Ecological Restoration after Natural Disasters

[0195] Forests often require assessment of their recovery after natural disasters such as fires, storms, or insect infestations, and traditional methods are inefficient due to the instability and complexity of the damaged areas.

[0196] 1. Data Collection

[0197] Multi-station laser scanning was set up in the disaster-affected area, and multi-station TLS scanning was carried out on key sample plots (such as the core disaster area or the key recovery area) to obtain point cloud data.

[0198] The station deployment method is flexible and can adapt to irregular terrain and the distribution of remaining trees.

[0199] 2. Data Processing

[0200] Point cloud processing of damaged and restored areas is performed using the method of this invention:

[0201] Extract the trunk points of surviving and fallen trees.

[0202] The distribution of tree trunk locations is reconstructed using triangular descriptors to monitor changes in spatial relationships between trees.

[0203] The post-disaster data was compared with historical point cloud data to analyze the progress of vegetation recovery.

[0204] 3. Results Output

[0205] The output area includes tree recovery rate, spatial density variation, and the number and distribution of fallen or broken trees.

[0206] Generate a 3D visualization model to intuitively present the restoration process.

[0207] - Provide quantitative indicators for ecological protection and restoration, and guide subsequent vegetation reconstruction and artificial intervention.

[0208] - Reduces the risks of manual on-site investigation and significantly improves monitoring efficiency and accuracy.

[0209] like Figure 1 As shown in the figure, the multi-station ground-based laser scanning point cloud registration method suitable for forest environments provided by this invention includes the following steps:

[0210] S101: Generate distance image;

[0211] S102: Trunk point continuity constraint;

[0212] Apply distance continuity constraints along the column direction to extract the stable portion of the point cloud;

[0213] S103: Trunk point orientation constraint;

[0214] Even with continuity constraints, some dense branches and scattered ground fragments remain in the extracted tree trunk points. Point clouds are clustered, and then the geometric properties of each cluster are used to remove branches, leaves, and ground points, while retaining the tree trunk points.

[0215] S104: Construct a triangle descriptor;

[0216] Triangles are used to describe the relative relationships between discrete tree trunk locations; multiple triangles are combined to form a triangular network to describe the overall distribution of tree trunks in the scan; the number of triangles depends on the size of the scanned plot and the tree trunk density.

[0217] S105: Coarse registration of 3D point clouds;

[0218] The algorithm uses efficient hash table search to find triangle matching pairs, calculates the similarity between matching pairs of scanning stations, and outputs the coarse registration parameters of the scan point cloud pairs if the similarity meets the threshold.

[0219] S106: Factor plot optimization;

[0220] A factor graph optimization method is introduced to model the inconsistencies between relative transformations at multiple stations, and globally consistent transformation parameters are obtained after overall optimization.

[0221] S107: Point cloud fine registration and pose estimation;

[0222] The relative pose transformation parameters optimized by factor graphs are used as initial values ​​and optimized by point-to-point ICP method. The pose of each scanning station in a unified coordinate system is derived by the optimized transformation parameters, and all scanning station point clouds are transformed to a unified coordinate system to form a consistent global point cloud.

[0223] The generated distance image provided in this embodiment of the invention:

[0224] (1) Ground-based lidar data acquisition;

[0225] TLS scanning stations are deployed according to the size of the forest plot. For a 20m×20m plot, a scanning method of one station in the middle and one station at each of the four corners can be adopted to obtain complete plot scanning data to the greatest extent.

[0226] (2) Converting 3D point clouds into distance images;

[0227] Processing single-site point cloud data, let the three-dimensional coordinates of any point i in the point cloud be (x, y, z). i ,y i ,z i It can also be represented as polar coordinates (α). i ,β i ,γ i The form of ) where α i ,β i These are the horizontal angle and the vertical angle, γ. i Let i be the distance from the scan point i to the scan center; then, the index of the scan point i in the distance image can be defined by formula (1);

[0228]

[0229] In the formula, row i ,col i Here, i represents the row and column indices of the scan point i in the distance image, respectively, and θ represents the angular resolution of the distance image.

[0230] Distance to each pixel in the image using γ i For padding, if multiple scan points correspond to the same index position after conversion, the pixel retains the information of the scan point closest to the center.

[0231] The tree trunk point continuity constraint provided in this embodiment of the invention:

[0232] (1) Define the search step size (Step) and the pixel difference threshold (R). diff ), Group pixel count threshold (Seg th The three parameters are set to 15 pixels, 0.03m, and 100 pixels respectively;

[0233] (2) For each column in the distance image, starting from the first point, compare the pixel value difference between the search point and the current point within the search step size Step. If it is less than the preset threshold R, diff If the search point is in the same group as the current point, then the search point and the current point are grouped together; otherwise, a new group is created.

[0234] (3) Repeat the above operation until all pixels in the distance image have been accessed;

[0235] (4) Count the number of pixels in each group. If it exceeds the preset threshold Seg thIf a pixel in a group is not found, it is retained and marked as 1; otherwise, it is discarded and marked as 0.

[0236] (5) Traverse all pixels. If the pixel is marked as 1, keep the corresponding scan point. Otherwise, if the pixel is marked as 0, discard the scan point.

[0237] The tree trunk point orientation constraint provided in this embodiment of the invention:

[0238] (1) Using Euclidean distance clustering, the point cloud data extracted in step 2 is processed and divided into different point cloud clusters;

[0239] (2) Perform principal component analysis (PCA) on each point cloud cluster and calculate the linear L. λ Planarity P λ Dispersion S λ Three parameters describe the geometric distribution of point cloud clusters in space, defined by formula (2);

[0240]

[0241] In the formula, λ1>λ2>λ3 are the eigenvalues ​​arranged in descending order, which correspond to the degree of change in the direction of the eigenvector;

[0242] (3) The tree trunk points exhibit a linear distribution in space, that is, they have a large linearity L. λ The ground point has a large planarity P λ The branch and leaf points have a large degree of dispersion. λ Therefore, this characteristic can be used to determine the type of extracted point cloud clusters. If a point cloud cluster satisfies condition L... λ If the value is greater than 0.95, then this point cloud cluster is considered to be a tree trunk; the distribution of point cloud clusters before and after applying geometric constraints;

[0243] (4) Using the extracted tree trunk point cloud cluster, extract the corresponding tree trunk position; its planar position (x,y) is the center of the point cloud cluster, and the z component is selected as the root point of the tree trunk, that is, the intersection of the tree trunk and the ground. The black dots at the bottom are the extracted tree trunk positions.

[0244] The construction of the triangle descriptor provided in this embodiment of the invention:

[0245] (1) Filter the triangles;

[0246] 1) For any given tree trunk location, search for 10 nearest neighbor points centered on it;

[0247] 2) Randomly select two nearest neighbor points and combine them with the current point to form a triangle;

[0248] 3) If the constructed triangle is an isosceles / equilateral triangle, remove it; otherwise, if the three sides are different, keep it.

[0249] 4) Repeat the above operation to traverse all tree trunk positions;

[0250] (2) Hash table storage

[0251] 1) Construct a hash function using the "division remainder method", defined by formula (3);

[0252] h(l1,l2,l3)=mod[mod(l1*m,N)+l2*m,N]+l3 (3)

[0253] In the formula, (l1, l2, l3) are the side lengths of any triangle, and are arranged in descending order (l1 > l2 > l3); m is a large prime number, and N represents the size of the hash table;

[0254] 2) Perform the above operation on each triangle in the triangulation so that each triangle can be projected onto a unique index in the hash table;

[0255] 3) All TLS scans in the sample plot are projected to the corresponding positions in the hash table; if a hash collision is encountered, the chaining method is used to handle it.

[0256] The coarse registration of 3D point clouds provided in this embodiment of the invention:

[0257] (1) Search for triangle matching pairs

[0258] 1) For any triangle in the scanning site to be searched, use a hash function to directly map the triangle descriptor to the corresponding index in the hash table;

[0259] 2) Find the triangle descriptor contained at the current index position, calculate the direction vector difference of the corresponding vertex using formula (4), and if it is less than the threshold of 0.1, keep the tree trunk matching pair; otherwise, discard it.

[0260]

[0261] In the formula, and These represent the direction vectors of the query trunk and the candidate trunk, respectively.

[0262] (2) Calculate the coarse parameters of the three-dimensional transformation

[0263] 1) Count the scanning station IDs corresponding to each tree trunk, vote, calculate the similarity, and define it by formula (5);

[0264]

[0265] In the formula, S q With S c These represent query scan and candidate scan, respectively.c N represents the number of votes for the candidate scan. q The number of tree trunks for candidate scans;

[0266] 2) If the similarity between the query scan and the candidate scan is greater than the preset threshold, it is assumed that there is an adjacency relationship between the two stations, and the three-dimensional transformation parameters are solved using the paired trunk positions.

[0267] The factor graph optimization:

[0268] (1) Factor plot modeling

[0269] For a specific forest plot, a multi-station TLS scan was performed. Subsequently, a factor graph model was established using the unknown pose of each station in a unified coordinate system as a node and the coarse transformation between adjacent stations as a factor constraint.

[0270] (2) Global optimization

[0271] By comprehensively considering all nodes and constraints in the factor graph, the problem is transformed into a least squares problem within a Bayesian estimation framework; the inconsistencies between relative transformations are reduced, and the optimal scanning station pose and relative transformation are solved.

[0272] The point cloud fine registration and pose estimation are as follows:

[0273] (1) Point cloud precision registration;

[0274] Using the relative pose transformation parameters optimized by the factor map as initial values, the transformation parameters between each effective paired scanning station are optimized and adjusted using a point-to-point ICP fine registration method.

[0275] (2) Global pose derivation;

[0276] Based on the optimized relative transformation parameters, the pose of each scanning station in a unified coordinate system is derived using a depth-first strategy.

[0277] (3) Unify the point cloud coordinate system;

[0278] After the above steps, the poses of each scanning station in a unified coordinate system are known and applied to each scanning point cloud, transforming it from an independent local coordinate system into a global point cloud under a unified coordinate system.

[0279] like Figure 2 As shown in the figure, an embodiment of the present invention provides a multi-station ground-based laser scanning point cloud registration system suitable for forest environments, comprising:

[0280] Distance image generation module, used to generate distance images;

[0281] The continuity constraint module is used for the continuity constraint of the trunk points; it performs distance continuity constraints from the column direction and extracts the stable part of the point cloud.

[0282] The orientation constraint module is used to constrain the orientation of tree trunk points; it removes the influence of still existing dense branches and leaves as well as scattered ground fragments.

[0283] The descriptor building module is used to construct triangular descriptors; triangles are used to describe the relative relationships between discrete tree trunk locations; multiple triangles are combined to form a triangular mesh, which is used to describe the overall distribution of tree trunks in the scan; the number of triangles depends on the size of the scanned plot and the tree trunk density;

[0284] The registration module is used for coarse registration of 3D point clouds. It uses efficient hash table search to find triangle matching pairs, calculates the similarity between matching pairs of scanning stations, and outputs the coarse registration parameters of the scanned point cloud pairs if the similarity meets the threshold.

[0285] The optimization module is used for factor graph optimization. It introduces the factor graph optimization method to model the inconsistencies between relative transformations of multiple stations and obtains globally consistent transformation parameters after overall optimization.

[0286] The pose estimation module is used for fine registration and pose estimation of point clouds. It uses the relative pose transformation parameters optimized by factor graph as initial values ​​and optimizes them by point-to-point ICP method. It derives the pose of each scanning station in a unified coordinate system through the optimized transformation parameters and transforms the point clouds of all scanning stations to a unified coordinate system to form a consistent global point cloud.

[0287] The multi-station ground-based laser scanning point cloud registration system of this invention mainly achieves accurate registration of multi-station point clouds in a forest environment through a modular approach. The following is a detailed description of its working principle:

[0288] 1. Distance Image Generation and Trunk Point Extraction

[0289] The system first generates a distance image using a distance image generation module, which performs preliminary processing on the point cloud data collected by each scanning station. Next, the system applies continuity constraints to the tree trunk points using a continuity constraint module, extracting points with continuous distances along the column direction and preserving a stable tree trunk point cloud. This step is further processed using a fast Euclidean distance clustering method to remove dense branches, leaves, and ground fragments, retaining only the tree trunk point cloud. This method leverages the relatively stable continuity of tree trunks in a forest environment, effectively reducing the interference of branches, leaves, and cluttered ground on point cloud registration.

[0290] 2. Orientation Constraints and Descriptor Construction

[0291] The orientation constraint module constrains the orientation of tree trunk points. Since the growth direction of tree trunks is typically relatively perpendicular in a forest, the system uses orientation constraints as a standard for removing residual branches and cluttered ground points. Based on this, the system constructs triangular descriptors in the descriptor construction module. Triangles are formed by describing the relative positions between discrete tree trunk points, and multiple triangles are combined into a triangular mesh. This triangular mesh characterizes the spatial distribution characteristics of tree trunks and has strong geometric stability. The number of triangles is determined by the size of the scanning area and the tree trunk density. This descriptor construction method helps improve the matching accuracy of tree trunk point clouds during registration.

[0292] 3. Coarse registration and factor plot optimization

[0293] In the registration module, the system utilizes a hash table for efficient searching. By finding matching triangle pairs, the similarity between scanning stations is calculated. Matching pairs that meet the similarity threshold are considered valid pairs for coarse registration, and coarse registration parameters are output. This registration method based on triangle descriptors and hash search can effectively achieve coarse registration between different scanning stations. Then, the system enters the optimization module, introducing a factor graph optimization method to model the relative transformation of multi-site cloud data, resolving inconsistencies caused by errors between stations. Through factor graph optimization, the system obtains globally consistent transformation parameters, providing accurate initial values ​​for the next step of fine registration.

[0294] 4. Fine registration and global pose estimation

[0295] In the pose estimation module, the system first uses the relative pose transformation parameters obtained from factor graph optimization as initial values, and then performs fine registration using a point-to-point ICP (Iterative Closest Point) method. Finally, the system derives the pose of each scanning station in a unified coordinate system and transforms all scanning station point cloud data to the same coordinate system, forming a consistent global point cloud. This method effectively reduces errors and ensures the registration accuracy of point clouds in complex forest environments, providing high-precision 3D point cloud data support for subsequent forest environment analysis and data processing.

[0296] Specific implementation of the present invention:

[0297] A multi-station ground-based laser scanning point cloud registration method and system suitable for forest environments, with the following specific implementation steps:

[0298] Step 1: Generate distance image

[0299] (1) Ground-based lidar data acquisition

[0300] TLS scanning stations are deployed according to the size of the forest plot. For example, for a 20m×20m plot, a scanning method with one station in the middle and one station at each of the four corners can be adopted to obtain complete plot scanning data to the greatest extent.

[0301] (2) Converting 3D point clouds into distance images

[0302] Processing single-site point cloud data, let the three-dimensional coordinates of any point i in the point cloud be (x, y, z). i ,y i ,z i It can also be represented as polar coordinates (α). i ,β i ,γ i The form of ) where α i ,β i These are the horizontal angle and the vertical angle, γ. i Let i be the distance from the scan point i to the scan center; then, the index of the scan point i in the distance image can be defined by formula (1).

[0303]

[0304] In the formula, row i ,col i Here, i represents the row and column indices of the scan point i in the distance image, respectively, and θ represents the angular resolution of the distance image.

[0305] Distance to each pixel in the image using γ i For padding, if multiple scan points correspond to the same index position after conversion, the pixel retains the information of the scan point closest to the center. Figure 3 The paper demonstrates the working principle of a ground-based laser scanner and the process of converting raw point cloud data into distance images.

[0306] Step 2: Continuity constraint of tree trunk points

[0307] The transformed distance image retains the geometric structure information of the original point cloud while reducing the data dimensionality. This invention applies distance continuity constraints along the column direction to extract the stable portion of the point cloud. Then, it further processes the image using a fast Euclidean distance clustering method and removes dense, scattered ground fragments through geometric constraints.

[0308] (1) Define the search step size (Step) and the pixel difference threshold (R). diff ), Group pixel count threshold (Seg th The three parameters were set to 15 pixels, 0.03m, and 100 pixels respectively.

[0309] (2) For each column in the distance image, starting from the first point, compare the pixel value difference between the search point and the current point within the search step size Step. If it is less than the preset threshold R, diff If the search point is in the same group as the current point, then the search point and the current point are grouped together. Otherwise, a new group is created.

[0310] (3) Repeat the above operation until all pixels in the distance image have been accessed.

[0311] (4) Count the number of pixels in each group. If it exceeds the preset threshold Seg th If a pixel in a group is not found in the group, it is retained and marked as 1; otherwise, it is discarded and marked as 0. Figure 4 As shown in the image.

[0312] (5) Traverse all pixels. If the pixel is marked as 1, keep the corresponding scan point. Otherwise, if the pixel is marked as 0, discard the scan point.

[0313] Step 3: Tree trunk point orientation constraint

[0314] Even among the tree trunk points extracted using continuity constraints, there are still some dense branches and leaves, as well as scattered ground fragments. These parts appear as a continuous distribution in the distance image, making them difficult to distinguish. However, their distribution in three-dimensional space is significantly different from that of the tree trunk points. For example, tree trunk points are clearly linearly distributed, while branch and leaf points are uniformly distributed in space, and the distribution of ground points tends to be planar.

[0315] (1) Using Euclidean distance clustering, the point cloud data extracted in step 2 is processed and divided into different point cloud clusters.

[0316] (2) Perform principal component analysis (PCA) on each point cloud cluster and calculate the linear L. λ Planarity P λ Dispersion S λ Three parameters describe the geometric distribution of point cloud clusters in space, defined by formula (2).

[0317]

[0318] In the formula, λ1>λ2>λ3 are the eigenvalues ​​arranged in descending order, which correspond to the degree of change in the direction of the eigenvector.

[0319] (3) The tree trunk points exhibit a linear distribution in space, that is, they have a large linearity L. λ The ground point has a large planarity P λ The branch and leaf points have a large degree of dispersion. λ Therefore, this invention utilizes this characteristic to determine the type of extracted point cloud clusters. If a point cloud cluster satisfies condition L... λ If the value is greater than 0.95, the point cloud cluster is considered to be a tree trunk. The distribution of the point cloud cluster before and after applying geometric constraints is as follows: Figure 5 As shown in the image.

[0320] (4) Using the extracted tree trunk point cloud clusters, extract the corresponding tree trunk positions. The planar position (x, y) is the center of the point cloud cluster, and the z-component is chosen as the root point of the tree trunk, i.e., the intersection of the tree trunk and the ground. For example... Figure 6 As shown in the image, the black dots at the bottom represent the extracted tree trunk locations.

[0321] Step 4: Construct the triangle descriptor

[0322] Triangles are used to describe the relative relationships between discrete tree trunk locations. Multiple triangles are then combined to form a triangular mesh, which describes the overall distribution of tree trunks during the scan. The number of triangles depends on the size of the scanned plot and the tree trunk density; therefore, how to store the triangle information is a crucial step. This invention utilizes a hash table to efficiently store the triangular mesh, with each triangle directly mapped to its corresponding location via a hash function.

[0323] (1) Filtering triangles

[0324] A triangle is the simplest geometric shape, described by only three sides, and possesses good three-dimensional rotation / translation invariance. Furthermore, besides isosceles / equilateral triangles, if two triangles have corresponding equal side lengths, then their three vertices also correspond one-to-one, such as... Figure 7 As shown in the image.

[0325] 1) For any tree trunk position, search for 10 nearest neighbor points centered on it.

[0326] 2) Select any two nearest neighbor points and combine them with the current point to form a triangle.

[0327] 3) If the constructed triangle is isosceles / equilateral, remove it. Otherwise, if the three sides are different, keep it, such as... Figure 7 As shown in the image.

[0328] 4) Repeat the above operation to traverse all tree trunk positions.

[0329] (2) Hash table storage

[0330] 1) Construct a hash function using the "division remainder method" and define it using formula (3).

[0331] h(l1,l2,l3)=mod[mod(l1*m,N)+l2*m,N]+l3 (3)

[0332] In the formula, (l1,l2,l3) are the side lengths of any triangle, and are arranged in descending order (l1>l2>l3); m is a large prime number, and N represents the size of the hash table.

[0333] 2) Perform the above operation on each triangle in the triangular network, so that each triangle can be projected onto a unique index in the hash table, such as... Figure 8 As shown in the image.

[0334] 3) All TLS scans in the sample plot are projected to the corresponding positions in the hash table. If a hash collision is encountered, the chaining method is used to handle it, such as... Figure 8 As shown in the image.

[0335] Step 5: Coarse registration of 3D point clouds

[0336] This invention relies on the efficient search operation of hash tables to find triangle matching pairs, calculate the similarity between scanning station matching pairs, and if the similarity meets the threshold, output the coarse registration parameters of the scanning point cloud pairs.

[0337] (1) Search for triangle matching pairs

[0338] 1) For any triangle in the scanned site to be searched, use a hash function to directly map the triangle descriptor to the corresponding index in the hash table, such as... Figure 9 As shown in (a).

[0339] 2) Find the triangle descriptor contained at the current index position, calculate the direction vector difference of the corresponding vertex using formula (4), and if it is less than the threshold of 0.1, retain the tree trunk matching pair; otherwise, discard it. Figure 9 As shown in (b).

[0340]

[0341] In the formula, and These represent the direction vectors of the query trunk and the candidate trunk, respectively.

[0342] (2) Calculate the coarse parameters of the three-dimensional transformation

[0343] 1) Count the scanning station IDs corresponding to each tree trunk, vote, calculate the similarity, and define it by formula (5).

[0344]

[0345] In the formula, S q With S c These represent query scan and candidate scan, respectively. c N represents the number of votes for the candidate scan. q The number of tree trunks for the candidate scan.

[0346] 2) If the similarity between the query scan and the candidate scan is greater than the preset threshold (30% in this invention), then it is considered that there is an adjacency relationship between the two stations, and the three-dimensional transformation parameters are solved using the paired trunk positions.

[0347] Step 6: Factor graph optimization

[0348] The 3D transformation parameters obtained from coarse point cloud registration exhibit significant inconsistencies, and errors accumulate gradually over multiple stations, making it difficult to provide reliable initial values ​​for subsequent fine registration steps. Therefore, this invention introduces a factor graph optimization method to model the inconsistencies between relative transformations across multiple stations, obtaining globally consistent transformation parameters after overall optimization.

[0349] (1) Factor plot modeling

[0350] For a specific forest plot, conduct multi-site TLS scans, such as Figure 10 As shown in (a). Subsequently, using the desired pose of each station scanned in a unified coordinate system as nodes, and the coarse transformation between adjacent stations as factor constraints, a factor graph model is established, as follows: Figure 10 As shown in (b).

[0351] (2) Global optimization

[0352] Considering all nodes and constraints in the factor graph, the problem is transformed into a least-squares problem within a Bayesian estimation framework. This reduces inconsistencies between relative transformations, allowing for the solution of the optimal scanning station pose and relative transformation.

[0353] Step 7: Point cloud fine registration and pose estimation

[0354] This invention uses the relative pose transformation parameters optimized by factor graphs as initial values ​​and optimizes them using a point-to-point ICP method. Subsequently, the poses of each scanning station in a unified coordinate system are derived using the optimized transformation parameters, and all scanning station point clouds are transformed to a unified coordinate system to form a globally consistent point cloud.

[0355] (1) Point cloud fine registration

[0356] Using the relative pose transformation parameters optimized by the factor map as initial values, the transformation parameters between each effective paired scanning station are optimized and adjusted using a point-to-point ICP fine registration method.

[0357] (2) Global pose derivation

[0358] Based on the optimized relative transformation parameters, the pose of each scanning station in a unified coordinate system is derived using a depth-first strategy.

[0359] (3) Unify the point cloud coordinate system

[0360] After the above steps, the poses of each scanning station in a unified coordinate system are known and applied to each scanning point cloud, transforming it from an independent local coordinate system into a global point cloud under a unified coordinate system.

[0361] Technical Solution and Flowchart for Multi-Station Ground-Based Laser Scanning Point Cloud Registration Method and System Applicable to Forest Environments Figure 11 .

[0362] Technical solutions of the embodiments of the present invention

[0363] Taking 12 forest plots from the self-collected dataset of Shennongjia National Nature Reserve in Hubei Province and the publicly available datasets released by ETH Zurich, the Finnish Geosciences Institute (FGI), and Tongji University as examples, the technical solution of the multi-station ground-based laser scanning point cloud registration method and system applicable to forest environments is as follows.

[0364] (1) Trunk point extraction

[0365] In forest sample plots, operators deploy multiple scanning stations from different angles to cover the entire plot area as completely as possible. For any given station's TLS scan data, tree trunk points are extracted. Then, the location corresponding to each tree trunk point is calculated. This invention uses the root point, i.e., the intersection of the tree trunk point and the ground, as a representative point to describe the tree trunk distribution. Below is an example of processing representative sample plots from four datasets: Shennongjia, ETH, FGI, and Tongji University, and the extracted tree trunk points and their corresponding locations are shown below. Figure 12 As shown in the image.

[0366] Figure 12 Each of the four datasets exhibits distinct characteristics. The Shennongjia dataset represents a typical subtropical natural forest with a chaotic and dense distribution of tree trunks, making extraction challenging. The ETH and FGI datasets represent northern natural forests with high density and a few shrubs. The Tongji University dataset, collected from the campus, represents a managed forest with low density and a regular distribution. This invention demonstrates good trunk extraction results in all four datasets, showing complete structure and accurate root point locations.

[0367] (2) Coarse registration of point clouds

[0368] Triangle descriptors are constructed based on the extracted trunk points. A hash table database is used to efficiently store triangle descriptors from all scans and search for pairs of triangles with the same name. Orientation constraints are added to the paired trunks, the similarity between two scan stations is calculated, the adjacency of two scan stations is determined, and the coarse registration 3D transformation parameters are output.

[0369] Subsequently, the inconsistency problem in the coarse registration parameters was modeled, with the pose of the scanning station to be determined as a node and the adjacent transformations between scanning stations as factor constraints, forming a factor graph. After optimization, overall consistent transformation parameters were obtained, providing reliable initial values ​​for fine point cloud registration.

[0370] Figure 13The paper illustrates the distribution of coarse registration errors in point clouds across four datasets, including rotation and translation errors. It also shows the trends of these errors before and after factor map optimization. The results demonstrate that factor map optimization effectively reduces the impact of coarse registration errors. In most sample plots, it achieves rotation errors of less than 20 mrad and translation errors of less than 30 cm.

[0371] (3) Point cloud precision registration

[0372] Based on the reliable initial transformation values ​​provided by the factor graph optimization, this invention uses a point-to-point ICP registration method for fine optimization and a depth-first strategy to derive the pose of each scanning station in a unified coordinate system. The error evaluation results are as follows: Figure 14 As shown in the image.

[0373] Figure 14 The paper presents the experimental results of precise registration of the present invention across four datasets and twelve sample plots. Extensive experimental testing demonstrates that the proposed method achieves a rotation error of 0-3 mrad and a translation error of 0-3 cm in most sample plots. The high-precision registration results demonstrate the superiority of the proposed method.

[0374] 2. Technical effects of the embodiments of the present invention

[0375] The embodiments of this invention conducted extensive experiments on 12 forest plots using self-collected datasets from the Shennongjia National Nature Reserve in Hubei Province, publicly available datasets from ETH Zurich, the Finnish Geophysical Institute (FGI), and Tongji University. The multi-station ground-based laser scanning point cloud registration method and system proposed in this invention, suitable for forest environments, achieved good experimental results, realizing automatic multi-station point cloud registration without any prior information. The overall point cloud after registration is shown below. Figure 15 As shown in the image, four datasets are presented as representative examples.

[0376] This invention provides a computer device, which includes a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the multi-station ground-based laser scanning point cloud registration method suitable for forest environments.

[0377] like Figure 2 As shown, this embodiment of the invention provides a multi-station ground-based laser scanning point cloud registration system suitable for forest environments, implementing the robust real-time positioning and mapping method applicable to complex forest understory environments. The system includes:

[0378] Distance image generation module, used to generate distance images;

[0379] The continuity constraint module is used for the continuity constraint of the trunk points; it performs distance continuity constraints from the column direction and extracts the stable part of the point cloud.

[0380] The orientation constraint module is used to constrain the orientation of tree trunk points; it removes the influence of still existing dense branches and leaves as well as scattered ground fragments.

[0381] The descriptor building module is used to construct triangular descriptors; triangles are used to describe the relative relationships between discrete tree trunk locations; multiple triangles are combined to form a triangular mesh, which is used to describe the overall distribution of tree trunks in the scan; the number of triangles depends on the size of the scanned plot and the tree trunk density;

[0382] The registration module is used for coarse registration of 3D point clouds. It uses efficient hash table search to find triangle matching pairs, calculates the similarity between matching pairs of scanning stations, and outputs the coarse registration parameters of the scanned point cloud pairs if the similarity meets the threshold.

[0383] The optimization module is used for factor graph optimization. It introduces the factor graph optimization method to model the inconsistencies between relative transformations of multiple stations and obtains globally consistent transformation parameters after overall optimization.

[0384] The pose estimation module is used for fine registration and pose estimation of point clouds. It uses the relative pose transformation parameters optimized by factor graph as initial values ​​and optimizes them by point-to-point ICP method. It derives the pose of each scanning station in a unified coordinate system through the optimized transformation parameters and transforms the point clouds of all scanning stations to a unified coordinate system to form a consistent global point cloud.

[0385] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.

[0386] This invention processes multi-station ground-based laser scanning (TLS) point cloud data collected in complex forest (artificial / natural forest) environments, achieving automatic registration of multi-station point clouds. To comprehensively evaluate registration performance, embodiments of this invention were tested on 12 plots from four datasets (Shennongjia National Nature Reserve in Hubei Province, ETH Zurich, FGI, and Tongji University). The accuracy evaluation results are as follows: Figure 13 and Table 1, Figure 14 As shown.

[0387] Figure 13 The coarse registration errors of this invention are shown in various datasets. (a) shows the rotation component, and (b) shows the translation component. Table 1 summarizes the quantitative evaluation data of the coarse registration errors in each dataset. Different colored bars in the figure represent different rotation / translation components, and the presence or absence of dashed shading inside the bars indicates the error value before and after factor plot optimization. Figure 13 As can be summarized in Table 1, the error amount is significantly reduced after factor map optimization, especially for cases with large coarse registration errors. For example, the overall rotation error of plot 1 decreased from 140 mrad to below 70 mrad, and the translation error component also decreased from 60 cm to below 30 cm. In addition, all other plots also showed good performance, demonstrating the superiority of the factor map optimization module in the system proposed in this invention.

[0388] Table 1. Quantitative evaluation results of coarse registration error (including four public datasets: Shennongjia, ETH, FGI, and Tongji University).

[0389]

[0390] Figure 14 The figure shows the registration errors of the present invention in various plots, with (a) and (b) representing the rotation and translation error components, respectively. As can be seen from the figure, the method proposed in this invention achieved good results in all 12 test plots. Regarding the rotation component, except for plot 4 where the error was slightly larger, the overall rotation error of the other plots was below 3 mrad. Regarding the translation component, the error of a single coordinate axis in all plots was around 3 cm. In particular, high-precision results within 1 cm were even achieved in the two plantations, plots 11 and 12, verifying the accuracy and robustness of the method proposed in this invention.

[0391] Finally, using the precise pose obtained by the method of this invention, the point cloud data of each scanning site are transformed into a unified coordinate system to form the global point cloud of the sample plot, as shown below. Figure 15As shown (one sample plot is selected as a representative for each dataset), the automatically registered multi-site cloud can be properly restored to each scan location and completely covers the entire sample plot. This achieves the objective of the proposed method well, namely, to realize efficient, accurate, and automatic registration of multi-site TLS scanning in complex forest environments without any prior information, manual operation, or special markings.

[0392] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A multi-station ground-based laser scanning point cloud registration method suitable for forest environments, characterized in that, Includes the following steps: Step 1: Generate distance image; Step 2: Continuity constraint of tree trunk points; Apply distance continuity constraints along the column direction to extract the stable portion of the point cloud; Step 3: Constrain the orientation of the tree trunk point; The tree trunk points extracted using continuity constraints still contain some dense branches and leaves as well as scattered ground fragments. These are then clustered into point cloud clusters. Subsequently, the geometric characteristics of each cluster are used to remove branches, leaves, and ground points while retaining the tree trunk points. Step 4: Construct the triangular descriptor; Triangles are used to describe the relative relationships between discrete tree trunk locations; multiple triangles are combined to form a triangular network to describe the overall distribution of tree trunks in the scan; the number of triangles depends on the size of the scanned plot and the tree trunk density. Step 5: Coarse registration of 3D point clouds; The algorithm uses efficient hash table search to find triangle matching pairs, calculates the similarity between matching pairs of scanning stations, and outputs the coarse registration parameters of the scan point cloud pairs if the similarity meets the threshold. Step 6: Factor graph optimization; A factor graph optimization method is introduced to model the inconsistencies between relative transformations at multiple stations, and globally consistent transformation parameters are obtained after overall optimization. Step 7: Point cloud fine registration and pose estimation; The relative pose transformation parameters optimized by factor graph are used as initial values ​​and optimized by point-to-point ICP method. The pose of each scanning station in a unified coordinate system is derived by the optimized transformation parameters, and all scanning station point clouds are transformed to a unified coordinate system to form a consistent global point cloud. The continuity constraint of the trunk points: (1) Define the search step size (Step) and the pixel difference threshold (R). diff ), Group pixel number threshold (Seg th The three parameters are set to 15 pixels, 0.03m, and 100 pixels respectively; (2) For each column in the distance image, starting from the first point, compare the pixel value difference between the search point and the current point within the search step size Step. If it is less than the preset threshold R, diff If the search point is in the same group as the current point, then the search point and the current point are grouped together; otherwise, a new group is created. (3) Repeat the above operation until all pixels in the distance image have been accessed; (4) Count the number of pixels in each group. If it exceeds the preset threshold Seg th If a pixel in a group is not found, it is retained and marked as 1; otherwise, it is discarded and marked as 0. (5) Traverse all pixels. If the pixel is marked as 1, keep the corresponding scan point; otherwise, if the pixel is marked as 0, discard the scan point. The tree trunk point orientation constraint: (1) Using Euclidean distance clustering, the point cloud data extracted in step 2 is processed and divided into different point cloud clusters; (2) Perform principal component analysis (PCA) on each point cloud cluster and calculate the linear L. λ Planarity P λ Dispersion S λ Three parameters describe the geometric distribution of point cloud clusters in space, defined by formula (2); In the formula, λ1>λ2>λ3 are the eigenvalues ​​arranged in descending order, which correspond to the degree of change in the direction of the eigenvector; (3) The tree trunk points exhibit a linear distribution in space, that is, they have a large linearity L. λ The ground point has a large planarity P λ The branch and leaf points have a large degree of dispersion. λ Therefore, this characteristic can be used to determine the type of extracted point cloud clusters. If a point cloud cluster satisfies condition L... λ If the value is greater than 0.95, then this point cloud cluster is considered to be a tree trunk; the distribution of point cloud clusters before and after applying geometric constraints; (4) Using the extracted tree trunk point cloud cluster, extract the corresponding tree trunk position; its planar position (x,y) is the center of the point cloud cluster, and the z component is selected as the root point of the tree trunk, that is, the intersection of the tree trunk and the ground. The black dots at the bottom are the extracted tree trunk positions. The construction of the triangle descriptor: (1) Select triangles; 1) For any given tree trunk location, search for 10 nearest neighbor points centered on it; 2) Randomly select two nearest neighbor points and combine them with the current point to form a triangle; 3) If the constructed triangle is an isosceles / equilateral triangle, remove it; otherwise, if the three sides are different, keep it. 4) Repeat the above operation to traverse all tree trunk positions; (2) Hash table storage 1) Construct a hash function using the "division remainder method", defined by formula (3); h(l1,l2,l3)=mod[mod(l1*m,N)+l2*m,N]+l3 (3) In the formula, (l1, l2, l3) are the side lengths of any triangle, and are arranged in descending order (l1 > l2 > l3); m is a large prime number, and N represents the size of the hash table; 2) Perform the above operation on each triangle in the triangulation so that each triangle can be projected onto a unique index in the hash table; 3) All TLS scans in the sample plot are projected to the corresponding positions in the hash table; if a hash collision is encountered, the chaining method is used to handle it; The factor graph optimization: (1) Factor plot modeling Perform multi-site TLS scans on a specific forest plot; Subsequently, a factor graph model was established using the unknown pose of each station scanned in a unified coordinate system as a node and the coarse transformation between adjacent stations as a factor constraint. (2) Global optimization By comprehensively considering all nodes and constraints in the factor graph, the problem is transformed into a least squares problem within a Bayesian estimation framework; by reducing the inconsistencies between relative transformations, the optimal scanning station pose and relative transformation are solved.

2. The multi-station ground-based laser scanning point cloud registration method applicable to forest environments as described in claim 1, characterized in that, The generated distance image: (1) Ground-based lidar data acquisition; TLS scanning stations are deployed according to the size of the forest plot. For a 20m×20m plot, a scanning method of one station in the middle and one station at each of the four corners can be adopted to obtain complete plot scanning data to the greatest extent. (2) Converting 3D point clouds into distance images; Processing single-site point cloud data, let the three-dimensional coordinates of any point i in the point cloud be (x, y, z). i ,y i ,z i It can also be represented as polar coordinates (α). i ,β i ,γ i The form of ) where α i ,β i These are the horizontal angle and the vertical angle, γ. i Let i be the distance from the scan point i to the scan center; then, the index of the scan point i in the distance image can be defined by formula (1); In the formula, row i ,col i Here, i represents the row and column indices of the scan point i in the distance image, respectively, and θ represents the angular resolution of the distance image. Distance to each pixel in the image using γ i For padding, if multiple scan points correspond to the same index position after conversion, the pixel retains the information of the scan point closest to the center.

3. The multi-station ground-based laser scanning point cloud registration method applicable to forest environments as described in claim 1, characterized in that, The coarse registration of the 3D point cloud: (1) Search for triangle matching pairs 1) For any triangle in the scanning site to be searched, use a hash function to directly map the triangle descriptor to the corresponding index in the hash table; 2) Find the triangle descriptor contained at the current index position, calculate the direction vector difference of the corresponding vertex using formula (4), and if it is less than the threshold of 0.1, keep the tree trunk matching pair; otherwise, discard it. In the formula, and These represent the direction vectors of the query trunk and the candidate trunk, respectively. (2) Calculate the coarse parameters of the three-dimensional transformation 1) Count the scanning station IDs corresponding to each tree trunk, vote, calculate the similarity, and define it by formula (5); In the formula, S q With S c These represent query scan and candidate scan, respectively. c N represents the number of votes for the candidate scan. q The number of tree trunks for the candidate scan; 2) If the similarity between the query scan and the candidate scan is greater than the preset threshold, it is assumed that there is an adjacency relationship between the two stations, and the three-dimensional transformation parameters are solved using the paired trunk positions. The point cloud fine registration and pose estimation are as follows: (1) Point cloud precision registration; Using the relative pose transformation parameters optimized by the factor map as initial values, the transformation parameters between each effective paired scanning station are optimized and adjusted using a point-to-point ICP fine registration method. (2) Global pose derivation; Based on the optimized relative transformation parameters, the pose of each scanning station in a unified coordinate system is derived using a depth-first strategy. (3) Unify the point cloud coordinate system; After the above steps, the poses of each scanning station in a unified coordinate system are known and applied to each scanning point cloud, transforming it from an independent local coordinate system into a global point cloud under a unified coordinate system.

4. A multi-station ground-based laser scanning point cloud registration system suitable for forest environments, implementing the multi-station ground-based laser scanning point cloud registration method for forest environments as described in any one of claims 1-3, characterized in that, The multi-station ground-based laser scanning point cloud registration system suitable for forest environments includes: Distance image generation module, used to generate distance images; The continuity constraint module is used for the continuity constraint of the trunk points; it performs distance continuity constraints from the column direction and extracts the stable part of the point cloud. The orientation constraint module is used to constrain the orientation of tree trunk points; it removes the influence of still existing dense branches and leaves as well as scattered ground fragments. The descriptor building module is used to construct triangular descriptors; triangles are used to describe the relative relationships between discrete tree trunk locations; multiple triangles are combined to form a triangular mesh, which is used to describe the overall distribution of tree trunks in the scan; the number of triangles depends on the size of the scanned plot and the tree trunk density; The registration module is used for coarse registration of 3D point clouds. It uses efficient hash table search to find triangle matching pairs, calculates the similarity between matching pairs of scanning stations, and outputs the coarse registration parameters of the scanned point cloud pairs if the similarity meets the threshold. The optimization module is used for factor graph optimization. It introduces the factor graph optimization method to model the inconsistencies between relative transformations of multiple stations and obtains globally consistent transformation parameters after overall optimization. The pose estimation module is used for fine registration and pose estimation of point clouds. It uses the relative pose transformation parameters optimized by factor graph as initial values ​​and optimizes them by the point-to-point ICP method. It derives the pose of each scanning station in a unified coordinate system through the optimized transformation parameters and transforms the point clouds of all scanning stations to a unified coordinate system to form a consistent global point cloud. The continuity constraint of the trunk points: (1) Define the search step size (Step) and the pixel difference threshold (R). diff ), Group pixel number threshold (Seg th The three parameters are set to 15 pixels, 0.03m, and 100 pixels respectively; (2) For each column in the distance image, starting from the first point, compare the pixel value difference between the search point and the current point within the search step size Step. If it is less than the preset threshold R, diff If the search point is in the same group as the current point, then the search point and the current point are grouped together; otherwise, a new group is created. (3) Repeat the above operation until all pixels in the distance image have been accessed; (4) Count the number of pixels in each group. If it exceeds the preset threshold Seg th If a pixel in a group is not found, it is retained and marked as 1; otherwise, it is discarded and marked as 0. (5) Traverse all pixels. If the pixel is marked as 1, keep the corresponding scan point; otherwise, if the pixel is marked as 0, discard the scan point. The tree trunk point orientation constraint: (1) Using Euclidean distance clustering, the point cloud data extracted in step 2 is processed and divided into different point cloud clusters; (2) Perform principal component analysis (PCA) on each point cloud cluster and calculate the linear L. λ Planarity P λ Dispersion S λ Three parameters describe the geometric distribution of point cloud clusters in space, defined by formula (2); In the formula, λ1>λ2>λ3 are the eigenvalues ​​arranged in descending order, which correspond to the degree of change in the direction of the eigenvector; (3) The tree trunk points exhibit a linear distribution in space, that is, they have a large linearity L. λ The ground point has a large planarity P λ The branch and leaf points have a large degree of dispersion. λ Therefore, this characteristic can be used to determine the type of extracted point cloud clusters. If a point cloud cluster satisfies condition L... λ If the value is greater than 0.95, then this point cloud cluster is considered to be a tree trunk; the distribution of point cloud clusters before and after applying geometric constraints; (4) Using the extracted tree trunk point cloud cluster, extract the corresponding tree trunk position; its planar position (x,y) is the center of the point cloud cluster, and the z component is selected as the root point of the tree trunk, that is, the intersection of the tree trunk and the ground. The black dots at the bottom are the extracted tree trunk positions. The construction of the triangle descriptor: (1) Select triangles; 1) For any given tree trunk location, search for 10 nearest neighbor points centered on it; 2) Randomly select two nearest neighbor points and combine them with the current point to form a triangle; 3) If the constructed triangle is an isosceles / equilateral triangle, remove it; otherwise, if the three sides are different, keep it. 4) Repeat the above operation to traverse all tree trunk positions; (2) Hash table storage 1) Construct a hash function using the "division remainder method", defined by formula (3); h(l1,l2,l3)=mod[mod(l1*m,N)+l2*m,N]+l3 (3) In the formula, (l1, l2, l3) are the side lengths of any triangle, and are arranged in descending order (l1 > l2 > l3); m is a large prime number, and N represents the size of the hash table; 2) Perform the above operation on each triangle in the triangulation so that each triangle can be projected onto a unique index in the hash table; 3) All TLS scans in the sample plot are projected to the corresponding positions in the hash table; if a hash collision is encountered, the chaining method is used to handle it; The factor graph optimization: (1) Factor plot modeling For a specific forest plot, a multi-station TLS scan was performed. Subsequently, a factor graph model was established using the unknown pose of each station in a unified coordinate system as a node and the coarse transformation between adjacent stations as a factor constraint. (2) Global optimization By comprehensively considering all nodes and constraints in the factor graph, the problem is transformed into a least squares problem within a Bayesian estimation framework; by reducing the inconsistencies between relative transformations, the optimal scanning station pose and relative transformation are solved.

5. A computer device, characterized in that, The computer device includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the steps of the multi-station ground-based laser scanning point cloud registration method for forest environments as described in any one of claims 1-3.

6. A computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the multi-station ground-based laser scanning point cloud registration method for forest environments as described in any one of claims 1-3.

7. An information data processing terminal, characterized in that, The information data processing terminal is used to implement the multi-station ground-based laser scanning point cloud registration system suitable for forest environments as described in claim 4.

Images