A multi-platform laser point cloud registration method for ginkgo plantations

Abstract

The present invention discloses a multi-platform laser point cloud registration method suitable for ginkgo plantations, and belongs to the field of forestry technology. It includes: S1: data acquisition and preprocessing: acquiring TLS and ULS point cloud data in the target area, and preprocessing the acquired data; S2: performing single-tree segmentation on the point cloud data of each platform to obtain the position information of standing tree positioning points; S3: using the ULS and TLS standing tree positioning features obtained by single-tree segmentation to perform point cloud coarse registration; S4: using the multi-platform point cloud registration results to perform multi-platform point cloud fine registration fusion based on the ICP algorithm; S5: evaluating the registration accuracy. Based on the forest environment, the present invention draws on the idea of ​​coarse-to-fine registration, and proposes a multi-platform point cloud registration algorithm suitable for ginkgo plantations, in order to provide high-precision and more complete three-dimensional point cloud data with unified coordinates.

Description

Technical Field

[0001] The invention belongs to the technical field of forestry, and in particular relates to a multi-platform laser point cloud registration method suitable for a ginkgo plantation. Background Art

[0002] Active remote sensing point cloud data has been widely used in forestry surveys, and the data redundancy and high-precision matching algorithm requirements brought by multi-platform data are increasing. ULS (unmanned aerial vehicle platform laser radar) can obtain a wide range of forest canopy structure parameters from top to bottom and improve the inversion accuracy, but the point cloud information of the lower part of the canopy is missing during the scanning process and cannot reflect the vertical structure of the forest stand. TLS (ground-based platform laser radar) has the advantages of high density and high precision, but the data collection range of TLS is limited. In forest stands with complex structures, it cannot fully reflect the upper canopy information. Multi-platform data fusion can make up for the defects in the two types of data collection and improve the quality of point clouds.

[0003] Among various registration and fusion algorithms, there are three commonly used point cloud registration frameworks, including target-based, feature-based, and point-based methods.

[0004] Target-based methods usually require external information to align lidar point clouds. However, in forest scenes, the external alignment information required by target-based methods is usually either unavailable, inaccurate, or difficult to obtain.

[0005] Feature-based methods are similar to target-based methods in that they use registration points, lines, and polygons to register multi-platform lidar point clouds, but these features are identified inside the point cloud, so feature-based methods are widely used in indoor and urban environments.

[0006] The forest environment is much more complex and irregular than the indoor and urban environments. It is difficult to find regular features like those in the indoor and urban environments, or these features may not exist at all. In some simple plots, the registration is usually performed by visual-based methods, that is, by experience-based judgment, which has poor robustness to highly complex and large-scale forest stands. The point-based method directly performs lidar point cloud registration based on the geometric information provided by the lidar points. In forest scenes where massive point clouds are generated after lidar scanning, iterative algorithms based on the relationship between point pairs are directly used to perform multi-platform point cloud fusion, which often falls into an error cycle. It is necessary to align the initial pose of the point cloud before running the algorithm. The better the initial value, the higher the running efficiency and accuracy of the point-based registration algorithm.

[0007] In 2020, Guan et al. proposed a novel multi-platform LiDAR data registration framework based on the unique spatial distribution of trees in forests in the article "A Novel Framework to Automatically Fuse Multiplatform LiDAR Data in Forest Environments Based on Tree Locations" in Volume 58 of IEEE Transactions on Geoscience and Remote Sensing. Tree pairs from multi-platform LiDAR data are considered as the only features required in the registration process. Tree pairs are identified from multi-platform LiDAR data through TIN matching strategy, and then the target point cloud is matched with the source point cloud, which has high robustness and high accuracy. However, the algorithm cannot run effectively when facing densely planted artificial forests, resulting in the disadvantage of high computational cost in some high-density artificial forest scenes, so it is necessary to improve this problem. Summary of the invention

[0008] Technical problem solved by the present invention: The present invention provides a multi-platform laser radar point cloud registration algorithm framework from coarse to fine, which takes ULS data as target point cloud and TLS data as source point cloud, establishes coarse registration based on standing tree positioning feature algorithm and fine registration of point cloud based on ICP algorithm, and proposes a multi-platform point cloud registration algorithm suitable for ginkgo plantations, in order to provide high-precision and more complete three-dimensional point cloud data with unified coordinates.

[0009] Technical solution: In order to solve the above technical problems, the technical solution adopted by the present invention is as follows:

[0010] A multi-platform laser point cloud registration method suitable for ginkgo plantations includes the following steps:

[0011] S1: Data acquisition and preprocessing: Acquire TLS and ULS point cloud data in the target area and preprocess the acquired data;

[0012] S2: Perform single tree segmentation on the point cloud data of each platform to obtain the location information of the standing tree positioning points;

[0013] S3: Coarse registration of point clouds using ULS and TLS standing tree positioning features obtained by single tree segmentation;

[0014] S4: Use the multi-platform point cloud registration results to perform multi-platform point cloud precise registration and fusion based on the ICP algorithm;

[0015] S5: Evaluate the registration accuracy.

[0016] Preferably, in S1, data acquisition and preprocessing: acquiring TLS and ULS point cloud data in the target area, and preprocessing the acquired data, specifically including:

[0017] The acquired data are stitched, reconstructed, denoised, ground point classified and elevation normalized. The acquired TLS point cloud is stitched at multiple sites, and the ground points are separated using elevation filtering. The acquired ULS point cloud is reconstructed and the ground points are separated. Based on the ground point separation, the TLS and ULS point clouds are respectively normalized in elevation to eliminate the influence of terrain undulations, and the point clouds of non-target trees on the ground are further removed in combination with visual interpretation.

[0018] Preferably, in S2, the point cloud data of each platform is segmented into individual trees to obtain the specific content of the standing tree positioning point position information, including:

[0019] S201: TLS single tree segmentation: The bottom-up PCS algorithm is used to find seed points by identifying the trunk diameter at breast height attribute, and the density-based noise application spatial clustering algorithm is used to detect the trunk, and then the single tree segmentation is performed based on the seed points;

[0020] S202: ULS single tree segmentation: A top-down PCS algorithm is used to identify the location of a single tree. The algorithm starts from the global highest point and estimates lower points based on the spacing threshold and minimum spacing rule. The point is determined by comparing the shortest distance to achieve single tree segmentation.

[0021] Preferably, in S3, the specific contents of performing point cloud coarse registration using ULS and TLS standing tree positioning features obtained by single tree segmentation include:

[0022] S301: extracting the plane X and Y position coordinates of each segmented tree in the ULS and TLS point clouds; taking each extracted tree position coordinate as a search point, constructing a circle with each search point as the center to run the KNN algorithm, thereby searching for neighboring tree positions;

[0023] S302: Perform Delaunay triangulation using the neighboring points searched within the circle to construct a TIN triangulation network;

[0024] S303: Annotate the triangle information of the TIN triangulated networks constructed on different platforms, identify equilateral triangles, isosceles triangles and highly collinear triangles during the annotation process, calculate the centroid of each triangle using the corner points of each triangle, and establish triangle index numbers starting from 1 from high to low according to the y value of the centroid coordinates on the plane coordinate axis;

[0025] S304: Calculate the area similarity and angle similarity of each pair of TIN triangulation networks constructed by TLS and ULS respectively, achieve similarity matching, and construct a TSM similarity matrix;

[0026] S305: Obtain a score matrix, select a TIN pair with a maximum score through the score matrix, and determine through a threshold value to ensure that there is a unique corresponding TIN pair with the maximum score between the TLS and ULS platforms;

[0027] S306: Define the Z value, and solve the rotation matrix and translation vector through singular value decomposition;

[0028] S307: The TLS point cloud is used as the source point cloud, the ULS point cloud is used as the target point cloud, the rotation matrix and translation vector are applied to the target point cloud, and the coordinates are transformed to the location of the source point cloud to achieve coarse alignment.

[0029] Preferably, in S304, the area similarity and angle similarity are calculated for each pair of TIN triangulation networks constructed by TLS and ULS respectively to achieve similarity matching. The specific contents of constructing the TSM similarity matrix include:

[0030] TS mn Represents the overall similarity between the mth triangle in the TIN constructed by the TLS platform and the nth triangle in the TIN constructed by the ULS platform. The overall similarity calculation formula is:

[0031]

[0032] The area similarity calculation formula is:

[0033]

[0034] The angle similarity calculation formula is:

[0035] I=(I A +I B +I C ) / 3,

[0036] Among them, S is the area similarity, I is the angle similarity, and I A ,I B ,I C Represents the angular similarity of triangle A, B, and C respectively; l a and l b are two sides of a triangle in the base TIN, l a ′ and l b ′ are the two sides of the triangle in the airborne TIN, C and C′ are l a With l b These two edges and l a 'with lb 'The angle between these two sides.

[0037] Preferably, in S305, a score matrix is ​​obtained, a TIN pair with a maximum score is selected through the score matrix, and a threshold is determined to ensure that there is a unique corresponding TIN pair with a maximum score between the TLS and ULS platforms. The specific contents include:

[0038] S3051: After the TSM matrix is ​​constructed, the overall maximum value in the matrix is ​​found by traversal, and the value is reassigned to 1, and the remaining elements of the row and column where the value is located are marked as 0. On the basis of retaining the above results, the maximum value in the remaining rows and columns is retrieved and assigned. This operation is continued until a score matrix consisting of 1s and 0s with equal weights is obtained, and the sum of the scores of 1 is the similarity score of the TIN pair;

[0039] S3052: Sort the TIN pairs by similarity scores and select the TIN pair with the maximum score; make a determination based on the defined score threshold to ensure that there is a unique corresponding TIN pair with the maximum score between the TLS and ULS platforms;

[0040] S3053: Obtain a set of tree pair coordinates from lidar datasets from different platforms, and use Open CV with default parameters to call the RANSAC algorithm to filter and optimize the matching tree pairs.

[0041] Preferably, in S306, the Z value is defined, and the specific contents of solving the rotation matrix and the translation vector by singular value decomposition include:

[0042] S3061: Define the Z value of tree locations, expressed as the average elevation of all ground points within a 1m buffer of the tree, in order to minimize the impact of different tree locations from different platforms;

[0043] S3062: Using the X, Y, and Z coordinates of the matching tree pairs in the multi-platform dataset, the rotation matrix and translation vector are solved by singular value decomposition. The specific formula is:

[0044]

[0045] Among them, X, Y, Z are the coordinates of the source point cloud, X', Y', Z' are the coordinates of the target point cloud, and R and T are the rotation matrix and translation vector respectively.

[0046] Preferably, in S4, the specific contents of using the multi-platform point cloud registration results to perform multi-platform point cloud precise registration fusion based on the ICP algorithm include:

[0047] S401: Preliminary alignment of multi-platform point clouds based on the rough registration matrix obtained by the standing tree positioning feature algorithm;

[0048] S402: For each point in the source point cloud, find the nearest point in the target point cloud and establish a matching relationship;

[0049] S403: Calculate a new transformation matrix using the matched point pairs, which maps the target point cloud into the space of the source point cloud;

[0050] S404: applying the calculated transformation to the target point cloud to update the position of the target point cloud;

[0051] S405: Iteration, repeat the above steps until convergence.

[0052] Preferably, in S5, the root mean square error is used to evaluate the registration accuracy, and the specific calculation formula is:

[0053]

[0054] Among them, RMSE represents the average distance error between the registered point cloud and the reference point cloud, n represents the number of point pairs considered in calculating the error, and y i represents the point position in the source point cloud, is the position of the midpoint of the target point cloud after registration.

[0055] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0056] (1) The present invention runs the KNN algorithm by using a circular frame to define the search range. By limiting the search radius, the overhead of calculating the distance between the target data point and all training data points is reduced, thereby improving the efficiency of the algorithm. In locally dense areas, combined with radius search, the local data distribution can be better captured.

[0057] (2) The present invention uses standing tree positioning features for point cloud fusion, which has the characteristic of strong robustness. Similar features can be extracted even from point cloud data taken at different times, in different weather conditions and under different lighting conditions. This robustness can make the results of multi-platform and multi-temporal fusion of point clouds more accurate and stable.

[0058] (3) The computational complexity of the standing tree positioning feature of the present invention is also relatively low. A better initial transformation is obtained between the source point cloud and the target point cloud after rough alignment, which reduces the time and error of manual intervention and can improve the speed and accuracy of fine alignment. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 is a flow chart of the present invention;

[0060] Figure 2 This is the front view of the TLS point cloud single tree segmentation result of the ginkgo in the study area;

[0061] Figure 3 This is a top view of the ULS point cloud single tree segmentation results of the ginkgo in the study area;

[0062] Figure 4 This is a schematic diagram of the TLS platform using a circle to define the KNN search range and construct a TIN triangulation network;

[0063] Figure 5 This is a schematic diagram of using a circle to define the KNN search range and construct a TIN triangulation network on the ULS platform;

[0064] Figure 6 It is a TIN triangulated network triangle information annotation map;

[0065] Figure 7 It is a schematic diagram of the similarity matrix TSM;

[0066] Figure 8 This is the result of the multi-platform point cloud rough registration based on the standing tree positioning feature algorithm;

[0067] Fig. 9 This is the result of precise registration based on the ICP algorithm;

[0068] Fig.10 is an example single wood image extracted from the coarse registration result;

[0069] Fig.11 is an example single tree image extracted from the fine registration result. DETAILED DESCRIPTION

[0070] The present invention is further illustrated below in conjunction with specific examples. The examples are implemented based on the technical solutions of the present invention. It should be understood that these examples are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0071] The present invention uses the urban area of ​​Nanjing, Jiangsu Province, China (32.078°N, 118.816°E) as the research area, and the main tree species in the research area, Ginkgo biloba, as the research object. In the research area, RIEGL VZ-400i ground-based laser radar is used to scan at 36 stations distributed in a fixed array using a fixed and repeated process to obtain TLS point cloud data. DJI M300RTK carrying DJI Zenmuse L1 laser radar is used to collect drone laser data to obtain ULS point cloud data.

[0072] like Figure 1 As shown, a multi-platform laser point cloud registration method suitable for ginkgo plantations includes the following steps:

[0073] S1: Data acquisition and preprocessing: The original data of TLS and ULS point clouds are obtained in the target area, and the obtained data are processed by stitching, reconstruction, denoising, ground point classification, elevation normalization, etc. The point cloud contains the spatial position coordinate information of each point. The obtained TLS point cloud is stitched at multiple sites, and the ground points are separated by elevation filtering. The obtained ULS point cloud is reconstructed and the ground points are separated. On the basis of ground point separation, the TLS and ULS point clouds are respectively normalized by elevation to eliminate the influence of terrain undulations, and the point clouds of non-target trees on the ground are further removed in combination with visual interpretation. The TLS elevation normalized point cloud and the ULS elevation normalized point cloud are used as the data source for multi-platform point cloud registration.

[0074] Table 1 below shows the parameters of the ground-based laser radar RIEGL VZ-400i, and Table 2 below shows the DJI drone setting parameters.

[0075] Table 1 TLS scanning parameters

[0076] Parameter name Parameter Value Scanning accuracy 5mm / 100m Laser ranging principle Pulse ranging method Laser emission frequency (kHz) 100～120 Measuring range(m) 1.5～800 Scanning speed (points / second) 500000 Scanning angle range (FOV) 100° vertical × 360° horizontal Laser safety level Laser Class (eye-safe laser) Machine size (width x height) / weight 206mm×308mm / 9.7kg

[0077] Table 2 Main parameter settings of UAV

[0078] Flight speed 5-8m / s Height of the photographed surface relative to the take-off point 0 Advanced Settings - Laser Overlap Overlap more than 80% Load Setting - Echo Mode Three echoes Load Setting-Sampling Frequency 160kHZ Load Setting - Scan Mode Repeat Scan Flight altitude 80-120 meters

[0079] S2: Perform single tree segmentation on the point cloud data of each platform to obtain the location information of the standing tree positioning points, including:

[0080] S201: TLS single tree segmentation, using a bottom-up point cloud segmentation algorithm (Point Cloud Segmentation, PCS), by identifying the trunk diameter attributes to find seed points, the present invention uses a density-based noise application spatial clustering algorithm (DBSCAN) to detect tree trunks, and uses a point cloud at a height of 1.2 to 1.3m as input data during trunk detection, and then performs single tree segmentation based on the seed points.

[0081] S202: ULS single tree segmentation uses a top-down PCS algorithm to identify the location of a single tree. The algorithm starts from the global highest point and estimates lower points based on the spacing threshold and minimum spacing rule. It uses the size of the shortest distance to determine the point's ownership, thereby achieving single tree segmentation.

[0082] S3: Use the ULS and TLS standing tree positioning features obtained by single tree segmentation to perform point cloud coarse registration, including:

[0083] S301: Extract the plane X, Y position coordinates of each segmented tree in the ULS and TLS point clouds; for TLS and ULS, use each extracted tree position coordinate as a search point, construct a circle with each search point as the center, and run the KNN algorithm to search for neighboring tree positions.

[0084] S302: For TLS and ULS, the neighboring points searched within the circle are used to perform Delaunay triangulation to construct a TIN triangulation network. To avoid generating the same TIN, the search points are not involved in the triangulation.

[0085] S303: Annotate the triangle information of the TIN triangulated networks constructed on different platforms, annotate the triangle information in the TIN, including the internal angles A, B, C and the lengths of each side, define the largest angle in the triangle as angle C, the second largest angle as angle B, the smallest angle as angle A, and the opposite side of angle A is side a (BC), the opposite side of angle B is side b (AC), and the opposite side of angle C is side c (AB). During the annotation process, identify equilateral triangles, isosceles triangles, and highly collinear triangles (two angles in the triangle are both <5 degrees), use the corner points of each triangle to calculate the centroid of each triangle, and establish the triangle index number starting from 1 from high to low according to the size of the centroid coordinates' y value on the plane coordinate axis.

[0086] S304: Calculate the area similarity S and the angle similarity I for each pair of TIN triangulation networks constructed by TLS and ULS, achieve similarity matching, and construct a TSM similarity matrix;

[0087] In order to improve the computational efficiency, the bidirectional KD tree algorithm is used to quickly find and calculate similar triangles between different platforms.

[0088] TS mn Represents the overall similarity between the mth triangle in the TIN constructed by the TLS platform and the nth triangle in the TIN constructed by the ULS platform. The overall similarity calculation formula is:

[0089]

[0090] The calculation formula of area similarity S is:

[0091]

[0092] The calculation formula of angle similarity I is:

[0093] I=(I A +I B +I C ) / 3,

[0094] in:

[0095]

[0096]

[0097]

[0098] Among them, S is the area similarity, I is the angle similarity, and I A ,I B ,I C Respectively represent the angular similarity of triangle A, B, and C, u(A), u(B), and u(C) represent the functions of triangle A, B, and C based on Gaussian distribution; l a and l b are two sides of a triangle in the base TIN, l a ′ and l b ′ are the two sides of the triangle in the airborne TIN, C and C′ are l a With l b These two edges and l a ′ and l b ′The angle between these two sides.

[0099] S305: Obtain a score matrix, select the TIN pair with the maximum score through the score matrix, and determine through a threshold to ensure that there is a unique corresponding TIN pair with the maximum score between the TLS and ULS platforms. The specific contents include:

[0100] S3051: After the TSM matrix is ​​constructed, the overall maximum value in the matrix is ​​found by traversal, and the value is reassigned to 1, and the remaining elements in the row and column where the value is located are marked as 0. On the basis of retaining the above results, the maximum value in the remaining rows and columns is retrieved and assigned. This operation is continued until a score matrix consisting of 1s and 0s with equal weights is obtained, and the sum of the scores of 1 is the similarity score of the TIN pair.

[0101] S3052: Sort the TIN pairs by similarity scores and select the TIN pair with the maximum score. Here, it is necessary to make a judgment based on the defined score threshold to ensure that there is a unique corresponding TIN pair with the maximum score between the TLS and ULS platforms.

[0102] S3053: From the above process, a set of tree pair coordinates can be obtained from the lidar datasets from different platforms. To ensure the quality of coarse registration, the RANSAC algorithm is called using Open CV with default parameters to filter and optimize the matched tree pairs. Considering that the distance measurements from different lidar platforms are highly accurate, the scale factor of the rotation matrix is ​​set to 1.

[0103] S306: Define the Z value, and solve the rotation matrix R and translation vector T through singular value decomposition. The specific contents include:

[0104] S3061: Define the Z value of tree locations, expressed as the average elevation of all ground points within a 1m buffer of the tree, to minimize the impact of different tree locations from different platforms.

[0105] S3062: Using the X, Y, and Z coordinates of the matching tree pairs in the multi-platform dataset, the rotation matrix R and the translation vector T are solved by singular value decomposition. The specific formula is:

[0106]

[0107] Among them, X, Y, Z are the coordinates of the source point cloud (TLS), X', Y', Z' are the coordinates of the target point cloud (ULS), and R and T are the rotation matrix and translation vector respectively.

[0108] S307: The TLS point cloud is used as the source point cloud, the ULS point cloud is used as the target point cloud, the rotation matrix R and the translation vector T are applied to the target point cloud, and the coordinates are transformed to the location of the source point cloud.

[0109] S4: Use the multi-platform point cloud registration results based on the standing tree positioning feature algorithm to perform point cloud precision registration based on the ICP algorithm. The specific contents include:

[0110] S401: Preliminary alignment of multi-platform point clouds is achieved based on the coarse registration matrix obtained based on the standing tree positioning feature algorithm.

[0111] S402: For each point in the source point cloud, find the nearest point in the target point cloud and establish a matching relationship.

[0112] S403: Calculate a new transformation matrix using the matched point pairs, which maps the target point cloud into the space of the source point cloud.

[0113] S404: Apply the calculated transformation to the target point cloud to update the position of the target point cloud.

[0114] S405: Iteration, repeat the above steps until convergence.

[0115] S5: Use the root mean square error (RMSE) to evaluate the registration accuracy: RMSE represents the average distance error between the registered point cloud and the reference point cloud. The square of the distance error of each point is calculated, and then the average is taken and the square root is taken. RMSE provides a quantitative indicator for measuring the overall matching degree between the two point clouds. A smaller RMSE value indicates a better registration effect. The calculation formula is:

[0116]

[0117] Where n represents the number of point pairs considered when calculating the error, yi represents the point position in the source point cloud, is the position of the midpoint of the target point cloud after registration.

[0118] Based on the feature that the positions of trees are generally unchanged, the present invention performs multi-platform laser point cloud registration of ginkgo plantations. The positions of individual trees in TLS and ULS are extracted by single tree segmentation. Figure 2 This is the front view of the single tree segmentation result of TLS. Figure 3 This is a top view of the ULS single tree segmentation result, where the TLS single tree correct detection rate reaches 100%, and the ULS single tree correct detection rate reaches 93%. Then, the KNN search range is defined by the circle frame to find the location points of the neighboring trees to construct the TIN irregular triangulation network, such as Figure 4 , Figure 5 As shown in Figure 2, the circle range diagram and triangulation diagram of TLS and ULS are shown respectively. Figure 6 The marked triangle information including side length and angle is traversed to calculate the similarity of each irregular triangulated network of TLS and ULS, and then construct the following Figure 7 The similarity matrix TSM shown in Figure 1 is used to obtain the optimal matching irregular triangulated network pair. The matching tree pairs are screened using the RANSAC algorithm, and then Figure 8 As shown in , the rotation matrix and translation vector are obtained by singular value decomposition, thereby achieving rough registration of point clouds. Fig. 9 As shown in FIG. 1 , the ICP algorithm is used to further perform point cloud fine registration based on the rough registration. Fig.10 This is an example tree extracted from the coarse registration result. The registration accuracy is calculated and the RMSE between TLS and ULS is 0.864m. Fig.11 This is an example of a single tree extracted from the precise registration result. The RMSE between TLS and ULS obtained after the registration accuracy calculation is 0.095m. The present invention is verified by an example. It can be seen that the present invention has the characteristics of high multi-platform point cloud registration accuracy in the ginkgo plantation. After ICP precise registration fusion, the final accuracy RMSE is 0.095m. It has certain practical value.

[0119] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A multi-platform laser point cloud registration method suitable for ginkgo plantations, characterized by: The following steps are involved: S1: Data acquisition and preprocessing: Acquire TLS and ULS point cloud data in the target area and preprocess the acquired data; S2: Perform single tree segmentation on the point cloud data of each platform to obtain standing tree positioning features; S201: TLS single tree segmentation: The bottom-up PCS algorithm is used to find seed points by identifying the trunk diameter at breast height attribute, and the density-based noise application spatial clustering algorithm is used to detect the trunk, and then the single tree segmentation is performed based on the seed points; S202: ULS single tree segmentation: A top-down PCS algorithm is used to identify the location of a single tree. The algorithm starts from the global highest point and estimates lower points based on the spacing threshold and minimum spacing rule. The point is determined by comparing the shortest distance, thereby achieving single tree segmentation. S3: Coarse registration of point clouds using ULS and TLS standing tree positioning features obtained by single tree segmentation; S4: Use the multi-platform point cloud registration results to perform multi-platform point cloud precise registration and fusion based on the ICP algorithm; S401: Preliminary alignment of multi-platform point clouds based on the rough registration matrix obtained by the standing tree positioning feature algorithm; S402: For each point in the source point cloud, find the nearest point in the target point cloud and establish a matching relationship; S403: Calculate a new transformation matrix using the matched point pairs, which maps the target point cloud into the space of the source point cloud; S404: applying the calculated transformation to the target point cloud to update the position of the target point cloud; S405: Iteration, repeating the above steps until convergence; S5: Evaluate the registration accuracy.

2. The multi-platform laser point cloud registration method applicable to ginkgo plantations according to claim 1, characterized in that: In S1, data collection and preprocessing: TLS and ULS point cloud data are acquired in the target area, and the acquired data are preprocessed, including: The acquired data are stitched, reconstructed, denoised, ground point classified and elevation normalized. The acquired TLS point cloud is stitched at multiple sites, and the ground points are separated using elevation filtering. The acquired ULS point cloud is reconstructed and the ground points are separated. Based on the ground point separation, the TLS and ULS point clouds are respectively normalized in elevation to eliminate the influence of terrain undulations, and the point clouds of non-target trees on the ground are further removed in combination with visual interpretation.

3. The multi-platform laser point cloud registration method applicable to ginkgo plantations according to claim 1 is characterized in that: In S3, the specific process of using the ULS and TLS standing tree positioning features obtained by single tree segmentation for point cloud coarse registration is as follows: S301: extracting the plane X and Y position coordinates of each segmented tree in the ULS and TLS point clouds; taking each extracted tree position coordinate as a search point, constructing a circle with each search point as the center to run the KNN algorithm, thereby searching for neighboring tree positions; S302: Perform Delaunay triangulation using the neighboring points searched within the circle to construct a TIN triangulation network; S303: Annotate the triangle information of the TIN triangulated networks constructed on different platforms, identify equilateral triangles, isosceles triangles and highly collinear triangles during the annotation process, calculate the centroid of each triangle using the corner points of each triangle, and establish triangle index numbers starting from 1 from high to low according to the y value of the centroid coordinates on the plane coordinate axis; S304: Calculate the area similarity and angle similarity of each pair of TIN triangulation networks constructed by TLS and ULS respectively, achieve similarity matching, and construct a TSM similarity matrix; S305: Obtain a score matrix, select a TIN pair with a maximum score through the score matrix, and determine through a threshold value to ensure that there is a unique corresponding TIN pair with the maximum score between the TLS and ULS platforms; S306: Define the Z value, and solve the rotation matrix and translation vector through singular value decomposition; S307: The TLS point cloud is used as the source point cloud, the ULS point cloud is used as the target point cloud, the rotation matrix and translation vector are applied to the target point cloud, and the coordinates are transformed to the location of the source point cloud to achieve coarse alignment.

4. The multi-platform laser point cloud registration method applicable to ginkgo plantations according to claim 3 is characterized by: In S304, the area similarity and angle similarity are calculated for each pair of TIN triangulation networks constructed by TLS and ULS respectively to achieve similarity matching. The specific contents of constructing the TSM similarity matrix include: TS mn Represents the overall similarity between the mth triangle in the TIN constructed by the TLS platform and the nth triangle in the TIN constructed by the ULS platform. The overall similarity calculation formula is: The area similarity calculation formula is: The angle similarity calculation formula is: I=(I A +I B +I C ) / 3, Among them, S is the area similarity, I is the angle similarity, and I A ,I B ,I C Represents the angular similarity of triangle A, B, and C respectively; l a and l b are the two sides of a triangle in the base TIN, l a ′ and l b ′ are the two sides of the triangle in the airborne TIN, C and C′ are l a With l b These two edges and l a ′ and l b ′The angle between these two sides.

5. The multi-platform laser point cloud registration method applicable to ginkgo plantations according to claim 3 is characterized by: In S305, a score matrix is ​​obtained, and a TIN pair with the maximum score is selected through the score matrix, and a threshold is determined to ensure that there is a unique corresponding TIN pair with the maximum score between the TLS and ULS platforms. The specific contents include: S3051: After the TSM matrix is ​​constructed, the overall maximum value in the matrix is ​​found by traversal, and the value is reassigned to 1, and the remaining elements of the row and column where the value is located are marked as 0. On the basis of retaining the above results, the maximum value in the remaining rows and columns is retrieved and assigned. This operation is continued until a score matrix consisting of 1s and 0s with equal weights is obtained, and the sum of the scores of 1 is the similarity score of the TIN pair; S3052: Sort the TIN pairs by similarity scores and select the TIN pair with the maximum score; make a determination based on the defined score threshold to ensure that there is a unique corresponding TIN pair with the maximum score between the TLS and ULS platforms; S3053: Obtain a set of tree pair coordinates from lidar datasets from different platforms, and use Open CV with default parameters to call the RANSAC algorithm to filter and optimize the matching tree pairs.

6. The multi-platform laser point cloud registration method applicable to ginkgo plantations according to claim 3 is characterized by: In S306, the Z value is defined, and the specific contents of solving the rotation matrix and the translation vector by singular value decomposition include: S3061: Define the Z value of tree locations, expressed as the average elevation of all ground points within a 1m buffer of the tree, in order to minimize the impact of different tree locations from different platforms; S3062: Using the X, Y, and Z coordinates of the matching tree pairs in the multi-platform dataset, the rotation matrix and translation vector are solved by singular value decomposition. The specific formula is: Among them, X, Y, Z are the coordinates of the source point cloud, X', Y', Z' are the coordinates of the target point cloud, and R and T are the rotation matrix and translation vector respectively.

7. The multi-platform laser point cloud registration method applicable to ginkgo plantations according to claim 1, characterized in that: In S5, the root mean square error is used to evaluate the registration accuracy. The specific calculation formula is: Among them, RMSE represents the average distance error between the registered point cloud and the reference point cloud, n represents the number of point pairs considered when calculating the error, yi represents the point position in the source point cloud, is the position of the midpoint of the target point cloud after registration.

Images