A natural forest multi-source point cloud registration method and system based on point cloud segmentation optimization

Abstract

This invention relates to the field of forest resource survey and monitoring technology, and discloses a method and system for multi-source point cloud registration in natural forests based on point cloud segmentation optimization. The method involves setting up sample plots and collecting data on individual tree diameter at breast height (DBH), tree height, and crown width. RTK is used to obtain the spatial coordinates of individual trees. UAV laser scanning and backpack laser scanning are employed for point cloud data acquisition and preprocessing in the study area. The ICP algorithm is used to align the source and target point clouds using rotation and translation. Point cloud registration is achieved by minimizing the error function. Evaluation indicators are used to quantify the accuracy and error of point cloud registration, and point cloud registration is evaluated. Therefore, this invention effectively improves the accuracy of point cloud registration through a segmented registration strategy. By fusing BLS and ULS multi-source point cloud data, it overcomes the limitations of single-sensor data, significantly improving the extraction accuracy of tree structural parameters. This method is not only applicable to point cloud registration in natural forests but also provides reliable data support and technical assurance for forest resource monitoring and ecological protection.

Description

Technical Field

[0001] This invention relates to the field of forest resource survey and monitoring technology, and in particular to a method and system for multi-source point cloud registration of natural forests based on point cloud segmentation optimization. Background Technology

[0002] Forests play a vital role in maintaining ecosystem nutrient cycling and mitigating global climate change. Accurate forest parameters (such as diameter at breast height (DBH) and tree height) are crucial for developing effective conservation and management policies, enhancing forest carbon sequestration capacity, and ensuring the sustainable development of forest ecosystems. Traditional forest surveys provide accurate data but are often time-consuming, labor-intensive, and can potentially disrupt forest ecosystems. In contrast, remote sensing surveys have become the mainstream method due to their efficiency, wide coverage, and timeliness. LiDAR technology, as a non-invasive laser scanning technique, can efficiently acquire high-resolution point cloud data and provide detailed three-dimensional forest information, making it an important, precise, and efficient tool for forest resource monitoring and ecological research.

[0003] LiDAR systems can be categorized into three main types based on their sensor mounting platforms: Ground-based LiDAR (TLS), Airborne LiDAR (ALS), and Unmanned Aerial Vehicle (ULS). TLS is primarily used for large-scale applications, such as snow-covered areas, rivers, and forests. ALS, with its top-down scanning perspective, is better suited for capturing tree canopy and height information, making it ideal for forest canopy data acquisition. ULS offers high flexibility and point cloud density, making it suitable for localized research. BLS, on the other hand, is widely used in practice due to its high mobility and ease of operation.

[0004] However, existing technologies have certain problems in the registration of natural forest point cloud data, especially in the fusion and registration of multi-source data. Due to the mutual occlusion of tree canopies, BLS has difficulty penetrating the canopy to capture the structure of individual trees, leading to errors in the estimation of tree height and diameter at breast height (DBH). ULS is also affected by canopy interference; the scattered distribution of low-level point clouds easily introduces noise, thus affecting the accuracy of parameter extraction. In point cloud registration, the Iterative Closest Point (ICP) algorithm is widely used. Its basic principle is to continuously adjust the positional relationship between the source and target point clouds through rotation and translation to minimize the distance between corresponding points and achieve high-precision alignment. The ICP algorithm can be divided into two types: point-to-point and point-to-plane. Among them, the point-to-point method directly minimizes the Euclidean distance between the source and target points, which is computationally simple and converges quickly. The point-to-plane method uses the normal vector of the target point to establish the geometric relationship between the source point and the target plane through the covariance matrix and singular value decomposition (SVD), which can improve the stability of rotation estimation to a certain extent. Both methods have their advantages and disadvantages, but when there is a lot of noise or inconsistent data quality, the ICP algorithm may still get stuck in a local optimum, leading to a decrease in registration accuracy.

[0005] Therefore, how to improve the registration accuracy of multi-source point cloud data, reduce noise interference and the impact of local optima in complex natural forest environments is a problem that needs to be solved by existing technologies. Summary of the Invention

[0006] This invention provides a method and system for multi-source point cloud registration in natural forests based on point cloud segmentation optimization, aiming to solve at least one of the above-mentioned technical problems.

[0007] To achieve the above objectives, this invention provides a method for multi-source point cloud registration in natural forests based on point cloud segmentation optimization, the method comprising the following steps:

[0008] S1: Set up sample plots and collect data on the diameter at breast height, tree height and crown width of individual trees, and obtain the spatial coordinates of individual trees using RTK;

[0009] S2: Point cloud data of the study area is collected by using UAV laser scanning and backpack laser scanning, and point cloud data from different data sources in the study area are preprocessed.

[0010] S3: The ICP algorithm is used for point cloud registration. Rotation and translation are used to align the source point cloud with the target point cloud. Point cloud registration is achieved by minimizing the error function.

[0011] S4: Root mean square error, average distance error, vertical error, and horizontal error are used as evaluation indicators for point cloud registration algorithms to quantify point cloud registration accuracy and error, and to evaluate point cloud registration.

[0012] Optionally, step S1: Set up sample plots and collect data on individual tree diameter at breast height (DBH), tree height, and crown width, and obtain the spatial coordinates of individual trees using RTK. This specifically includes:

[0013] S11: Establish 9 sample plots with 3 density types; wherein, the density types include: Type I with a density of 700-1100 trees / hectare, Type II with a density of 1100-1400 trees / hectare, and Type III with a density of 1400-1800 trees / hectare;

[0014] S12: Collect the diameter at breast height (DBH), tree height, and crown width of each target tree species in each sample plot. Use an RTK rover station to record the location of the target tree species and use differential correction for accurate positioning.

[0015] Optionally, step S2: Point cloud data of the study area is collected using UAV laser scanning and backpack laser scanning. Preprocessing of point cloud data from different data sources in the study area is performed, specifically including:

[0016] S21: Use a drone laser scanning device to acquire ULS point cloud data of the study area, and use a backpack laser scanning device to acquire BLS point cloud data of the study area.

[0017] S22: Perform preprocessing on the ULS point cloud data and BLS point cloud data collected in the study area, including data computation, cropping, denoising, normalization and ground point separation.

[0018] Optionally, step S3: Point cloud registration is performed using the ICP algorithm, which aligns the source point cloud with the target point cloud using rotation and translation, and achieves point cloud registration by minimizing the error function. Specifically, this includes:

[0019] S31: The ICP algorithm is used to align the ULS point cloud data (source point cloud) with the BLS point cloud data (target point cloud) using rotation and translation matrices, and point cloud registration is achieved by minimizing the error function; wherein, the expression for the point cloud registration is as follows:

[0020]

[0021]

[0022]

[0023] In the formula, It is the optimal rotation matrix. It is the optimal translation matrix. It is a collection of source point clouds. It is a collection of target point clouds. It is a rotation matrix. It is a translation matrix. It represents the number of iterations.

[0024] Optionally, the ICP algorithm adopts the point-to-surface ICP algorithm, which is based on the covariance matrix and singular value decomposition, and calculates the rotation and translation parameters through the source point and the target surface normal vector.

[0025] Optionally, the ICP algorithm is a point-to-point ICP algorithm, which calculates the rotation and translation parameters by directly minimizing the Euclidean distance between corresponding point pairs.

[0026] Optionally, in step S31, the ICP algorithm is used to align the ULS point cloud data (as the source point cloud) with the BLS point cloud data (as the target point cloud) using rotation and translation matrices. Specifically, this includes:

[0027] S311: Under the ICP algorithm, a segmented registration strategy is adopted, with backpack laser scanning to capture the tree trunk and UAV laser scanning to obtain information on the tree crown and tree height;

[0028] S312: Perform height filtering on the ULS point cloud data, segment the data using a preset segmented registration strategy, gradually remove low-level points, use each segment as the registration point cloud, use the BLS point cloud as the reference point cloud, iteratively adjust the rotation matrix and translation matrix, find the corresponding point pairs to achieve point cloud registration.

[0029] Optionally, in step S312, the iteration condition is: stop when the change in RMSE is less than 0.001m and the change in MDErr is less than 0.0001m, or when a maximum of 50 iterations have been performed.

[0030] Optionally, step S4: Use root mean square error, average distance error, vertical error, and horizontal error as evaluation metrics for the point cloud registration algorithm. The specific expressions are as follows:

[0031]

[0032]

[0033]

[0034]

[0035] In the formula, RMSE represents the root mean square error, MDErr represents the average distance error, VErr represents the vertical error, and HErr represents the horizontal error. The distance between each pair of corresponding points. It is the number of corresponding point pairs. These are the coordinates of the measurement point. These are the reference coordinates.

[0036] Furthermore, to achieve the above objectives, the present invention also provides a multi-source point cloud registration system for natural forests based on point cloud segmentation optimization, comprising:

[0037] The data acquisition module is used to set up sample plots and collect data on the diameter at breast height (DBH), tree height, and crown width of individual trees, and to obtain the spatial coordinates of individual trees using RTK.

[0038] The processing module is used to collect point cloud data of the study area using UAV laser scanning and backpack laser scanning, and to preprocess point cloud data from different data sources in the study area.

[0039] The registration module is used to perform point cloud registration using the ICP algorithm. It aligns the source point cloud with the target point cloud by rotation and translation, and achieves point cloud registration by minimizing the error function.

[0040] The evaluation module is used to evaluate point cloud registration algorithms by employing root mean square error, average distance error, vertical error, and horizontal error as evaluation metrics. It quantifies the accuracy and error of point cloud registration and performs point cloud registration evaluation.

[0041] The beneficial effects of this invention are as follows: It proposes a multi-source point cloud registration method and system for natural forests based on point cloud segmentation optimization. This involves setting up sample plots and collecting data on individual tree diameter at breast height (DBH), tree height, and crown width, combined with RTK to obtain the spatial coordinates of individual trees. UAV laser scanning and backpack laser scanning are used to collect point cloud data from different data sources in the study area, and the point cloud data from different sources are preprocessed. The ICP algorithm is used for point cloud registration, aligning the source and target point clouds using rotation and translation, and achieving point cloud registration by minimizing the error function. Root mean square error, average distance error, vertical error, and horizontal error are used as evaluation indicators for the point cloud registration algorithm to quantify the accuracy and error of point cloud registration and to evaluate the registration. Therefore, this invention effectively improves the accuracy of point cloud registration through a segmented registration strategy, especially in the complex environment of natural forests. By fusing BLS and ULS multi-source point cloud data, it overcomes the limitations of single-sensor data and significantly improves the extraction accuracy of tree structural parameters. This method is not only applicable to point cloud registration in natural forests but also provides reliable data support and technical assurance for forest resource monitoring and ecological protection. Attached Figure Description

[0042] Figure 1 This is a flowchart illustrating the multi-source point cloud registration method for natural forests based on point cloud segmentation optimization in an embodiment of the present invention.

[0043] Figure 2 This is a schematic diagram showing the location, species, and density of individual trees within the sample plot in an embodiment of the present invention.

[0044] Figure 3 This is a schematic diagram of the preprocessing flow of BLS and ULS point cloud data in an embodiment of the present invention;

[0045] Figure 4 This is a schematic diagram of the technical process for optimizing the point cloud registration strategy of the ICP algorithm in an embodiment of the present invention;

[0046] Figure 5 This is a schematic diagram of the RMSE values ​​of the registration results based on ICP (point-to-plane) in an embodiment of the present invention;

[0047] Figure 6 This is a schematic diagram of the ICP (point-to-plane) point cloud registration result in an embodiment of the present invention;

[0048] Figure 7 This is a schematic diagram of the RMSE values ​​of the registration results based on ICP (point-to-point) in an embodiment of the present invention;

[0049] Figure 8 This is a schematic diagram of the point cloud registration result based on the ICP (point-to-point) algorithm in an embodiment of the present invention;

[0050] Figure 9 This is a schematic diagram illustrating the number of iterations and processing time for registration of different land parcels in an embodiment of the present invention;

[0051] Figure 10 This is a schematic diagram illustrating different registration types and tree density classification evaluation indicators in embodiments of the present invention;

[0052] Figure 11 This is a schematic diagram of the horizontal cross-section point cloud results under different registration methods in the embodiments of the present invention;

[0053] Figure 12 This is a schematic diagram of the structure of the natural forest multi-source point cloud registration system based on point cloud segmentation optimization in an embodiment of the present invention. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0055] This invention provides a method for multi-source point cloud registration in natural forests based on point cloud segmentation optimization, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the multi-source point cloud registration method for natural forests based on point cloud segmentation optimization, as described in an embodiment of the present invention.

[0056] In this embodiment, a method for registering multi-source point clouds of natural forests based on point cloud segmentation optimization includes the following steps:

[0057] S1: Set up sample plots and collect data on the diameter at breast height, tree height and crown width of individual trees, and obtain the spatial coordinates of individual trees using RTK;

[0058] S2: Point cloud data of the study area is collected by using UAV laser scanning and backpack laser scanning, and point cloud data from different data sources in the study area are preprocessed.

[0059] S3: The ICP algorithm is used for point cloud registration. Rotation and translation are used to align the source point cloud with the target point cloud. Point cloud registration is achieved by minimizing the error function.

[0060] S4: Root mean square error, average distance error, vertical error, and horizontal error are used as evaluation indicators for point cloud registration algorithms to quantify point cloud registration accuracy and error, and to evaluate point cloud registration.

[0061] It should be noted that existing technologies have certain problems in the registration of natural forest point cloud data, especially in the process of multi-source data fusion and registration. Due to the mutual occlusion of tree canopies, BLS has difficulty penetrating the canopy to capture the structure of individual trees, leading to errors in the estimation of tree height and diameter at breast height (DBH). ULS is also affected by canopy interference; the scattered distribution of low-level point clouds easily introduces noise, thus affecting the accuracy of parameter extraction. In point cloud registration, the Iterative Closest Point (ICP) algorithm is widely used. Its basic principle is to continuously adjust the positional relationship between the source and target point clouds through rotation and translation to minimize the distance between corresponding points and achieve high-precision alignment. The ICP algorithm can be divided into two types: point-to-point and point-to-plane. Among them, the point-to-point method directly minimizes the Euclidean distance between the source and target points, which is computationally simple and converges quickly. The point-to-plane method uses the normal vector of the target point to establish the geometric relationship between the source point and the target plane through the covariance matrix and singular value decomposition (SVD), which can improve the stability of rotation estimation to a certain extent. Both methods have their advantages and disadvantages, but when there is a lot of noise or inconsistent data quality, the ICP algorithm may still get stuck in a local optimum, leading to a decrease in registration accuracy.

[0062] To address the aforementioned issues, this embodiment establishes sample plots and collects data on individual tree diameter at breast height (DBH), height, and crown width. It then uses RTK to obtain the spatial coordinates of individual trees and employs UAV laser scanning and backpack laser scanning to collect and preprocess point cloud data for the study area. The ICP algorithm is used to align the source and target point clouds through rotation and translation. Point cloud registration is achieved by minimizing the error function. Evaluation metrics are used to quantify the accuracy and error of point cloud registration, and point cloud registration is evaluated. Therefore, this invention effectively improves the accuracy of point cloud registration through a segmented registration strategy. By fusing multi-source point cloud data from BLS and ULS, it overcomes the limitations of single-sensor data, significantly improving the extraction accuracy of tree structural parameters. This approach is not only applicable to point cloud registration in natural forests but also provides reliable data support and technical assurance for forest resource monitoring and ecological protection.

[0063] To explain this application more clearly, a specific application example of a multi-source point cloud registration method for natural forests based on point cloud segmentation optimization is provided below.

[0064] 1. Overview of the study area:

[0065] This study area is located in Zhenyuan County, Pu'er City, Yunnan Province (100°21′-101°31′E, 23°24′-24°22′N). It has complex terrain, an average altitude of 1376m, and a subtropical monsoon climate with annual precipitation ranging from 1100mm to 2780mm and average annual temperature from 17°C to 21°C. The forest coverage in this area is approximately 76.33%, including natural forests and plantations dominated by *Pinus kesiya*. *Pinus kesiya* var. *langbianensis*, as the dominant species in the region, is adapted to local ecological conditions and has significant ecological and economic value, promoting biodiversity and ecological stability.

[0066] 2. Data source:

[0067] (1) Data acquisition from sample plots:

[0068] Nine representative 30m × 30m plots were selected in Wuyi Village, Zhenyuan County in February 2023. In each plot, all living trees with a diameter at breast height (DBH) greater than 5cm were marked and their parameters were measured. DBH was recorded at a height of 1.3m on each tree using a DBH measuring tape, along with the tree species information. For tree height (H), measurement locations with clear views of the tree base and crown apex were selected, and the heights of the two points were measured and recorded using a calibrated laser rangefinder. Each measurement was repeated three times, and the average value was used as the final result. Tree locations were recorded using an RTK rover station, and differential correction technology was employed for precise positioning. Based on stand density characteristics, nine plots were designed and classified into three types according to density: Type I (700-1100 trees / hectare), Type II (1100-1400 trees / hectare), and Type III (1400-1800 trees / hectare). Each type included three plots, numbered S1 to S9. Figure 2 (a) This shows the relative positions of individual trees in plots S1-S9 (circular center point, size proportional to diameter at breast height), with gray representing Pinus sylvestris and red representing other broad-leaved tree species; (b) This is a statistical chart of the number of trees in plots S1-S9. The slope information of the plots is listed in Table 1.

[0069] Table 1 Slope Information of Sample Plots

[0070] Smaple polt ID Mean Slope Smaple polt ID Mean Slope Smaple polt ID Mean Slope S1 11° S4 30° S7 17° S2 19° S5 14° S8 10° S3 24° S6 32° S9 33°

[0071] (2) LiDAR data acquisition and preprocessing:

[0072] BLS point cloud data was acquired using a LiBackpack device equipped with a DGC50H sensor, while ULS point cloud data was collected using a DJI M300-220N drone equipped with an HS40P sensor. Ground operators used backpack equipment to travel within the sample plot along a "Z"-shaped route at a constant speed of 2 m / s. The drone flew at a terrain-following altitude of 100 m, with a speed of 4 m / s and an 80% overlap in both forward / backward and lateral directions. After removing ground points, the density of the ULS point cloud was approximately 160 points / m³, and the density of the BLS point cloud was approximately 800 points / m³. A 5 m buffer was established to ensure the integrity of the edge tree point cloud data. The collected point cloud data underwent preprocessing; the specific process is detailed in [link to documentation]. Figure 3 First, using attitude angle and acceleration data from the inertial measurement unit (IMU), the point cloud is converted into a global coordinate system with three-dimensional position coordinates, eliminating errors caused by equipment motion and attitude changes. Next, the point cloud is clipped using plot boundary vectors and a 5m buffer. Invalid noise points are removed using the 5±σ rule. Then, the cloth simulation filtering (CSF) algorithm is applied, with a classification threshold of 0.50m, a grid size of 0.10m, and 500 iterations, to separate ground points from vegetation points. The point cloud is normalized based on the ground points, eliminating scale and unit differences caused by slope variations. Finally, only the normalized vegetation point cloud is exported to reduce data volume and prevent convergence problems of local optima during registration.

[0073] 3. Point cloud registration using the ICP algorithm:

[0074] In step 3, the ICP algorithm is used for fine registration. The algorithm aligns the source point cloud with the target point cloud using rotation and translation matrices (Formulas (1) and (2)) and achieves point cloud registration by minimizing the error function (Formula (3)). The ICP algorithm includes two methods: point-to-point and point-to-surface. The two methods are similar in principle, but differ in the calculation methods of rotation and translation vectors: the point-to-point method directly minimizes the Euclidean distance between corresponding point pairs; the point-to-surface method is based on the covariance matrix and singular value decomposition (SVD), and calculates the rotation and translation parameters through the relationship between the source point and the target surface normal vector. In order to improve the registration accuracy, the BLS point cloud is set as the target point cloud and the ULS point cloud is set as the source point cloud. The ICP algorithm is used for fine registration to ensure the integrity and accuracy of the structural parameters of a single tree. The calculation formula is as follows:

[0075] (1)

[0076] (2)

[0077] (3)

[0078] In the formula: It is the optimal rotation matrix. It is the optimal translation matrix. It is a collection of source point clouds. It is a collection of target point clouds. It is a rotation matrix. It is a translation matrix. It represents the number of iterations.

[0079] 4. Segmented registration strategy:

[0080] Under the optimal ICP algorithm, a segmented registration strategy is adopted. BLS accurately captures the trunk structure, while ULS excels at acquiring crown and tree height information. However, directly registering point clouds above the ground may increase noise around the trunk, interfering with parameter extraction. To address this issue, a segmented registration strategy is proposed. Based on the basic characteristics of Pinus kesiya var. langbianensis forests (average tree height between 12m and 18m, with the base of the crown around 8m) and the ICP (point-to-point) method, five registration types were designed: 0-0m, 0-2m, 0-4m, 0-6m, and 0-8m. The xy type indicates that BLS uses point clouds above xm, and ULS uses point clouds above ym. Figure 4 In the optimized ICP algorithm, the iteration condition is set as follows: stop when the change in RMSE is less than 0.001m and the change in MDErr is less than 0.0001m, or after a maximum of 50 iterations. The technical process first performs height filtering on the ULS point cloud, gradually eliminating low-level points by segmentation, with each segment serving as the registration point cloud and the BLS point cloud as the reference point cloud. Next, the ICP process includes initializing the point cloud matrix, using KDtree to find the nearest neighbor, calculating the Euclidean distance, solving for the rigid transformation parameters, calculating the rotation and translation matrices through singular value decomposition (SVD), decentering, applying the rigid transformation, and iterating. Figure 4 ).

[0081] 5. Point cloud registration evaluation:

[0082] Root mean square error (RMSE), mean distance error (MDErr), vertical error (VDErr), and horizontal error (HErr) are used as evaluation metrics for point cloud registration algorithms to quantify the accuracy and error of point cloud registration. The calculation formulas are as follows:

[0083] RMSE= (4)

[0084] (5)

[0085] (6)

[0086] (7)

[0087] In the formula: The distance between each pair of corresponding points. It is the number of corresponding point pairs. These are the coordinates of the measurement point. These are the reference coordinates.

[0088] 6. Results and Analysis:

[0089] (1) ICP (point-to-plane) point cloud registration results:

[0090] Registration results for all sample plots were obtained through multiple iterations under the same constraints. The average RMSE for type I tree density was 0.37m (S1: 0.31m, S2: 0.37m, S3: 0.44m); the average RMSE for type II was 0.29m (S4: 0.16m, S5: 0.38m, S6: 0.33m); and the average RMSE for type III was 0.36m (S7: 0.29m, S8: 0.44m, S9: 0.37m). Figure 5 As shown, the RMSE values ​​for all tree density types are below 0.50m, with S4 achieving the highest accuracy and S8 the lowest.

[0091] Figure 6 (a) shows the overlay effect after BLS and ULS point cloud registration. This combination effectively compensates for the lack of canopy structure in the BLS data, improves the overall density, and enhances the integrity of the canopy. The geometry of the point cloud remains consistent before and after registration, but negative values ​​appear near the ground, indicating good local registration but uneven global alignment. Figure 6 (b) shows the S8 registration result with the highest RMSE. The density of the BLS point cloud is higher than that of the ULS point cloud, and different datasets show different levels of detail. The rotated and translated ULS point cloud matches the BLS point cloud relatively well, but the initial registration and global alignment are not ideal. Figure 6 (c) shows that S4 has good initial alignment and a small RMSE. Figure 6 (d) shows the registration cross-sectional views of S4 and S8. In S8, the ULS point cloud portion is lower than the ground, while this problem does not exist in S4. This phenomenon stems from the rigid transformation in ULS point cloud registration, which prioritizes optimizing the alignment accuracy of the canopy region, but may cause the point cloud to shift downwards, resulting in negative values ​​and affecting the overall registration accuracy. S8 had poorer initial alignment conditions, leading to a more severe problem of negative values. Figure 6 (e) highlights the registration of individual trees, where the reference point cloud mainly relies on the upper canopy points, while the ULS point cloud near the trunk and ground becomes a source of interference, emphasizing the importance of visual verification in the registration quality assessment.

[0092] (2) ICP (point-to-point) point cloud registration results:

[0093] Variations in RMSE error index at different locations, such as Figure 7 As shown, all RMSE values ​​were below 0.50m. The average RMSE for type I tree density was 0.36m (S1: 0.31m, S2: 0.34m, S3: 0.44m); for type II it was 0.29m (S4: 0.16m, S5: 0.39m, S6: 0.33m); and for type III it was 0.34m (S7: 0.29m, S8: 0.44m, S9: 0.29m). Numerically, the ICP (point-to-point) and ICP (point-to-matrix) registration results were quite similar, with ICP (point-to-point) performing slightly better. Notably, in plot S9, the RMSE decreased to 0.29m, an improvement of 0.08m.

[0094] Point cloud registration results show that, Figure 8 In (a), the point clouds of all tree density types are generally well aligned. Figure 8 (b) shows the registration effect of plot S4, where the point cloud remains highly consistent in local areas. The BLS and ULS data have very low noise and no obvious anomalies after registration, and the overall error is well controlled. Figure 8 (c) shows that the BLS canopy point cloud was effectively supplemented, and although noise still exists near the ground, it represents a slight improvement compared to the ICP (point-to-point) method. Furthermore, Figure 8 (d) illustrates the changes in individual trees before and after registration. Due to the complex geometry of individual trees in the point cloud, morphological changes at different angles, occlusion effects, and lighting conditions, the ICP algorithm may struggle to accurately match corresponding points, leading to increased errors. Compared to the registered point cloud and the ULS point cloud, the BLS point cloud, captured through a single data source, exhibits better representation of individual tree trunk structures. This is because, after rigid transformation, low-lying vegetation points and misalignment points introduced during registration become noise, interfering with the extraction of individual tree parameters and negatively impacting registration quality.

[0095] (3) ICP point cloud registration optimization results:

[0096] A segmented registration strategy is adopted under the optimal ICP algorithm (point-to-point). To comprehensively analyze the changes in RMSE, this study introduces three additional error metrics: MDErr, VErr, and HErr. Figure 9This paper presents the iteration count and time efficiency results of various segmented registration methods under different tree densities. The results show that increasing the segment height does not necessarily reduce registration time. Registration time is affected not only by segment height but also by the structural complexity of the point cloud, the matching speed of corresponding points, and the computational cost during the segmentation process. Furthermore, tree density is not the only factor affecting registration time. Although high-density plots generally require more computation time, other factors such as plot slope and point cloud complexity also significantly influence the registration process. Therefore, the relationship between tree density and registration time is not a simple positive correlation. In conclusion, point matching speed is a key factor in registration, but in some high-density or structurally complex point clouds, the matching speed may decrease, leading to a longer registration time.

[0097] Figure 10 The changes in registration error indices under different segmentation types are shown. The results indicate that HErr is relatively stable, while RMSE, MDErr, and VDErr show similar trends, suggesting that vertical transformation plays a dominant role in optimizing registration accuracy. In particular, height alignment from 0-0m to 0-2m resulted in the largest error reduction. Segmented registration effectively reduced RMSE, with the reduction rates being: Type I (0.10m-0.15m), Type II (0.05m), and Type III (0.01m-0.05m). Although the error changes in plots S4, S7, and S9 were relatively small, improvements were observed under the 0-2m type, requiring further analysis in conjunction with visual inspection results.

[0098] The fusion effect of point cloud registration at different heights shows that the canopy structure consistently benefits from the complementarity of point cloud information. For example... Figure 11 As shown, (a) presents the side view before registration and the top view after fusion in plot S8; (b) presents the ULS point clouds of individual trees at different horizontal heights and the fused profile; (c) presents the effect diagrams of individual trees before and after fusion. By adjusting the registration type, interference from low-level point cloud data is reduced, and the alignment accuracy of the canopy region is enhanced, especially in areas with low tree density (such as types I and II), where the improvement of the segmentation strategy is more obvious. The segmented ICP method effectively reduces registration errors and achieves more complete canopy structures in plots S1, S2, S3, S5, S6, and S8, successfully avoiding interference from irrelevant ULS points. Although the improvement is smaller in some plots (such as S4, S7, and S9), the point cloud registration accuracy is further improved by removing low-level points and optimizing the alignment of the canopy region. This study verifies the effectiveness of the segmented fusion strategy in point cloud registration, provides support for tree parameter extraction, and ensures the integrity of the canopy structure.

[0099] In summary, this invention discloses a multi-source point cloud registration strategy for natural forests based on point cloud segmentation optimization. By comparing the point-to-point and point-to-area methods of the Iterative Closest Point (ICP) algorithm, the results show that both can achieve high-precision point cloud registration. However, the point-to-point method is more stable in the complex environment of natural forests and less affected by rigid transformation offsets, thus it is preferred as the final method. Based on this, the segmented registration strategy effectively improves registration accuracy, especially in the alignment optimization of the canopy area, significantly reducing noise interference around the ground and tree trunks. This strategy can effectively fuse BLS and ULS point cloud data, improving the integrity and accuracy of forest structure. Simultaneously, by optimizing the ICP algorithm, unnecessary low-level point cloud data is reduced, further improving registration accuracy and providing reliable data support for extracting forest structure parameters in complex natural forest environments.

[0100] Reference Figure 12 , Figure 12 This is a schematic diagram of the structure of the multi-source point cloud registration system for natural forests based on point cloud segmentation optimization, according to an embodiment of the present invention.

[0101] like Figure 12 As shown, the multi-source point cloud registration system for natural forests based on point cloud segmentation optimization proposed in this embodiment of the invention includes:

[0102] The data acquisition module 10 is used to set up sample plots and collect data on the diameter at breast height (DBH), tree height, and crown width of individual trees, and to obtain the spatial coordinates of individual trees using RTK.

[0103] Processing module 20 is used to collect point cloud data of the study area using UAV laser scanning and backpack laser scanning, and to preprocess point cloud data from different data sources in the study area.

[0104] The registration module 30 is used to perform point cloud registration using the ICP algorithm. It aligns the source point cloud with the target point cloud by rotation and translation, and achieves point cloud registration by minimizing the error function.

[0105] Evaluation module 40 is used to evaluate point cloud registration algorithms by employing root mean square error, average distance error, vertical error, and horizontal error as evaluation metrics to quantify point cloud registration accuracy and error, and to evaluate point cloud registration.

[0106] Other embodiments or specific implementations of the natural forest multi-source point cloud registration system based on point cloud segmentation optimization of the present invention can be referred to the above-described method embodiments, and will not be repeated here.

[0107] It is understood that in the description of this specification, references to terms such as "one embodiment," "another embodiment," "other embodiments," or "first embodiment to Nth embodiment," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0108] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0109] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A natural forest multi-source point cloud registration method based on point cloud segmentation optimization, characterized in that, The method comprises the following steps: S1: layout sample plot and collect single-tree diameter at breast height, tree height and crown width, and obtain single-tree spatial coordinates in combination with RTK; S2: point cloud data collection in the research area is performed by using unmanned aerial vehicle laser scanning and backpack laser scanning, and the point cloud data of different data sources in the research area is preprocessed; S3: point cloud registration is performed by using an ICP algorithm, source point cloud and target point cloud are aligned by using rotation and translation, and point cloud registration is realized by minimizing an error function; specifically, the following steps are included: S31: ULS point cloud data as source point cloud and BLS point cloud data as target point cloud are aligned by using a rotation matrix and a translation matrix by using the ICP algorithm, and point cloud registration is realized by minimizing an error function; wherein, an expression of the point cloud registration is specifically as follows: 、 、 、 wherein, is the optimal rotation matrix, is the optimal translation matrix, is the source point cloud set, is the target point cloud set, is the rotation matrix, is the translation matrix, is the number of iterations; Wherein, the ICP algorithm adopts a point-to-point ICP algorithm, and the point-to-point ICP algorithm calculates rotation and translation parameters by directly minimizing the Euclidean distance between corresponding point pairs; Step S31, specifically includes: S311: a segmented registration strategy is adopted under the ICP algorithm, the trunk is captured by the backpack laser scanning, and the crown and the height information are obtained by the unmanned aerial vehicle laser scanning; S312: height filtering is performed on the ULS point cloud data, the segmented processing is performed by the preset segmented registration strategy, and the low-layer points are gradually removed, each segment is used as a registration point cloud, the BLS point cloud is used as a reference point cloud, the rotation matrix and the translation matrix are iteratively adjusted, the corresponding point pairs are found, and the point cloud registration is realized; S4: root mean square error, average distance error, vertical error and horizontal error are used as evaluation indexes of the point cloud registration algorithm, the point cloud registration accuracy and error are quantified, and the point cloud registration evaluation is performed.

2. The natural forest multi-source point cloud registration method based on point cloud segmentation optimization according to claim 1, wherein, Step S1: layout sample plot and collect single-tree diameter at breast height, tree height and crown width, and obtain single-tree spatial coordinates in combination with RTK, specifically including: S11: nine sample plots of three density types are laid out; wherein, the density types include: type I with a density of 700-1100 trees per hectare, type II with a density of 1100-1400 trees per hectare, and type III with a density of 1400-1800 trees per hectare; S12: single-tree diameter at breast height, tree height and crown width of target tree species in each sample plot are collected, the positions of the target tree species are recorded by using an RTK mobile station, and differential correction is adopted for precision positioning.

3. The natural forest multi-source point cloud registration method based on point cloud segmentation optimization according to claim 2, wherein, Step S2: point cloud data collection in the research area is performed by using unmanned aerial vehicle laser scanning and backpack laser scanning, and the point cloud data of different data sources in the research area is preprocessed, specifically including: S21: ULS point cloud data of the research area is obtained by using an unmanned aerial vehicle laser scanning device, and BLS point cloud data of the research area is obtained by using a backpack laser scanning device; S22: for the ULS point cloud data and the BLS point cloud data collected in the research area, preprocessing including data solution, cropping, denoising, normalization and ground point separation is performed.

4. The natural forest multi-source point cloud registration method based on point cloud segmentation optimization according to claim 1, wherein, The ICP algorithm adopts a point-to-surface ICP algorithm, the point-to-surface ICP algorithm calculates rotation and translation parameters based on a covariance matrix and singular value decomposition by using source points and target surface normal vectors.

5. The natural forest multi-source point cloud registration method based on point cloud segmentation optimization according to claim 1, wherein, In step S312, the iteration condition is that the RMSE change value is less than 0.001 m and the MDErr change value is less than 0.0001 m, or a maximum of 50 iterations is performed.

6. The natural forest multi-source point cloud registration method based on point cloud segmentation optimization according to claim 5, wherein, Step S4: The root mean square error, the mean distance error, the vertical error and the horizontal error are used as the evaluation indexes of the point cloud registration algorithm, and the expressions are as follows: 、 、 、 、 where RMSE denotes the root mean square error, MDErr denotes the mean distance error, VErr denotes the vertical error, HErr denotes the horizontal error, is the distance between each pair of corresponding points, is the number of pairs of corresponding points, is the coordinate of the measured point, is the coordinate of the reference.

7. A natural forest multi-source point cloud registration system based on point cloud segmentation optimization, configured to perform the natural forest multi-source point cloud registration method based on point cloud segmentation optimization according to any one of claims 1-6. It comprises: The acquisition module is used for laying out sample plots and collecting single-tree diameter at breast height, tree height and crown width, and obtaining the spatial coordinates of single trees in combination with RTK; The processing module is used for collecting point cloud data of the research area by using unmanned aerial vehicle laser scanning and backpack laser scanning, and pre-processing the point cloud data of different data sources of the research area; The registration module is used for performing point cloud registration by using the ICP algorithm, aligning the source point cloud and the target point cloud by using rotation and translation, and realizing point cloud registration by minimizing the error function; The evaluation module is used for using the root mean square error, the mean distance error, the vertical error and the horizontal error as the evaluation indexes of the point cloud registration algorithm, quantifying the point cloud registration accuracy and error, and evaluating the point cloud registration.

Images