Model and method for forest scene airborne and ground platform laser point cloud registration

Abstract

The invention relates to the technical field of three-dimensional point cloud data processing, and discloses a model and method for forest scene airborne and ground platform laser point cloud registration, and the method comprises the following steps: carrying out the preprocessing of point cloud data; model construction, wherein tree position key point extraction, vertical layering virtual point construction and tree vertical layering aggregation descriptor construction are carried out in sequence; during vertical layering virtual point construction, the length of a predefined vertical axis is adopted, it is ensured that the vertical axes of all trees are evenly divided into N layers within the same height range, and the center point of each layer serves as a virtual construction point; feature matching: fault-tolerant matching selection, geometric consistency screening and random sampling consistency are carried out in sequence; and point cloud registration: calculating a rotation matrix based on the matching pair and carrying out ICP precise registration in sequence. The method does not depend on complete and accurate extraction of tree position key points, the constructed combination features are high in description capability, and robustness to regularly distributed woods is high.

Description

Technical Field

[0001] The present invention relates to the technical field of three-dimensional point cloud data processing, and in particular to a model and method for airborne and ground platform laser point cloud registration of forest scenes. Background Art

[0002] With the advancement of modern forestry refined management and ecological research, LiDAR (Light Detection and Ranging) technology has become an indispensable key tool for forest inventory with its significant advantages of automation, fast and efficient, and high-precision measurement. Among them, the UAV laser scanning system and the ground laser scanning system (including ground-based laser scanners and mobile laser scanners) work together to accurately capture the detailed information such as the branch shape and spatial distribution of trees in the forest, laying a solid data foundation for core applications such as tree parameter extraction, biomass estimation, vegetation type classification, and pest and disease monitoring. However, the complex three-dimensional spatial structure of the forest poses a huge challenge to point cloud data processing. Due to differences in observation angles, the point cloud data collected by drones and ground-based LiDAR systems show significant heterogeneity in density, coverage, and spatial distribution. In addition, the mutual occlusion of tree canopies leads to missing or redundant data in some areas, making it difficult for a single-platform LiDAR point cloud to fully and accurately characterize the complete structural information of the forest. Only by deeply integrating drone and ground-based LiDAR point cloud data can a comprehensive and high-precision modeling of forest ecosystems be achieved. In this process, point cloud data registration, as a key prerequisite for fusion, directly determines the accuracy and reliability of subsequent analysis. At present, relevant research has been conducted on point cloud registration. For example, patent CN113205548A, an automated registration method and system for forest area drones and ground-based point clouds, and patent CN118799363A, a drone and ground-based forest point cloud registration method, both provide their own registration methods, which can achieve the purpose of registration to a certain extent. However, both methods have high requirements on the quality of the collected initial point cloud data, and are heavily dependent on the completeness and accuracy of the extraction of key points of tree positions, which are difficult to achieve in many practical application scenarios.

[0003] Therefore, point cloud registration technology in forest scenes still has many bottlenecks that need to be solved: First, the registration method based on geographic reference is highly dependent on accurate geographic coordinate information to align point cloud data from different platforms. However, in forest environments, especially in dense forests with high canopy density, the severe obstruction of satellite signals by the canopy will lead to a significant decrease in absolute positioning accuracy, which greatly limits the scope of application and effectiveness of this method. Second, due to the inherently weak structural characteristics of forest scenes, the local feature distributions of different key points show a high degree of similarity. It is difficult for a single feature descriptor to accurately depict the unique attributes of each key point, resulting in a low accuracy of feature matching and seriously hindering the smooth progress of point cloud registration. Third, affected by the interference of tree crown occlusion, local descriptors constructed based on the positions and attributes of trees (such as diameter at breast height and tree height) are difficult to achieve accurate matching in environments with relatively fixed layouts such as plantations due to the extremely high similarity of local features, directly affecting the overall accuracy of point cloud registration. Therefore, developing an airborne and ground platform laser point cloud registration model and method suitable for forest scenes and breaking through the bottlenecks of existing technologies have important practical significance and application value for improving the level of forest resource monitoring and ecological research. Summary of the Invention

[0004] The present invention aims to provide a model and method for airborne and ground platform laser point cloud registration in forest scenes to solve the problems in the existing feature extraction method based on tree positions for UAV and ground LiDAR point cloud registration in forest scenes, which depends on the complete and accurate extraction of tree position key points, and the constructed simple feature description ability is not strong, and the robustness to regularly distributed forests is low.

[0005] To solve the above problems, the present invention adopts the following technical solutions: Solution 1: A method for airborne and ground platform laser point cloud registration in forest scenes, including the following steps: Step 1, preprocess the point cloud data; Step 2, model construction, successively perform tree position key point extraction, vertical layered virtual point construction, and tree vertical layered aggregation descriptor construction; when constructing the vertical layered virtual points, a predefined vertical axis length is used to ensure that the vertical axes of all trees are evenly divided into N layers within the same height range, and the center point of each layer is used as a virtual construction point; Step 3, feature matching, successively perform fault-tolerant matching selection, geometric consistency screening, and random sample consensus; Step 4, point cloud registration, successively calculate the rotation matrix based on the matching pairs and perform ICP precise registration.

[0006] Preferably, in step two, the construction of the virtual points is specifically as follows: standardize the downsampled point cloud, use the slicing method to extract the lower trunk area of the tree, locate the trunk position to extract the horizontal coordinates (X, Y), construct two-dimensional key points, use the average elevation of all ground points in the buffer to represent its Z coordinate, and construct three-dimensional key points; for each two-dimensional key point, construct a vertical axis through this point and evenly divide it into N layers. The center point of each layer is the virtual point constructed for the corresponding layer of the key point. Introduce a unified axis height parameter to ensure that the vertical axis lengths of all trees are the same. The j-th layer virtual point of the i-th key point The calculation formula is: , is the unified axis height parameter, is the number of layers evenly divided.

[0007] Preferably, the construction of the tree vertical layer aggregation descriptor in step two includes virtual point feature extraction and feature order splicing; the virtual point feature extraction is specifically as follows: based on the virtual points of each layer constructed by the key points, use the Fast Point Feature Histogram (FPFH) to extract the local geometric features of the virtual points of each layer in the standardized point cloud. If a certain layer only contains virtual points and lacks effective point cloud data, it is considered that the feature extraction of this layer fails, mark it as an invalid layer and skip the calculation of this layer, and perform L1 norm standardization processing on the successfully extracted features. Preferably, the feature order splicing is specifically as follows: splice the standardized features of each layer of the same axis calculated successfully in the order from bottom to top or from top to bottom to generate the tree vertical layer aggregation descriptor.

[0008] Preferably, the fault-tolerant matching selection in step three is specifically as follows: for the key point sets of the given source point cloud and target point cloud, calculate the Manhattan distance between the features of the associated tree vertical layer aggregation descriptors, and select the first n most similar points of the source point cloud key points and target point cloud key points to construct a matching pair candidate set, where n is set to 3.

[0009] Preferably, the geometric consistency screening in step three is specifically as follows: introduce a geometric consistency analysis method, obtain the geometric consistency score GCS by calculating the geometric consistency distance between the matching pairs; construct a GCS matrix GCSM for voting and screening according to the score, identify the maximum score item in the matrix, set all elements in its row and column to zero, and repeat the operation until all unmarked matrix elements become zero to optimize the matching pairs. Preferably, the calculation of the rotation matrix in step four is specifically as follows: use the corresponding set of key points extracted, calculate the rigid transformation parameters for converting the target point cloud to the source point cloud coordinate system according to the matching pairs, set the scale factor of the rotation matrix to 1, and calculate the rotation matrix and translation vector.

[0010] Solution 2: A model for airborne and ground platform laser point cloud registration in forest scenarios, including a hierarchical vertical aggregation feature model: A vertical hierarchical virtual point construction module for normalizing the downsampled point cloud, extracting key points in the lower trunk area of trees, constructing a vertical axis and generating multiple layers of virtual points, and introducing a unified axis height parameter to ensure the same length of the vertical axis; A tree vertical hierarchical aggregation descriptor construction module for extracting local geometric features based on virtual points, splicing the successfully extracted coaxial normalized features in sequence to form a tree vertical hierarchical aggregation descriptor.

[0011] Preferably, it further includes a fault-tolerant multi-level matching pair screening model: An initial matching set construction module for constructing an initial matching relationship set of key points of the source point cloud and the target point cloud using the Manhattan distance; A geometric consistency screening module for voting and screening by calculating the geometric consistency score GCS and constructing a GCS matrix GCSM to optimize the matching pairs; An abnormal matching pair elimination module for further eliminating abnormal matching pairs using the RANSAC algorithm to obtain high-quality matching pairs. The principle of the present invention: The present invention realizes the registration of airborne and ground platform laser point clouds in forest scenarios through a complete set of method processes and model systems. At the method level, first, the collected point cloud data is preprocessed. Through point cloud downsampling, ground separation, and point cloud normalization, concise and representative data is provided for subsequent processing. Then, it enters the model construction stage. First, key points are extracted, and then virtual points are constructed. By extracting key points of the tree trunk, constructing a vertical axis and generating vertical hierarchical virtual points in layers, the unified axis height ensures correspondence. Then, features are extracted and spliced based on the virtual points to form hierarchical vertical aggregation features. Subsequently, feature matching is performed. An initial matching set is constructed using the Manhattan distance, the matching pairs are screened through geometric consistency analysis, and then abnormal matching pairs are eliminated using the RANSAC algorithm to determine the accurate matching relationship. Finally, the rotation matrix is calculated based on the matching pairs, and the ICP algorithm is combined to achieve accurate point cloud registration. In terms of the model, the hierarchical vertical aggregation feature model starts from the vertical shape of the tree. The vertical hierarchical virtual point construction module obtains virtual points, and the tree vertical hierarchical aggregation descriptor construction module further forms unique feature descriptions. The robust fault-tolerant multi-level matching pair screening model, through the initial matching set construction module, geometric consistency screening module, and abnormal matching pair elimination module, screens out high-quality matching pairs layer by layer, providing a reliable basis for point cloud registration. The two complement each other, are closely combined with the method steps, and jointly complete the point cloud registration task. Compared with the prior art, the present invention does not rely on the complete and accurate extraction of key points of tree positions, constructs a combined feature with strong description ability, and has high robustness to regularly distributed forests.

[0012] Advantages of the present invention: High precision: The hierarchical vertical aggregation feature model fully explores the morphological features of trees in the vertical direction, fuses features at different heights, makes the feature vector highly unique, reduces mis-matching caused by similar trees, and at the same time uses features at different heights to make up for the impact of incomplete data; The robust fault-tolerant multi-level matching pair screening model uses a progressive strategy, gradually applies more stringent constraint conditions and screening rules through relaxed matching, tightens the matching criteria, effectively eliminates mis-matches, and obtains an accurate matching relationship, thereby achieving high-precision point cloud registration. Strong robustness: In the process of virtual point construction, aiming at the problems of missing lower parts of UAV point clouds and difficult to obtain accurate tree heights by ground LiDAR, the method adopted can effectively filter low-quality data, overcome the influence of tree height differences by unifying the axis height, and enhance the correspondence between the same-name virtual points in the airborne and ground platform point clouds by using the unified virtual axis height, thereby improving the consistency of the same-name features between different sources; In the feature matching stage, the fault-tolerant multi-level matching pair screening strategy can adapt to the challenges brought by partial overlap, density difference and repeated structural features in the forest scene, enabling the model to still maintain good performance in complex environments. Wide adaptability: The method and model comprehensively consider various complex factors in the forest scene. Whether it is the perspective difference and data heterogeneity between UAV and ground LiDAR point clouds, or problems such as canopy occlusion, there are corresponding processing mechanisms, and it is applicable to different types of forest environments, including regularly distributed forests and natural forests with complex structures. Outstanding innovation: Breaking through the limitations of the existing tree position feature extraction, a hierarchical vertical aggregation feature model and a robust fault-tolerant multi-level matching pair screening framework are proposed, solving the point cloud registration problem in the forest scene from a new perspective, and providing a brand-new technical idea and solution for this field. Brief Description of the Drawings

[0013] Figure 1 It is a flowchart of an embodiment of the present invention.

[0014] Figure 2 It is a comparison chart of the registration results of five survey areas in an embodiment of the present invention.

[0015] Figure 3 It is the virtual points constructed at the same position of the airborne and ground platform point clouds in an embodiment of the present invention, where blue: airborne platform point cloud; green: ground platform point cloud; red: two-dimensional key points; purple: vertically stratified virtual points. Detailed Embodiment

[0016] The following is a further detailed description through specific embodiments: The embodiment is basically as shown in the appendix Figure 1 as follows: The method for airborne and ground platform LiDAR point cloud registration in forest scenes includes the following: In the first step, preprocess the point cloud data, successively perform point cloud downsampling, ground separation, and point cloud normalization. Use Gaussian filtering to remove noise points, use the Cloth Simulation Filter (CSF) algorithm to separate ground and non-ground point clouds, and normalize the non-ground point clouds.

[0017] In the second step, construct the model, successively perform extraction of key points for tree positions, construction of vertically stratified virtual points, and construction of tree vertical stratification aggregation descriptors; among them, the construction of vertically stratified virtual points includes extraction of virtual point features and concatenation of feature orders.

[0018] In the third step, perform feature matching, successively perform selection of fault-tolerant matching, geometric consistency screening, and random sample consensus.

[0019] In the fourth step, perform point cloud registration, successively calculate the rotation matrix based on the matching pairs and perform ICP fine registration.

[0020] The present invention aims at the limitations of the existing feature extraction methods based on tree positions in the registration of UAV and ground LiDAR point clouds in forest scenes, which rely on the complete and accurate extraction of key points, and the constructed simple feature description ability is not strong, and the robustness to regularly distributed forests is low. By deeply exploring the structural similarity of forest scenes, an innovative tree point matching model for LiDAR point cloud registration between airborne and ground platforms in weak-structure forest scenes is proposed. The model mainly includes an innovatively proposed tree vertical stratification aggregation descriptor for describing the vertical structure form of trees and an innovatively proposed robust fault-tolerant multi-level matching pair screening framework.

[0021] The model for LiDAR point cloud registration between airborne and ground platforms in forest scenes of the present invention includes a hierarchical vertical aggregation feature model and a robust fault-tolerant multi-level matching pair screening model.

[0022] Hierarchical vertical aggregation feature model: For the description of the vertical form of trees, it includes two parts: extraction of vertically stratified construction points and construction of tree vertical aggregation descriptors. The former constructs a vertical axis through key points and generates virtual points to ensure correspondence by unifying the axis height parameters; the latter extracts local geometric features from the virtual points, and after standardization, they are concatenated to form hierarchical vertical aggregation features. This model fuses features at different heights, improves the uniqueness of the feature vector, reduces false matches, and can also use features at different heights to make up for the influence of invalid features in some layers caused by incomplete data. Robust Fault-Tolerant Multi-Level Matching Pair Screening Model: In view of the problem of easy generation of false matches in forest scene point cloud registration, this model is designed. Adopting a progressive strategy, using relaxed matching to gradually impose more stringent constraint conditions and screening rules, tightening the matching criteria, obtaining an accurate matching relationship. First, construct an initial matching set through the Manhattan distance, then calculate the GCS through geometric consistency analysis, construct the GCSM for voting and screening, and finally use the RANSAC algorithm to eliminate abnormal matching pairs, obtain high-quality matching pairs, and then estimate the transformation matrix to achieve preliminary alignment of the point cloud.

[0023] Technical Highlights of the Invention: (1)Virtual Point Construction: Traditional methods have problems such as poor feature similarity due to the heterogeneity of UAV and ground LiDAR point clouds, the difference between homologous key points and the corresponding difference of homologous key points. Introduce a method of constructing a standardized height axis and virtual points along the axis, that is, adopt the method of predefined vertical axis length to ensure that the vertical axes of all trees are evenly divided into N layers within the same height range, and the center point of each layer is used as a virtual construction point. This method effectively overcomes the influence of LiDAR data occlusion or uneven quality, and significantly improves the matching robustness of homologous points between UAV and ground platform point clouds, especially showing excellent virtual point correspondence ability in complex forest scenes.

[0024] (2)Construction of Tree Vertical Layer Aggregation Descriptor: Construct a hierarchical vertical aggregation feature that aggregates features of multiple height layers. By hierarchically extracting the geometric features of the neighborhood of the construction points and standardizing them, a feature vector that is spliced layer by layer is formed. This framework improves the matching accuracy through the following advantages: Feature Complementarity: Integrate features of different height layers, enhance the uniqueness of the feature vector, and reduce the probability of false matches caused by similar tree shapes; Feature Uniqueness: Fuse features of different heights, making the feature vector have higher uniqueness. Reduce false matches caused by high tree similarity of a single descriptor and improve the accuracy of matching.

[0025] Feature Robustness: For the local feature extraction errors of single descriptors in different layers caused by point cloud heterogeneity, matching can still be carried out with the help of existing features of different heights, effectively compensating for the influence of incomplete data or point cloud heterogeneity.

[0026] (3)Fault-Tolerant Multi-Level Matching: Design a three-level matching process including Manhattan distance preliminary screening, geometric consistency voting matrix, and RANSAC optimization: This strategy can significantly reduce the number of false matches, improve the matching rate and the robustness of the registration task, and introduce a fault-tolerant mechanism, so that more correct matching pairs are included in the screening set, significantly improving the matching rate. This method can be applied to the point cloud registration tasks of partially overlapping and regularly distributed forest scenes.

[0027] The specific implementation process is as follows: The first step is the preprocessing of point cloud data Process the original point cloud data in the order of existing point cloud downsampling, ground separation, and point cloud normalization. Point cloud downsampling reduces the data density and subsequent computational volume; ground separation distinguishes between ground and non-ground points; point cloud normalization unifies the data scale; key point extraction selects representative points for subsequent operations. The second step is model construction Extraction of key points for tree positions and construction of vertically stratified virtual points: First, standardize the downsampled point cloud. Use the slicing method to accurately extract the lower trunk area of the tree, locate the trunk position to obtain the horizontal coordinates (X, Y), construct two-dimensional key points, and determine the Z coordinate with the average elevation of the ground points in the buffer zone to construct three-dimensional key points. Although the UAV LiDAR point cloud may cause missing lower point clouds and missed key point extraction due to occlusion, low-quality data can be filtered. Subsequently, for each key point Construct a vertical axis through this point and divide it evenly into N layers. The center point of each layer is the virtual point V. Considering that it is difficult to accurately obtain the tree height due to occlusion or data quality issues of ground LiDAR, introduce a unified axis height parameter to ensure that the vertical axis lengths of all trees are the same. The virtual point at the jth layer of the ith key point The calculation formula is: (Formula 1). is the unified axis height parameter, is the number of layers evenly divided.

[0028] Extraction of virtual point features: Based on the virtual points, use the Fast Point Feature Histogram (FPFH) to extract the local geometric features of each layer of virtual points. If a certain layer contains only virtual points and no valid point cloud data, mark it as an invalid layer and skip the calculation of this layer. Standardize the successfully extracted features. See the vertically stratified virtual points in Figure 3 .

[0029] Feature order splicing: Splice the successfully calculated features of each layer in the order from bottom to top or from top to bottom to form a vertical hierarchical aggregation descriptor for the tree. Splice the successfully calculated K layers of standardized valid features on the same axis in the order from bottom to top or from top to bottom to form a vertical hierarchical aggregation descriptor V for the tree.

[0030] (Formula 2).

[0031] The third step is feature matching Fault-tolerant matching selection: For the key point set of the source point cloud and the key point set of the target point cloud , Calculate the vertical hierarchical aggregation descriptor features of the trees

[0032] And The Manhattan distance between them, select the first n (n = 3 in this experiment) most similar points to construct a candidate set of matching pairs, and the formula is: (Formula 3).

[0033] Indicates And The Manhattan distance between them, Indicates selecting the first n most similar key points, and n is set to 3 in this experiment.

[0034] Geometric consistency screening: Introduce geometric consistency analysis, and obtain the geometric consistency score GCS by calculating the geometric consistency distance between matching pairs.

[0035] Suppose there are And , two matching pairs, and their geometric distance calculation formula is: (Formula 4); Cumulative geometric consistency score The calculation formula is: (Formula 5).

[0036] Construct a GCS matrix (GCSM) according to the score for voting and screening. Identify the item with the maximum score in the matrix, set the elements in its row and column to zero, and repeat this operation until all unmarked elements are zero to optimize the matching pairs. Is an indicator function. If , then return 1, otherwise return 0. Is the geometric consistency threshold. This score reflects the degree of spatial distribution consistency between matching pairs. The higher the score, the more credible the matching pair.

[0037] Next, construct a GCS matrix (GCSM) for voting selection.

[0038] Each row of the matrix represents a key point in the source point cloud, and each column corresponds to a key point in the target point cloud. For each pair of matched key points, if its geometric consistency score is not zero, its score is filled into the corresponding position of the matrix; if the matching pair is not selected or the geometric consistency score is zero, the corresponding position of the matrix is ​​filled with zero. Furthermore, by identifying and marking the maximum score item in the matrix, and setting all elements of the row and column where the score item is located to zero, it is ensured that the matching pair is not selected repeatedly. Then, among the remaining unmarked matrix elements, continue to identify the new maximum geometric consistency score item and set the elements of the corresponding rows and columns to zero. This process is repeated until all unmarked matrix elements become zero.

[0039] Through the above screening strategy, higher quality matching pairs can be effectively screened. Finally, a high-quality matching set C is formed. G .

[0040] Random sampling consistency: Use the RANSAC algorithm to further eliminate abnormal matching pairs from the preliminary screening to obtain high-quality matching pairs C F . First, the initial correspondence generation is performed, and the algorithm traverses the source point cloud K S For each key point in the target point cloud K, for each source key point, calculate its T The similarity between all key point descriptors in , select the Top-N similar target key points from these similarity results, and add these correspondences to the initial correspondence set C I middle.

[0041] Then enter the GCSM screening stage, first build a size (n S ×n T ) of the zero matrix GCSM, where n S and n T Represents the number of key points of the source point cloud and the target point cloud respectively. I Each corresponding relationship c i,j , calculate its geometric consistency score G i,j And store the score in the corresponding position of the matrix GCSM. After that, process the matrix row by row, find the maximum element in the matrix each time, and add the corresponding key point pair to the candidate corresponding set C. G In the example, the row and column of the element are set to zero at the same time to ensure that each source key point and target key point can only form a one-to-one correspondence at most.

[0042] Finally, for the candidate corresponding set C GUsing the RANSAC algorithm, corresponding point pairs with geometric consistency are screened based on the inlier threshold θ, and these screened inlier pairs form the final inlier correspondence set C F and returned. The entire matching algorithm adopts a progressive strategy, using relaxation matching to gradually impose more stringent constraint conditions and screening rules, tightening the matching criteria, and obtaining an accurate matching relationship.

[0043] Step 4, Point cloud registration Calculating the rotation matrix based on the matching pairs: According to the extracted key point correspondence set, the optimal rigid transformation matrix parameters for converting the target point cloud to the source point cloud coordinate system are calculated using the matching pairs. Assuming the scale factor of the rotation matrix is 1, the rough registration process is expressed as: (Equation 6).

[0044] where X, Y, Z are the coordinates of the source point cloud, , , are the coordinates of the target point cloud, is the rotation matrix, is the translation vector. ICP precise registration: The iterative closest point (ICP) algorithm is used for the tree trunk area to achieve precise registration.

[0045] Experimental example Using the registration model and method of the present invention, experiments are carried out on 5 sample plot survey areas. The experimental details are shown in Figure 2 . The pictures show the rough and fine registration results of 5 representative forest maps, as well as local detailed information. Ⅰ to Ⅴ represent the rough alignment (left, a) and fine registration (right, b) results of the point clouds of the unmanned aerial vehicle (blue) and the ground platform (green) on the test plots. From left to right, the first column and the fourth column show the top views of the registration details, the second column and the fifth column show the side views of the registration details, and the third column shows the local registration situation of the rough registration. The results show that accurate spatial alignment has been achieved between the airborne and ground platforms. Among them, the 5 sample plots cover different tree species, growth stages, regional overlaps, tree densities, tree layouts, and collection times. The sample plot characteristics are summarized as follows: 1. Tree maturity: Mature trees (sample plots Ⅰ, Ⅱ, Ⅳ, and Ⅴ); saplings (sample plot Ⅲ) 2. Overlap rate: Complete overlap (sample plots Ⅰ, Ⅱ, Ⅲ, and Ⅴ); partial overlap (sample plot Ⅳ) 3. Tree density: Sparse (sample plots Ⅰ, Ⅱ, Ⅳ, and Ⅴ); dense (sample plot Ⅲ) 4. Spatial layout: Regular (sample plots Ⅰ - Ⅳ); irregular (sample plot Ⅴ) 5. Acquisition time: March (plots I-IV); August (plot V) 6. Tree species: Rubber forest (plots I-IV); Eucalyptus and plane trees (plot V) Despite considerable variations in forest conditions, the proposed framework achieved consistent and accurate point cloud alignment in all plots, demonstrating strong robustness and generalization capabilities. The overall average point-to-point residual was 0.1 m. Among the 5 plots, plot III had the highest accuracy, mainly due to the sparse tree canopy during data acquisition, which enabled a more complete structural representation in both airborne and ground platform point clouds. Additionally, the upright growth pattern of rubber saplings (plot V) resulted in strong spatial consistency between homologous key points. However, the reduced feature significance due to the similar height morphology among individual trees in this plot led to a lower feature recall rate compared to plots I and II. Severe canopy occlusion in plot V caused significant structural incompleteness in the airborne point cloud data, resulting in the extraction of only a limited number of robust key points. This greatly reduced the quality of the initial correspondences and ultimately the accuracy of the coarse registration.

[0046] Figure 2 The specific registration experiment results of the corresponding present invention are shown in Table 1. Table 1 records the recall rate and registration accuracy of each plot after point cloud registration of 5 different plots by the model method of the present invention.

[0047] Table 1

[0048] Comparative example While the registration method TPM of the present invention completed the registration, three commonly used registration methods, ERLD, CBAR, and TEASR++, were selected from the prior art and, like the present invention, were used to register the same initial point cloud data obtained from the same 5 survey areas, and the comparative experiment results shown in Table 2 were obtained.

[0049] ERLD is a registration method specifically for tree position data. It constructs translation and rotation invariant local descriptors based on the planar coordinates (X, Y) of individual trees to generate a set of initial matching pairs, then estimates the rotation and translation parameters in the two-dimensional plane for each candidate matching pair, selects the optimal transformation by maximizing the number of matching tree pairs, and further optimizes the transformation parameters.

[0050] CBAR is a clustering-based automatic registration method. It utilizes the local structure of tree trunks in ULS and TLS point clouds, generates multi-level tree structure maps through hierarchical clustering, performs layer-by-layer registration using FPFH features, and selects the optimal transformation to achieve accurate registration.

[0051] TEASER++ is a verifiable 3D point cloud registration algorithm designed specifically for high outlier rates. It enhances outlier robustness through truncated least squares optimization, decouples scale, rotation, and translation estimation using graph theory methods, and achieves efficient and globally optimal or near-optimal registration results through semidefinite relaxation and splitting optimization.

[0052] Table 2

[0053] As shown in Table 2, only the registration method TPM of the present invention can achieve stable and reliable registration under various conditions such as different tree species, growth stages, regional overlap, tree density, spatial arrangement, and acquisition time. While ERLD, CBAR, and TEASR++ can only complete registration in specific survey areas.

[0054] As shown in Table 2, the ERLD method fails in plots IV and V. The maximum rough registration error in plot IV reaches 189.7 meters. This failure is mainly attributed to its limited feature representation ability and strong dependence on a single descriptor. This method constructs a rotation matrix based on its own orientation features and selects the optimal transformation by maximizing the number of inliers, which is prone to error in cases of structural ambiguity or repetition - especially obvious in plot IV. In plot V, due to severe occlusion of the tree crown in the ULS data, a large number of key points are lost, and the method fails again. These results highlight the problems of weak feature robustness and poor stability of this method in complex scenarios with severely incomplete key points.

[0055] The CBAR method shows a relatively low rough registration error in most plots (such as plots I - IV), but fails in plot V with a maximum error of 21.9 meters. This instability is mainly due to the severely incomplete structure of the low-altitude point cloud obtained by the UAV in plot V. In a forest scenario with strong structural heterogeneity, relying on a single descriptor may lead to biased feature representation, resulting in incorrect matching and increased registration error.

[0056] The TEASER++ method generally shows stable performance, but there is a significant decrease in accuracy. The maximum error in plot IV is 0.95 meters, and in plot V is 0.47 meters. This method uses downsampled points as key points and extracts local geometric features from their neighborhoods using the FPFH descriptor. Especially in large-scale forest plots, this leads to a large amount of computational overhead and poor correspondence of downsampled key points.

[0057] In contrast, the registration method proposed by the present invention exhibits stable performance in all five test areas, with small coarse registration errors and extremely low errors (RMSE is only 0.06 meters) maintained in the complex scenarios of Plot IV and Plot V, highlighting its robustness and generalization ability in complex forest scenarios. First, by integrating the structural features of point clouds at different height levels, the method effectively solves the failure problem caused by inconsistent structures of a single descriptor, enhancing the adaptability of the model to changes in the forest vertical structure. Second, extracting key points based on the trunk position ensures good spatial correspondence and avoids the matching interference in low-quality areas, thereby improving the registration robustness. Finally, a multi-level matching pair screening strategy that combines feature similarity and geometric consistency for hierarchical filtering significantly improves the reliability of the matching pairs, providing strong support for the subsequent fine registration stage.

[0058] The above are only embodiments of the present invention, and common general technical knowledge such as specific technical solutions and / or characteristics known in the solution are not described in detail herein. It should be noted that for those skilled in the art, without departing from the technical solution of the present invention, several deformations and improvements can still be made, and these should also be regarded as the protection scope of the present invention, which will not affect the implementation effect of the present invention and the practicality of the patent. The protection scope required by this application should be subject to the content of its claims, and the specific implementation manners described in the specification can be used to interpret the content of the claims.

Claims

1. A method for airborne and ground platform laser point cloud registration in forest scenarios, characterized in that, It includes the following steps: Step 1: Preprocess the point cloud data; Step 2: Model construction, successively performing extraction of key points of tree positions, construction of vertically stratified virtual points, and construction of tree vertical stratification aggregation descriptors; when constructing the vertically stratified virtual points, a predefined vertical axis length is used to ensure that the vertical axes of all trees are evenly divided into N layers within the same height range, and the center point of each layer is used as the virtual construction point; Step 3: Feature matching, successively performing fault-tolerant matching selection, geometric consistency screening, and random sample consensus; Step 4: Point cloud registration, successively calculating the rotation matrix based on the matching pairs and performing ICP precise registration.

2. The method for airborne and ground platform laser point cloud registration in forest scenes according to claim 1, characterized in that In Step 2, the construction of the virtual points is specifically as follows: Standardize the downsampled point cloud, use the slicing method to extract the lower trunk area of the tree, locate the trunk position to extract the horizontal coordinates (X, Y), construct two-dimensional key points, and use the average elevation of all ground points in the buffer to represent its Z coordinate to construct three-dimensional key points; for each two-dimensional key point, construct a vertical axis passing through this point and evenly divide it into N layers, and the center point of each layer is the virtual point constructed for the corresponding layer of the key point. Introduce a unified axis height parameter to ensure that the vertical axis lengths of all trees are the same. The j-th virtual point of its i-th key point The calculation formula is as follows: ， is the unified axis height parameter, is the number of layers evenly divided.

3. The method for airborne and ground platform laser point cloud registration in a forest scene according to claim 2, wherein The construction of the tree vertical stratification aggregation descriptor in Step 2 includes virtual point feature extraction and feature order splicing; the virtual point feature extraction is specifically as follows: Based on the virtual points of each layer constructed by the key points, use the fast point feature histogram FPFH to extract the local geometric features of the virtual points of each layer in the standardized point cloud. If a certain layer only contains virtual points and lacks effective point cloud data, it is considered that the feature extraction of this layer fails, mark it as an invalid layer and skip the calculation of this layer, and perform L1 norm standardization processing on the successfully extracted features.

4. The method for airborne and ground platform laser point cloud registration in forest scenes according to claim 3, characterized in that, The feature order splicing is specifically as follows: Splice the standardized features of each layer of the same axis calculated successfully in the order from bottom to top or from top to bottom to generate the tree vertical stratification aggregation descriptor.

5. The method for airborne and ground platform laser point cloud registration in forest scenes according to claim 1, characterized in that, The fault-tolerant matching selection in Step 3 is specifically as follows: For the key point sets of the given source point cloud and target point cloud, calculate the Manhattan distance between the associated tree vertical stratification aggregation descriptor features, and select the top n most similar points of the source point cloud key points and target point cloud key points to construct a matching pair candidate set, where n is set to 3.

6. The method for airborne and ground platform laser point cloud registration in forest scenes according to claim 1, characterized in that, The geometric consistency screening in Step 3 is specifically as follows: Introduce a geometric consistency analysis method, obtain the geometric consistency score GCS by calculating the geometric consistency distance between the matching pairs; construct a GCS matrix GCSM for voting screening according to the score, identify the maximum score item in the matrix, set all elements in its row and column to zero, and repeat the operation until all unmarked matrix elements become zero to optimize the matching pairs.

7. The method for airborne and ground platform laser point cloud registration in forest scenarios according to claim 1, characterized in that Calculating the rotation matrix based on the matching pairs in Step 4 is specifically as follows: Using the extracted corresponding set of key points, calculate the rigid transformation parameters for converting the target point cloud to the source point cloud coordinate system according to the matching pairs, set the scale factor of the rotation matrix to 1, and calculate the rotation matrix and translation vector.

8. A model for airborne and ground platform laser point cloud registration in forest scenarios, characterized in that, It includes a hierarchical vertical aggregation feature model: The vertical hierarchical virtual point construction module is used to standardize the downsampled point cloud, extract key points in the lower trunk area of the tree, construct a vertical axis and generate multiple layers of virtual points, and introduce a unified axis height parameter to ensure the consistency of the vertical axis length; The tree vertical hierarchical aggregation descriptor construction module is used to extract local geometric features based on the virtual points, and splice the successfully extracted coaxial standardized features in sequence to form the tree vertical hierarchical aggregation descriptor.

9. The model for airborne and ground platform laser point cloud registration in forest scenes according to claim 8, wherein It also includes a fault-tolerant multi-level matching pair screening model: The initial matching set construction module is used to construct an initial matching relationship set of the key points of the source point cloud and the target point cloud by using the Manhattan distance; The geometric consistency screening module is used to calculate the geometric consistency score GCS, construct a GCS matrix GCSM for voting screening, and optimize the matching pairs; The abnormal matching pair elimination module is used to further eliminate abnormal matching pairs by using the RANSAC algorithm to obtain high-quality matching pairs.

Images