Model and method for airborne and ground-based laser point cloud registration in forest scenes

Abstract

The present invention relates to the technical field of three-dimensional point cloud data processing, and discloses a model and method for airborne and ground-based laser point cloud registration of forest scenes. The method comprises the following steps: preprocessing point cloud data; model construction, sequentially performing tree position key point extraction, vertical layered virtual point construction, and tree vertical layered aggregate descriptor construction; when constructing the vertical layered virtual point, using predefined vertical axis lengths to ensure that the vertical axes of all trees are evenly divided into N layers within the same height range, with the center point of each layer serving as a virtual construction point; feature matching, sequentially performing fault-tolerant matching selection, geometric consistency screening, and random sampling consistency; and point cloud registration, sequentially performing matching pair-based rotation matrix calculation and ICP precise registration. The present invention does not rely on the complete and accurate extraction of tree position key points, and the constructed combined feature has strong descriptive capabilities and is highly robust to regularly distributed forests.

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 precision management and ecological research, LiDAR (Light Detection and Ranging) technology has become an indispensable tool for forest inventory due to its significant advantages in automation, speed, efficiency, and high-precision measurement. UAV laser scanning systems, working in tandem with terrestrial laser scanning systems (including ground-based laser scanners and mobile laser scanners), can accurately capture detailed information such as tree trunk morphology and spatial distribution within a forest. This provides a solid data foundation for core applications such as tree parameter extraction, biomass estimation, vegetation type classification, and pest and disease monitoring.

[0003] However, the complex three-dimensional spatial structure of forests presents significant challenges for point cloud data processing. Due to differences in observation angles, point cloud data collected by drones and ground-based LiDAR systems exhibit significant heterogeneity in density, coverage, and spatial distribution. Furthermore, inter-canopy occlusion by trees results in missing or redundant data in some areas, making it difficult for a single-platform LiDAR point cloud to comprehensively and accurately represent the complete structure of the forest. Only by deeply integrating drone and ground-based LiDAR point cloud data can a comprehensive, high-precision modeling of forest ecosystems be achieved. In this process, point cloud data registration, a critical prerequisite for fusion, directly determines the accuracy and reliability of subsequent analysis.

[0004] Currently, relevant research has been conducted on point cloud registration. For example, patent CN113205548A, "An Automated Registration Method and System for UAV and Ground-Based Point Clouds in Forest Areas," and patent CN118799363A, "A Method for Registration of UAV and Ground-Based Forest Point Clouds," both provide their own registration methods, which, to a certain extent, achieve registration goals. However, both methods place high demands on the quality of the initial point cloud data collected and rely heavily on the complete and accurate extraction of key points at tree locations, making them difficult to implement in many practical application scenarios.

[0005] Therefore, point cloud registration technology in forest scenes still has many bottlenecks that need to be solved:

[0006] First, georeferencing-based registration methods rely heavily on precise geographic coordinate information to align point cloud data from different platforms. However, in forest environments, especially dense forests with high canopy density, the canopy severely blocks satellite signals, leading to a significant decrease in absolute positioning accuracy, which greatly limits the scope and effectiveness of this method.

[0007] Second, the inherent weak structural characteristics of forest scenes cause the local feature distributions of different key points to be highly similar. A single feature descriptor cannot accurately describe the unique attributes of each key point, resulting in low feature matching accuracy and seriously hindering the smooth progress of point cloud registration.

[0008] Third, due to the obstruction of tree crowns, local descriptors constructed based on tree positions and attributes (such as diameter at breast height and tree height) are difficult to achieve accurate matching in environments with relatively fixed layout rules such as artificial forests due to the extremely high similarity of local features, which directly affects the overall accuracy of point cloud registration.

[0009] It can be seen that the development of airborne and ground platform laser point cloud registration models and methods suitable for forest scenes can break through existing technical bottlenecks and have important practical significance and application value for improving the level of forest resource monitoring and ecological research. Summary of the Invention

[0010] The present invention aims to provide a model and method for the registration of airborne and ground-based laser point clouds in forest scenes, so as to solve the problem that the existing tree position-based feature extraction method in the registration of forest scene UAV and ground-based LiDAR point clouds relies on the complete and accurate extraction of tree position key points, and the constructed simple feature description ability is weak and has low robustness to regularly distributed forests.

[0011] In order to solve the above problems, the present invention adopts the following technical solutions:

[0012] Solution 1: A method for airborne and ground-based laser point cloud registration of forest scenes, including the following steps:

[0013] Step 1: preprocess the point cloud data;

[0014] Step 2: Model construction, which involves extracting key points from tree locations, constructing vertical layered virtual points, and constructing tree vertical layered aggregate descriptors. 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, with the center point of each layer serving as a virtual construction point.

[0015] Step 3: Feature matching, which involves fault-tolerant matching selection, geometric consistency screening, and random sampling consistency.

[0016] Step 4: Point cloud registration, which involves calculating the rotation matrix based on the matching pairs and performing ICP precise registration.

[0017] Preferably, in step 2, the virtual point construction 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 and extract the horizontal coordinates (X, Y), construct a two-dimensional key point, use the average elevation of all ground points in the buffer zone to represent its Z coordinate, and construct a three-dimensional key point; for each two-dimensional key point, construct a vertical axis through the point, and divide it evenly into N layers, the center point of each layer is the virtual point constructed in the corresponding layer of the key point, introduce a unified axis height parameter to ensure that the vertical axis length of all trees is consistent, and the virtual point of the i-th key point and the j-th layer is The calculation formula is:

[0018] ,

[0019] is a unified axis height parameter, is the number of evenly divided layers.

[0020] Preferably, the construction of the vertical hierarchical aggregated descriptor of the trees in step 2 includes virtual point feature extraction and feature sequence splicing; the virtual point feature extraction is specifically: taking each layer of virtual points constructed by key points as a benchmark, using the fast point feature histogram FPFH to extract the local geometric features of each layer of virtual points in the standardized point cloud, if a layer only contains virtual points but lacks valid point cloud data, it is determined that the feature extraction of this layer has failed, it is marked as an invalid layer and the calculation of this layer is skipped, and the successfully extracted features are subjected to L1 norm normalization processing.

[0021] Preferably, the feature sequence splicing is specifically: splicing the successfully calculated standardized features of each layer of the coaxial line in a bottom-up or top-down order to generate a tree vertical layered aggregation descriptor.

[0022] Preferably, the fault-tolerant matching selection in step three is specifically as follows: for a given set of key points of a source point cloud and a target point cloud, the Manhattan distance between their associated tree vertical hierarchical aggregation descriptor features is calculated, and the top n most similar points of the source point cloud key points and the target point cloud key points are selected to construct a matching pair candidate set, where n is set to 3.

[0023] Preferably, the geometric consistency screening in step three is specifically as follows: introducing a geometric consistency analysis method, obtaining a geometric consistency score GCS by calculating the geometric consistency distance between matching pairs; constructing a GCS matrix GCSM according to the score for voting screening, identifying the maximum score item in the matrix, setting all elements in its row and column to zero, repeating the operation until all unmarked matrix elements become zero, and optimizing the matching pairs.

[0024] Preferably, the calculation of the rotation matrix based on the matching pairs in step 4 is specifically as follows: using the extracted key point correspondence set, according to the matching pairs, the rigid transformation parameters of the target point cloud converted to the source point cloud coordinate system are calculated, the scale factor of the rotation matrix is set to 1, and the rotation matrix and translation vector are calculated.

[0025] Solution 2: Model for airborne and ground-based laser point cloud registration of forest scenes, including a hierarchical vertical aggregation feature model:

[0026] The vertical layered virtual point construction module is used to standardize the downsampled point cloud, extract key points in the lower trunk area of the tree, construct the vertical axis and generate multiple layers of virtual points. A unified axis height parameter is introduced to ensure the consistency of the vertical axis length.

[0027] The tree vertical hierarchical aggregate descriptor construction module is used to extract local geometric features based on virtual points, and sequentially splice the successfully extracted coaxial standardized features to form a tree vertical hierarchical aggregate descriptor.

[0028] Preferably, a fault-tolerant multi-level matching pair screening model is also included:

[0029] The initial matching set construction module is used to construct the initial matching relationship set of key points between the source point cloud and the target point cloud using Manhattan distance;

[0030] The geometric consistency screening module is used to calculate the geometric consistency score GCS, construct the GCS matrix GCSM for voting screening, and optimize the matching pairs;

[0031] The abnormal matching pair elimination module is used to further eliminate abnormal matching pairs using the RANSAC algorithm to obtain high-quality matching pairs.

[0032] Principle of the present invention:

[0033] This invention achieves airborne and ground-based laser point cloud registration of forest scenes through a comprehensive set of methodological processes and model systems. At the methodological level, the collected point cloud data is first preprocessed. Through point cloud downsampling, ground separation, and point cloud normalization, concise and representative data is provided for subsequent processing. The model construction phase then begins, with key point extraction and virtual point construction. By extracting key points from tree trunks, constructing vertical axes, and generating vertical layered virtual points in layers, the axis heights are unified to ensure correspondence. Features are then extracted and spliced based on the virtual points to form layered vertical aggregate features. Feature matching is then performed, using Manhattan distance to construct an initial matching set. Matching pairs are screened through geometric consistency analysis, and the RANSAC algorithm is used to eliminate anomalous matching pairs and determine the accurate matching relationship. Finally, a rotation matrix is calculated based on the matching pairs, and the ICP algorithm is combined to achieve precise point cloud registration.

[0034] In terms of the model, the hierarchical vertical aggregation feature model starts from the vertical morphology of trees. The vertical hierarchical virtual point construction module obtains virtual points, and the tree vertical hierarchical aggregation descriptor construction module further forms a unique feature description. 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 rejection module, layer by layer screens high-quality matching pairs, providing a reliable basis for point cloud registration. The two complement each other and are closely integrated with the method steps to jointly complete the point cloud registration task.

[0035] Compared with the existing technology, the present invention does not rely on the complete and accurate extraction of key points of tree positions. The constructed combined features have strong description capabilities and are highly robust to regularly distributed forests.

[0036] Advantages of the present invention:

[0037] High accuracy: The layered vertical aggregation feature model fully exploits the vertical morphological characteristics of trees and integrates features at different heights to make the feature vector highly unique, reducing false matches caused by tree similarities. At the same time, it uses features at different heights to compensate for the impact of incomplete data. The robust fault-tolerant multi-level matching screening model uses a progressive strategy to gradually impose stricter constraints and screening rules using relaxed matching, tightening the matching criteria, effectively eliminating false matches, and obtaining accurate matching relationships, thereby achieving high-precision point cloud registration.

[0038] Strong robustness: During the virtual point construction process, the method adopted can effectively filter out low-quality data to address the problems of missing lower parts of drone point clouds and difficulty in obtaining accurate tree heights using ground LiDAR. By unifying the axis height, the influence of tree height differences is overcome. The use of a unified virtual axis height enhances the correspondence between virtual points with the same name in the airborne and ground platform point cloud data, thereby improving the consistency of features with the same name 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 differences, and repeated structural features in forest scenes, enabling the model to maintain good performance even in complex environments.

[0039] Wide adaptability: The method and model comprehensively consider the various complex factors of forest scenes. Whether it is the perspective differences between drone and ground LiDAR point clouds, data heterogeneity, or canopy occlusion, there are corresponding processing mechanisms. It is applicable to different types of forest environments, including regularly distributed trees and complex natural forests.

[0040] Outstanding innovation: It breaks through the limitations of existing tree position feature extraction, proposes a hierarchical vertical aggregation feature model and a robust fault-tolerant multi-level matching pair screening framework, solves the problem of forest scene point cloud registration from a new perspective, and provides new technical ideas and solutions for this field. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 Flowchart of an embodiment of the present invention.

[0042] Figure 2 This is a comparison chart of the registration results of five measurement areas according to an embodiment of the present invention.

[0043] Figure 3 These are virtual points constructed at the same location for 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 layered virtual points. DETAILED DESCRIPTION

[0044] The following is further described in detail through specific implementation methods:

[0045] The embodiment is basically as shown in the attached Figure 1 The method for airborne and ground-based laser point cloud registration of forest scenes includes the following:

[0046] The first step is to preprocess the point cloud data, sequentially performing point cloud downsampling, ground separation, and point cloud normalization. Gaussian filtering is used to remove noise points, and the Cloth Simulation Filter (CSF) algorithm is used to separate ground and non-ground point clouds. The non-ground point clouds are then normalized.

[0047] The second step is model construction, which involves extracting key points of tree positions, constructing vertical layered virtual points, and constructing vertical layered aggregation descriptors of trees. The vertical layered virtual point construction includes virtual point feature extraction and feature sequence splicing.

[0048] The third step is feature matching, which involves fault-tolerant matching selection, geometric consistency screening, and random sampling consistency.

[0049] The fourth step is point cloud registration, which involves calculating the rotation matrix and ICP fine registration based on the matching pairs.

[0050] This paper addresses the limitations of existing tree location-based feature extraction methods for UAV and ground-based LiDAR point cloud registration in forest scenes, which rely on the complete and accurate extraction of key points, have limited capability for constructing simple feature descriptions, and are less robust to regularly distributed trees. By deeply exploring the structural similarities of forest scenes, this paper proposes an innovative tree point matching model for airborne and ground-based LiDAR point cloud registration in weakly structured forests. The model primarily includes an innovative tree vertical hierarchical aggregation descriptor that describes the vertical structural morphology of trees, and an innovative robust, fault-tolerant, multi-level matching pair screening framework.

[0051] The model of the present invention for airborne and ground platform laser point cloud registration of forest scenes includes a hierarchical vertical aggregation feature model and a robust fault-tolerant multi-level matching pair screening model.

[0052] The layered vertical aggregation feature model describes the vertical morphology of trees and consists of two parts: vertical layered construction point extraction and tree vertical aggregation feature descriptor construction. The former constructs a vertical axis from key points and generates virtual points, using a unified axis height parameter to ensure correspondence. The latter extracts local geometric features from the virtual points, standardizes them, and then splices them to form a layered vertical aggregation feature. This model integrates features at different heights to enhance the uniqueness of feature vectors and reduce mismatches. It also uses features at different heights to compensate for invalid features in some layers caused by incomplete data.

[0053] A robust, fault-tolerant multi-stage matching pair screening model was designed to address the problem of mismatches in forest scene point cloud registration. This model employs a progressive strategy, leveraging relaxed matching to gradually impose stricter constraints and screening rules, tightening the matching criteria to obtain precise matching relationships. The model first constructs an initial set of matches using Manhattan distance, then calculates the GCS through geometric consistency analysis, constructs a GCSM for voting screening, and finally uses the RANSAC algorithm to eliminate outliers, obtaining high-quality matches. This model then estimates the transformation matrix for initial point cloud alignment.

[0054] Technical highlights of the present invention:

[0055] (1) Virtual point construction: Traditional methods suffer from poor feature similarity due to the heterogeneity of UAV and ground LiDAR point clouds, resulting in poor correspondence between key points with the same name and key points with the same name. This paper introduces a method of standardizing the height axis and constructing virtual points along the axis. That is, by pre-defining the length of the vertical axis, 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, significantly improves the matching robustness of the same-name points between the UAV and ground platform point clouds, and shows excellent virtual point correspondence capabilities, especially in complex forest scenes.

[0056] (2) Construction of vertical hierarchical aggregation descriptors for trees: Construct hierarchical vertical aggregation features that aggregate features from multiple height layers. By extracting the geometric features of the neighborhood of the construction points in layers and normalizing them, a layer-by-layer splicing feature vector is formed. This framework improves matching accuracy through the following advantages:

[0057] Feature complementarity: Integrate features from different height layers to enhance the uniqueness of feature vectors and reduce the probability of mismatching due to similar tree morphologies;

[0058] Feature uniqueness: By fusing features at different heights, the feature vector becomes more unique. This reduces mismatches caused by high tree similarity using a single descriptor and improves matching accuracy.

[0059] Feature robustness: The local feature extraction errors caused by point cloud heterogeneity in single descriptors at different layers can still be matched with existing features at different heights, effectively compensating for the impact of incomplete data or point cloud heterogeneity.

[0060] (3) Fault-tolerant multi-level matching: Design a three-level matching process that includes Manhattan distance screening, geometric consistency voting matrix, and RANSAC optimization:

[0061] This strategy significantly reduces the number of false matches, improving the matching rate and robustness of the registration task. Furthermore, it introduces a fault-tolerant mechanism, allowing more correct matches to be included in the screening set, significantly improving the matching rate. This method is applicable to the registration of partially overlapping and regularly distributed forest scene point clouds.

[0062] The specific implementation process is as follows:

[0063] The first step is point cloud data preprocessing

[0064] The raw point cloud data is processed in the order of point cloud downsampling, ground separation, and point cloud normalization. Point cloud downsampling reduces data density and reduces subsequent computational effort; ground separation distinguishes ground and non-ground points; point cloud normalization unifies the data scale; and key point extraction selects representative points for subsequent operations.

[0065] Step 2: Model construction

[0066] Extraction of key points at tree locations and construction of vertical layered 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 using 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 the lower point cloud to be missing due to occlusion and key points to be missed, low-quality data can be filtered. Then, for each key point A vertical axis is constructed through this point and divided into N layers. The center point of each layer is the virtual point V. Considering that the ground LiDAR is difficult to accurately obtain the tree height due to occlusion or data quality, a unified axis height parameter is introduced to ensure that the vertical axis length of all trees is consistent. The virtual point of the jth layer of its i-th key point is The calculation formula is:

[0067] (Formula 1). is a unified axis height parameter, is the number of evenly divided layers.

[0068] Virtual point feature extraction: Based on virtual points, the fast point feature histogram (FPFH) is used to extract local geometric features of virtual points in each layer. If a layer only contains virtual points but no valid point cloud data, it is marked as an invalid layer and the calculation of this layer is skipped. The successfully extracted features are normalized. Figure 3 .

[0069] Feature sequence splicing: Splice the successfully calculated features of each layer in a bottom-up or top-down order to form a tree vertical layered aggregation descriptor. , spliced in a bottom-up or top-down order to form a tree vertical hierarchical aggregation descriptor V.

[0070] (Formula 2).

[0071] The third step is feature matching

[0072] Error-tolerant matching selection: For the source point cloud key point set

[0073] and target point cloud keypoint set

[0074] ,

[0075] Calculate the tree vertical hierarchical aggregation descriptor features

[0076]

[0077] and The Manhattan distance between the two points is used to select the first n (n=3 in this experiment) most similar points to construct a matching pair candidate set. The formula is:

[0078] (Formula 3).

[0079] express and The Manhattan distance between Indicates selecting the top n most similar key points. In this experiment, n is set to 3.

[0080] Geometric consistency screening: Geometric consistency analysis is introduced to obtain the geometric consistency score GCS by calculating the geometric consistency distance between matching pairs.

[0081] Assume existence and , the geometric distance calculation formula for two matching pairs is:

[0082] (Formula 4);

[0083] Cumulative geometric consistency score The calculation formula is:

[0084] (Formula 5).

[0085] A GCS matrix (GCSM) is constructed based on the scores for voting screening. The maximum score item is identified in the matrix, and the elements in its row and column are set to zero. This operation is repeated until all unmarked elements are zero to optimize the matching pairs.

[0086] is the indicator function, if , then it returns 1, otherwise it returns 0. is the geometric consistency threshold. This score reflects the degree of spatial distribution consistency between matching pairs. The higher the score, the more reliable the matching pair.

[0087] Next, the GCS matrix (GCSM) is constructed for voting selection.

[0088] 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 in 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.

[0089] Through the above screening strategy, we can effectively screen higher quality matching pairs. Finally, we form a high-quality matching set C G .

[0090] 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 .

[0091] First, the initial correspondence is generated, 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.

[0092] 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 where the element is located 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.

[0093] Finally, for the candidate corresponding set C G Using the RANSAC algorithm, the corresponding point pairs with geometric consistency are screened out based on the inlier threshold θ. These screened inlier pairs constitute the final inlier correspondence set C. F The entire matching algorithm adopts a progressive strategy, using relaxed matching to gradually impose stricter constraints and screening rules, tightening the matching criteria to obtain an accurate matching relationship.

[0094] Step 4: Point Cloud Registration

[0095] Calculate the rotation matrix based on the matching pairs: According to the corresponding set of extracted key points, the matching pairs are used to calculate the optimal rigid transformation matrix parameters for transforming the target point cloud into the source point cloud coordinate system. Assuming the rotation matrix scale factor is 1, the coarse registration process is expressed as:

[0096] (Formula 6).

[0097] Among them, X, Y, Z are the coordinates of the source point cloud, , , is the coordinate of the target point cloud, is the rotation matrix, is the translation vector.

[0098] ICP precise registration: The iterative closest point (ICP) algorithm is used for the tree trunk area to achieve precise registration.

[0099] Experimental example

[0100] The registration model and registration method of the present invention were used to conduct experiments on five plots. Figure 2, the pictures show the coarse and fine registration results of 5 representative forest maps, as well as local detailed information. Ⅰ to Ⅴ represent the coarse alignment (left, a) and fine registration (right, b) results of the UAV (blue) and ground platform (green) point clouds on the test plot. From left to right, the 1st and 4th columns show the top views of the registration details, the 2nd and 5th columns show the side views of the registration details, and the 3rd column shows the local registration of the coarse registration. The results show that accurate spatial alignment has been achieved between the point clouds of the airborne and ground platforms. Five of the sample plots cover different tree species, growth stages, regional overlap, tree density, tree layout and collection time. The sample plot characteristics are summarized as follows:

[0101] 1. Tree maturity: Mature trees (plots I, II, IV, and V); saplings (plot III)

[0102] 2. Overlap rate: Complete overlap (plots I, II, III, and V); Partial overlap (plot IV)

[0103] 3. Tree density: sparse (plots I, II, IV, and V); dense (plot III)

[0104] 4. Spatial layout: Regular (plots I-IV); Irregular (plot V)

[0105] 5. Collection time: March (sample plots I-IV); August (sample plot V)

[0106] 6. Tree species: Rubber forest (plots I-IV); eucalyptus and sycamore (plot V)

[0107] Despite considerable variation in forest conditions, the proposed framework achieved consistently accurate point cloud alignment across all plots, demonstrating strong robustness and generalization. The overall average point-to-point residual error was 0.1 m. Among the five plots, Plot III achieved the highest accuracy, primarily due to the sparse tree canopy during data collection, which resulted in a more complete structural representation in both the airborne and ground-based point clouds. Furthermore, the upright growth pattern of rubber tree seedlings (Plot V) resulted in strong spatial consistency between keypoints with the same name. However, the high morphological similarity between individual trees in this plot reduced feature saliency, resulting in lower feature recall compared to Plots I and II. Severe canopy occlusion in Plot V resulted in significant structural incompleteness in the airborne point cloud data, resulting in the extraction of a limited number of robust keypoints. This significantly reduced the quality of the initial alignments and ultimately the accuracy of the coarse registration.

[0108] Figure 2 The corresponding specific registration experiment results of the present invention are shown in Table 1. Table 1 records the recall rate and registration accuracy of each sample plot after performing site cloud registration on 5 different sample plots using the model method of the present invention.

[0109] Table 1

[0110]

[0111] Comparative Example

[0112] While the registration method TPM of the present invention completes the registration, three commonly used registration methods, ERLD, CBAR, and TEASR++, are selected from the prior art to register the same initial point cloud data obtained from the same five measurement areas as in the present invention, and the comparative experimental results are obtained as shown in Table 2.

[0113] ERLD is a registration method specifically for tree location data. It constructs translation- and rotation-invariant local descriptors based on the plane coordinates (X, Y) of individual trees to generate an initial set of matching pairs. It 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.

[0114] CBAR is a clustering-based automatic registration method that utilizes the local structure of tree trunks in ULS and TLS point clouds, generates multi-level tree structure maps through hierarchical clustering, uses FPFH features for layer-by-layer registration, and selects the optimal transformation to achieve accurate registration.

[0115] TEASER++ is a verifiable 3D point cloud registration algorithm designed for high outlier rates. It enhances outlier robustness through truncated least-squares optimization, decouples scale, rotation, and translation estimation using graph theory, and achieves efficient and globally optimal or near-optimal registration results through semidefinite relaxation and splitting optimization.

[0116] Table 2

[0117]

[0118] As shown in Table 2, only the proposed registration method, TPM, achieves stable and reliable registration across a variety of conditions, including tree species, growth stages, regional overlap, tree density, spatial arrangement, and acquisition time. ERLD, CBAR, and TEASR++, on the other hand, can only achieve registration within specific measurement areas.

[0119] As shown in Table 2, the ERLD method failed registration in Sites IV and V, with the maximum coarse registration error reaching 189.7 meters in Site IV. This failure is primarily attributed to its limited feature representation capabilities and strong reliance on a single descriptor. This method constructs a rotation matrix based on its own directional features and selects the optimal transformation by maximizing the number of inliers, which is prone to errors in cases of ambiguous or repetitive structures—particularly evident in Site IV. In Site V, the method failed again due to the loss of a large number of key points due to severe occlusion by tree canopies in the ULS data. These results highlight the method's weak feature robustness and poor stability in complex scenes with severely incomplete key points.

[0120] The CBAR method demonstrated low coarse registration errors in most plots (e.g., plots I-IV), but failed in plot V, with a maximum error of 21.9 meters. This instability was primarily due to the severe structural incompleteness of the low-altitude point cloud acquired by the drone in plot V. In forest scenes with strong structural heterogeneity, reliance on a single descriptor can lead to biased feature representation, resulting in mismatches and increased registration error.

[0121] The TEASER++ method performed consistently well overall, but showed a significant decrease in accuracy, with maximum errors of 0.95 meters for plot IV and 0.47 meters for plot V. This method uses downsampled points as keypoints and extracts local geometric features from their neighborhood using the FPFH descriptor. This can lead to significant computational overhead, especially in large forest plots, and poor correspondence between downsampled keypoints.

[0122] In contrast, the registration method proposed in the present invention showed stable performance in all five test areas, with small coarse registration errors, and maintained extremely low errors (RMSE was only 0.06 meters) in the complex scenes of sample sites IV and V, highlighting its robustness and generalization ability in complex forest scenes. First, by integrating the structural features of point clouds at different height layers, this method effectively solves the failure problem of a single descriptor due to structural inconsistency, and enhances the model's adaptability to changes in the vertical structure of the forest. Secondly, the extraction of key points based on the position of tree trunks ensures good spatial correspondence and avoids matching interference in low-quality areas, thereby improving the robustness of the registration. Finally, a multi-level matching pair screening strategy that combines feature similarity and geometric consistency for hierarchical filtering significantly improves the reliability of matching pairs, providing strong support for the subsequent fine registration stage.

[0123] The above is only an embodiment of the present invention, and the common knowledge such as the specific technical solutions and / or characteristics in the solution are not described in detail here. It should be pointed out that for those skilled in the art, without departing from the technical solution of the present invention, several variations and improvements can be made, which should also be regarded as the scope of protection of the present invention, and these will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the description can be used to interpret the content of the claims.

Claims

1. A method for airborne and ground-based laser point cloud registration of forest scenes, characterized in that: The following steps are involved: Step 1: preprocess the point cloud data; Step 2: Model construction, which involves extracting key points from tree locations, constructing vertical layered virtual points, and constructing tree vertical layered aggregate descriptors. 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, with the center point of each layer serving as a virtual construction point. Extracting key points at tree locations and constructing vertically layered virtual points: The downsampled point cloud is first standardized. A slicing method is used to accurately extract the lower trunk area of the tree. The trunk position is located to obtain horizontal coordinates (X, Y) to construct 2D key points. The Z coordinate is determined using the average elevation of the ground points within the buffer zone to construct 3D key points. Then, for each key point A vertical axis is constructed through this point and divided into N layers. The center point of each layer is the virtual point V. Considering that the ground LiDAR is difficult to accurately obtain the tree height due to occlusion or data quality, a unified axis height parameter is introduced to ensure that the vertical axis length of all trees is consistent. The virtual point of the jth layer of the i-th key point is The calculation formula is: ; is a unified axis height parameter, is the number of evenly divided layers; The construction of vertical hierarchical aggregate descriptors for trees includes virtual point feature extraction and feature sequence splicing; Virtual point feature extraction: Taking virtual points as the benchmark, the fast point feature histogram is used to extract the local geometric features of virtual points in each layer; If a layer only contains virtual points but no valid point cloud data, it is marked as an invalid layer and the calculation of this layer is skipped. The successfully extracted features are normalized; Feature sequence splicing: The successfully calculated features of each layer are spliced in a bottom-up or top-down order to form a vertical hierarchical aggregation descriptor of the tree; The successfully calculated coaxial K-layer normalized effective features , and spliced in a bottom-up or top-down order to form a tree vertical hierarchical aggregation descriptor V: ； Step 3: Feature matching, which involves fault-tolerant matching selection, geometric consistency screening, and random sampling consistency. Error-tolerant matching selection: For the source point cloud key point set and target point cloud keypoint set ， Calculate the tree vertical hierarchical aggregation descriptor features and The Manhattan distance between the two points is used to select the first n most similar points to build a matching pair candidate set. The formula is: ； express and The Manhattan distance between Indicates selecting the top n most similar key points; Geometric consistency screening: Geometric consistency analysis is introduced to obtain the geometric consistency score GCS by calculating the geometric consistency distance between matching pairs; exist and , the geometric distance calculation formula for two matching pairs is: ； Cumulative geometric consistency score The calculation formula is: ； Based on the scores, a GCS matrix GCSM is constructed for voting screening. The maximum score item is identified in the matrix, and the row and column elements are set to zero. This operation is repeated until all unmarked elements are zero, thus optimizing the matching pairs. is the indicator function, if , then return 1, otherwise return 0; is the geometric consistency threshold; this score reflects the degree of consistency in spatial distribution between matching pairs. The higher the score, the more reliable the matching pair. Next, construct the GCS matrix GCSM for voting selection; 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 matching key points, if their geometric consistency score is not zero, their score is filled in the corresponding position of the matrix. If the matching pair is not selected or the geometric consistency score is zero, fill the corresponding position of the matrix with zero; then, identify and mark the maximum score item in the matrix and set all elements of the row and column where the score item is located to zero to ensure that the matching pair is not repeatedly selected; then, continue to identify the new maximum geometric consistency score item among the remaining unmarked matrix elements and set the elements of the corresponding row and column to zero; this process is repeated until all unmarked matrix elements become zero; Random sampling consistency: First, the initial correspondence is generated, 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; 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; for the initial corresponding set C 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; then, 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; Finally, for the candidate corresponding set C G Using the RANSAC algorithm, the corresponding point pairs with geometric consistency are screened out based on the inlier threshold θ. These screened inlier pairs constitute the final inlier correspondence set C. F And return; the entire matching algorithm adopts a progressive strategy, using loose matching to gradually impose stricter constraints and screening rules, tighten the matching criteria, and obtain an accurate matching relationship; Step 4: Point cloud registration, which involves calculating the rotation matrix based on the matching pairs and performing ICP precise registration.

2. The method for airborne and ground-based laser point cloud registration of forest scenes according to claim 1 is characterized in that: The calculation of the rotation matrix based on the matching pairs described in step 4 is specifically as follows: using the extracted key point correspondence set, according to the matching pairs, the rigid transformation parameters of the target point cloud to the source point cloud coordinate system are calculated, the scale factor of the rotation matrix is set to 1, and the rotation matrix and translation vector are calculated.

3. The method for airborne and ground-based laser point cloud registration of forest scenes according to claim 2 is characterized in that: The iterative closest point ICP algorithm is used for the tree trunk area to achieve accurate registration.

4. The method for airborne and ground-based laser point cloud registration of forest scenes according to claim 1 is characterized in that: n is set to 3.

5. A model for airborne and ground-based laser point cloud registration of forest scenes, characterized by: The method for airborne and ground platform laser point cloud registration of forest scenes according to claim 1 includes a hierarchical vertical aggregation feature model: The vertical layered virtual point construction module is used to standardize the downsampled point cloud, extract key points in the lower trunk area of the tree, construct the vertical axis and generate multiple layers of virtual points. A unified axis height parameter is introduced to ensure the consistency of the vertical axis length. The tree vertical hierarchical aggregate descriptor construction module is used to extract local geometric features based on virtual points, and sequentially splice the successfully extracted coaxial standardized features to form a tree vertical hierarchical aggregate descriptor.

6. The model for airborne and ground-based laser point cloud registration of forest scenes according to claim 5 is characterized in that: Also included are fault-tolerant multi-level matched-pair screening models: The initial matching set construction module is used to construct the initial matching relationship set of key points between the source point cloud and the target point cloud using Manhattan distance; The geometric consistency screening module is used to calculate the geometric consistency score GCS, construct the 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 using the RANSAC algorithm to obtain high-quality matching pairs.

Images