Multi-slice adaptive clustering and cluster split-merge damaged moso bamboo tree recognition method

Abstract

The present application relates to the field of three-dimensional point cloud data processing, and discloses a damaged bamboo single tree recognition method based on multi-slice adaptive clustering and cluster splitting and merging, mainly solving the problem of low recognition rate and easy mis-detection of traditional algorithms caused by bending, dumping and cross adhesion of disaster-stricken bamboo. The method extracts multiple overlapping slices from the point cloud and performs adaptive density clustering; geometric analysis is performed on the initial cluster, and the adhesion cluster meeting the stretching feature is accurately split by using the probability mixture model and the Bayesian information criterion; the sub-cluster is merged across the slices based on spatial proximity and geometric similarity; finally, the multi-peak distribution characteristics of the candidate object in the horizontal projection are tested to suppress false merging and complete deduplication. The present application breaks through the recognition bottleneck of extremely irregular shaped trees and realizes high-precision extraction of complex distorted trees.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional point cloud data processing, and discloses a method for identifying damaged bamboo stalks by multi-slice adaptive clustering and cluster splitting and merging. Background Technology

[0002] Existing technologies can achieve relatively accurate individual tree identification when processing point cloud data of trees with regular shapes and good site conditions. However, for forest stands with complex growing environments and significant inter-tree interactions, especially after disasters, the identification accuracy drops significantly. For example, trees bent or fallen due to windstorms, pests, or diseases, as well as physical adhesion between multiple trees due to branch crossings, make it difficult for traditional clustering or segmentation algorithms based on the assumption of geometric regularity to effectively distinguish individuals.

[0003] Specifically, most current single-tree identification algorithms exhibit multiple shortcomings when dealing with highly irregular bamboo forests. On the one hand, they generally rely on pre-defined canopy or trunk models, making it difficult to adapt to the severely bent, tilted, or even creeping postures of individual bamboo trees caused by disasters, resulting in incorrect fitting or segmentation. On the other hand, in high-density bamboo forests where individuals are intertwined and adhered to each other, clustering methods based on point cloud connectivity or simple spatial distance are prone to misidentifying multiple bamboo trees as a whole, while algorithms that are exceptionally sensitive to the local morphology of damaged bamboo may incorrectly segment multiple parts within a single bamboo tree into multiple individuals.

[0004] In summary, current 3D point cloud-based single-tree identification technologies suffer from general deficiencies in robustness and accuracy when dealing with complex scenarios such as extremely irregular shapes, severe cross-adhesion, and high-density distribution of bamboo and other trees damaged by disasters. This results in low recognition rates for traditional methods, and a tendency to produce false detections such as undersegmentation and oversegmentation, severely limiting the practicality of rapid and accurate investigation and assessment of damaged trees based on point cloud data. Summary of the Invention

[0005] The purpose of this invention is to provide a method for identifying damaged bamboo culms using multi-slice adaptive clustering and cluster splitting and merging. The core innovation lies in effectively solving the common problems of under-segmentation and over-segmentation in traditional methods when dealing with irregularly shaped, high-density interwoven, and damaged, adhered bamboo culms by combining multi-slice hierarchical processing, adaptive clustering, and a refined cluster splitting strategy based on a probabilistic mixture model. Specifically, this invention reduces processing complexity through vertical multi-slice extraction, achieves accurate density clustering using adaptive neighborhood radii, and introduces a probabilistic mixture model to identify and split independent sub-clusters within geometrically anomalous clusters. Finally, by combining cross-slice association and a multi-peak verification mechanism, it achieves high-precision and robust identification of damaged bamboo culms in complex environments.

[0006] The objective of this invention can be achieved through the following technical solutions: A method for identifying damaged bamboo culms using multi-slice adaptive clustering and cluster splitting and merging includes the following steps: Step S1: Extract multiple overlapping horizontal point cloud slices from the 3D point cloud data along the vertical direction to construct multi-layer spatial sampling data; Step S2: For each horizontal point cloud slice, the spatial distance between each point in the point cloud and a group of its nearest neighbors is calculated to form a distance distribution. A preset statistical quantity is extracted from the distance distribution as an adaptive neighborhood radius parameter. Then, density clustering is performed on the slice point cloud to generate an initial point cloud cluster. Step S3: Perform geometric morphology analysis on each initial point cloud cluster. When it is determined that its shape meets the preset stretching or irregular features, use a probabilistic mixture model to fit it to multiple components. Evaluate the model quality before and after splitting according to the Bayesian information criterion. If and only if the evaluation result of the multi-component fitting model is better than that of the single component, split the initial point cloud cluster into multiple independent sub-clusters. Step S4: Based on the spatial proximity and geometric similarity criteria, multiple initial point cloud clusters that are determined to originate from the same physical entity in different horizontal point cloud slices are merged into a single tree candidate object. Step S5: For multiple candidate trees that are spatially adjacent, examine the distribution pattern of their corresponding original point clouds on the horizontal projection plane; if the distribution pattern shows significant multi-peak characteristics, it is determined to be multiple independent trees and merging is suppressed; otherwise, the overlapping candidate trees are finally merged or deduplicated, and the identified trees are output.

[0007] Preferably, in step S1, the extraction of multiple overlapping horizontal point cloud slices along the vertical direction to construct multi-layer spatial sampling data specifically includes: Within a preset range of breast height, multiple consecutive horizontal point cloud slices are extracted with a fixed slice thickness and a fixed step distance. Adjacent horizontal point cloud slices have some overlapping areas in vertical space. Each extracted horizontal point cloud slice is spatially downsampled using a voxel grid filter, and the data points in each voxel grid are averaged into a single representative point to obtain the multi-layer spatial sampling data.

[0008] Preferably, in step S2, the spatial distance between each point in the point cloud and a set of its nearest neighbors is calculated to form a distance distribution, and a preset statistic is extracted from this distance distribution as an adaptive neighborhood radius parameter, specifically: The Euclidean distance between each data point in each horizontal point cloud slice and its Kth nearest neighbor is calculated using a K-dimensional tree data structure. The Euclidean distances of all data points are arranged in ascending order to construct a one-dimensional distance distribution sequence. The distance values ​​at preset percentile positions in the distance distribution sequence are extracted and set as the adaptive neighborhood radius parameter of the horizontal point cloud slice.

[0009] Preferably, in step S2, density clustering is performed on the sliced ​​point cloud to generate an initial point cloud cluster, specifically including: The adaptive neighborhood radius parameter and a set minimum number of core points are input into the density-based spatial clustering algorithm with accompanying noise. The algorithm iterates through all data points in the corresponding horizontal point cloud slice, divides the density-connected data points into the same set, removes outlier noise points that have not been divided, and outputs each independent set as the initial point cloud cluster.

[0010] Preferably, in step S3, performing geometric morphology analysis on each initial point cloud cluster to determine whether its shape satisfies preset stretching or irregular features specifically involves: A covariance matrix is ​​constructed for the coordinates of all three-dimensional points in the initial point cloud cluster, and principal component analysis is performed to extract the first principal component eigenvalue and the second principal component eigenvalue of the covariance matrix. The ratio of the first principal component eigenvalue to the second principal component eigenvalue is calculated. When the ratio is greater than the set aspect ratio threshold, the initial point cloud cluster is determined to satisfy the stretching or irregular feature.

[0011] Preferably, in step S3, the multi-component fitting of the model using a probabilistic mixture model and the evaluation of the model's performance before and after splitting based on the Bayesian information criterion specifically include: Using a Gaussian mixture model as the probabilistic mixture model, the number of components in the Gaussian mixture model is set to one and two, respectively. Two-step maximum likelihood estimation fitting is performed on the initial point cloud clusters that satisfy the stretching or irregular characteristics. The Bayesian information criterion values ​​are calculated for single-component fitting and two-component fitting, respectively. The Bayesian information criterion values ​​are jointly determined by the log-likelihood function value and the model parameter penalty term.

[0012] Preferably, in step S4, the process of merging candidates into a single tree based on spatial proximity and geometric similarity criteria includes determining spatial proximity. The specific process is as follows: Calculate the geometric center coordinates of each initial point cloud cluster and the split independent sub-clusters in different horizontal point cloud slices; traverse the geometric center coordinates of each cluster between adjacent horizontal point cloud slices and calculate the Euclidean distance. When the Euclidean distance is less than the set spatial displacement threshold, the corresponding cluster is determined to satisfy the spatial proximity criterion.

[0013] Preferably, in step S4, the process of merging candidates into a single candidate tree based on spatial proximity and geometric similarity criteria includes determining geometric similarity. The specific process is as follows: Two-dimensional circle fitting is performed on clusters that satisfy the spatial proximity criterion using the least squares method to obtain the estimated cross-sectional radius of each cluster; the absolute value of the difference in cross-sectional radius between corresponding clusters between adjacent horizontal point cloud slices is calculated, and when the absolute value is less than the set radius tolerance threshold, the corresponding cluster is determined to satisfy the geometric similarity criterion and a merging operation is performed.

[0014] Preferably, in step S5, the step of examining the distribution pattern of the corresponding original point cloud on the horizontal projection plane and determining that the distribution pattern exhibits significant multi-peak characteristics specifically involves: The three-dimensional data points contained in multiple spatially adjacent single-tree candidate objects are uniformly projected vertically onto a two-dimensional horizontal plane; a two-dimensional probability density function surface of the projected point set is constructed using the kernel density estimation method; the number of local maxima on the two-dimensional probability density function surface is detected, and when the number of local maxima is greater than one, it is determined that the distribution pattern exhibits significant multimodal characteristics.

[0015] Preferably, in step S5, the final merging or deduplication of overlapping candidate objects to output the identified individual trees specifically includes: For multiple candidate objects that do not exhibit multi-peak characteristics and overlap in space, the bounding box volume and point cloud density of each candidate object are calculated. The candidate object with the highest point cloud density is retained and redundant objects are removed. The final retained single-tree candidate objects are subjected to three-dimensional cylindrical parametric fitting, and the spatial three-dimensional coordinates and diameter at breast height parameters of the corresponding damaged moso bamboo single tree are extracted and output.

[0016] The beneficial effects of this invention are: The method for identifying damaged bamboo stalks provided by this invention can significantly improve the accuracy and robustness of identifying irregularly shaped trees in complex environments. By integrating strategies such as multi-slice processing, adaptive clustering, and probabilistic hybrid model splitting, it effectively addresses extreme cases such as bending, tilting, and multiple intertwined and adhered bamboo stalks, solving the undersegmentation and oversegmentation problems that are prone to occur in traditional methods. This ensures that even under conditions of highly complex forest stand structure, individual damaged bamboo stalks can be accurately distinguished and identified.

[0017] This invention achieves a high degree of automation and intelligence in the identification of damaged individual bamboo trees. Traditional methods often require significant manual intervention or parameter adjustments to adapt to different forest stand conditions, resulting in low efficiency. This method, however, minimizes reliance on prior knowledge and human experience through density clustering with adaptive radius parameters, cluster splitting driven by information criteria, and cross-slice merging based on spatial geometric similarity criteria. This enables large-scale point cloud data processing and damaged forest tree surveys to be conducted more efficiently and conveniently.

[0018] This invention significantly improves the precision and decision support capabilities of forestry resource surveys and disaster assessments. By providing high-precision location and individual information of damaged bamboo stalks, this method offers reliable data support for forestry departments to conduct disaster loss assessments, formulate precise restoration plans, and monitor stand recovery. This helps managers to more quickly and accurately grasp the specific situation of damaged forest land, thereby making more scientific and effective management decisions and enhancing the capacity for sustainable forest development. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the method for identifying damaged bamboo stalks using multi-slice adaptive clustering and cluster splitting and merging, as described in this invention. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0021] Example: Figure 1 As shown, this invention discloses a method for identifying damaged bamboo stalks using multi-slice adaptive clustering and cluster splitting and merging. It aims to address the shortcomings in robustness and accuracy of existing technologies when dealing with complex scenarios such as extremely irregular morphology, severe cross-adhesion, and high-density distribution of bamboo and other trees caused by disasters. This embodiment will elaborate on the core steps and technical details of the invention to clearly demonstrate its innovation and practicality.

[0022] In this embodiment, step S1 includes constructing multi-layer spatial sampling data, and the specific implementation method is as follows: This invention first extracts a series of overlapping horizontal slices from the original 3D point cloud data to reduce the dimensionality and complexity of subsequent processing. The original 3D point cloud data P consists of N discrete points, i.e. Each point Having three-dimensional coordinates .

[0023] Based on the actual needs of forestry surveys and the growth characteristics of moso bamboo, we have determined a preset range for diameter at breast height (DBH). For example, in Chinese forestry standards, DBH is usually measured at 1.3 meters above the ground, but considering that damaged moso bamboo may bend or fall, its effective DBH may vary. Therefore, this invention sets the DBH range to a relatively broad range, for example, from 0.5 meters to 2.5 meters above the ground. This ensures that even if the bamboo is bent or tilted to some extent, the effective identification area near its diameter at breast height can still be fully covered.

[0024] Within this range of breast height, we set a fixed slice thickness. and a fixed step distance Slice thickness This determines the vertical extent of each horizontal slice. Step distance The initial height interval between adjacent slices is determined. To ensure sufficient contextual information between slices for subsequent association and merging, we intentionally allow adjacent horizontal point cloud slices to partially overlap in vertical space. This overlap is [amount missing]. - For example, if Set to 0.1 meters. If the overlap is set to 0.05 meters, the overlap area will be 0.05 meters. This overlap design is crucial for ensuring the continuity and robustness of cross-slice merging, preventing individual bamboo shoots from being incorrectly segmented due to slice boundaries.

[0025] The specific slice extraction process is as follows: from the lowest point of the breast height range... Beginning, with Slide upwards at intervals. For each starting height... Extract a horizontal point cloud slice , which contains all that satisfy conditional point This process generates M consecutive horizontal point cloud slices. .

[0026] To reduce the data volume of each point cloud slice, decrease the computational burden, and smooth local noise, each extracted horizontal point cloud slice... Voxel grid filters are used for spatial downsampling. Voxel grid downsampling is a commonly used point cloud preprocessing technique that divides the point cloud space into a series of small three-dimensional cubic units, i.e., voxels, and then replaces all points within a voxel with a representative point within that voxel.

[0027] We set a voxel grid size. For example, it can be set to 0.02 meters, which is sufficient to capture local details of the bamboo stalk while significantly reducing the number of points. For each slice All its points are assigned to the corresponding voxel grid. For any non-empty voxel grid... The set of points contained within it is denoted as Its representative point The coordinates are obtained by calculating the geometric center of all points within the voxel, i.e., the average coordinates of all points: ,in, Representing a voxel grid Number of interior points. This operation grouped multiple points within each voxel grid into a single representative point, effectively reducing point cloud density and standardizing the data.

[0028] After downsampling, each horizontal point cloud slice Converted into a downsampled slice point cloud Ultimately, we obtain a multi-layered spatially sampled downsampled point cloud dataset. ,in This represents the point cloud of the j-th slice after downsampling.

[0029] In this embodiment, step S2 includes adaptive density clustering, and the specific implementation method is as follows: After obtaining the downsampled multi-layered point cloud data, this step is performed on each horizontal point cloud slice. Density clustering is performed to identify individual bamboo plants or clusters of bamboo. To overcome the sensitivity of the traditional DBSCAN algorithm to a fixed neighborhood radius, this invention introduces an adaptive neighborhood radius calculation method.

[0030] The traditional DBSCAN algorithm requires the user to manually set a global neighborhood radius. This approach often proves ineffective when processing point cloud data with uneven density. Because the point cloud density in affected bamboo forests can vary significantly due to factors such as the curvature and sparseness of the bamboo or the sensor scanning angle, a fixed... Values ​​that can lead to undersegmentation or oversegmentation. To address this issue, the present invention provides a solution for each slice. Independently compute an adaptive neighborhood radius .

[0031] For each downsampled horizontal point cloud slice Where L is the number of points in the slice, we first need to efficiently calculate the distance between each point and its nearest neighbors. The KD-Tree data structure plays a crucial role here, enabling fast nearest neighbor search. Using the KD-Tree, we calculate the slice... Each data point Its first Euclidean distance between the nearest neighbors , It is a preset integer, for example, it can be set to 5. Select the first... The nearest neighbor distance, as a measure of local density, can reflect the density of a region better than using the nearest neighbor distance directly, because it takes into account the local neighborhood of a point.

[0032] The first of all data points The nearest neighbor Euclidean distances constitute a one-dimensional distance distribution sequence. To eliminate the influence of outliers and capture the distribution characteristics of most points, we use distance distribution sequences. Extracting values ​​at the preset percentile The distance value of the location is used as the slice. Adaptive neighborhood radius parameter For example, it can be It is set to 90%. This means that 90% of the points, their first... The nearest neighbor distance is less than or equal to , ,in, Represents the computation sequence of Percentiles. This adaptive method ensures... It can be adjusted according to the actual point cloud density distribution of each slice, which improves the robustness of clustering.

[0033] Get each slice Adaptive neighborhood radius parameter Then, along with a preset minimum number of core points. The input is fed into the density-based spatial clustering application with accompanying noise algorithm DBSCAN. It is typically set to a small integer. It defines a minimum threshold of the number of points around a core point to form a cluster.

[0034] The working principle of the DBSCAN algorithm is as follows: Randomly select an unvisited point: the algorithm starts from... Randomly select a point that is not marked as visited; check the core point condition: if the point... The neighborhood contains at least If there are 1 point, then that point is marked as a core point; Cluster expansion: If it is a core point, then create a new cluster and apply it to all its nodes. Points within the neighborhood are added to the cluster. Then, steps 2 and 3 are recursively performed on all points within its neighborhood until the cluster can no longer expand; Boundary points and noise points: if a point The number of points in the neighborhood is less than But it is located at a certain core point If a point is within the neighborhood, it is a boundary point. If a point is neither a core point nor a boundary point, it is marked as a noise point and discarded; Repeat: Repeat the above process until all points have been visited.

[0035] Each slice It was divided into a series of independent initial point cloud clusters, forming an ensemble. ,in It is the k-th initial point cloud cluster of the j-th slice. These clusters represent potential bamboo individuals or point cloud clumps formed due to adhesion at this slice level.

[0036] In this embodiment, step S3 includes refined cluster splitting based on a probabilistic mixture model, specifically implemented as follows: The initial point cloud clusters obtained in step S2 may contain "composite clusters" formed by multiple bamboo trees being closely connected in space, which can lead to undersegmentation problems. The core objective of this step is to identify and separate these composite clusters to achieve more refined individual tree identification.

[0037] For each initial point cloud cluster First, we need to determine whether it has "stretched or irregular features," which are usually characteristics of multiple bamboo individuals sticking together. We then examine the clusters... Principal component analysis (PCA) is performed on the three-dimensional coordinates of all points. PCA is a statistical method for dimensionality reduction of data. It projects the observed data onto a new set of orthogonal bases through orthogonal transformation. These bases are given by the eigenvectors of the data's covariance matrix. The magnitude of the eigenvalues ​​reflects the variance of the data along the corresponding eigenvector directions.

[0038] The specific steps are as follows: Constructing the covariance matrix: Calculating the clusters The covariance matrix of all points around their centroid ; Eigenvalue decomposition: for Eigenvalue decomposition yields three eigenvalues. And the corresponding eigenvectors. We sort the eigenvalues ​​in descending order, i.e. , The most prominent direction of the corresponding data distribution Corresponding to the secondary direction, Corresponding to the least significant direction; calculating the aspect ratio: This invention uses the ratio of the eigenvalues ​​of the first two principal components to measure the degree of cluster "stretching". Calculating the ratio The larger this ratio, the longer the data is distributed in one direction and the narrower it is in the vertical direction, exhibiting a clear stretched shape; Characteristic determination: When Greater than the set aspect ratio threshold At that time, it is determined that the initial point cloud cluster satisfies the stretched or irregular characteristics. For example, it can be set to... The value is 10. This means that the primary axis of the cluster is much longer than the secondary axis, and it is likely not a single cylindrical cross-section, but rather composed of cross-sections of two or more bamboo plants bonded together. Only clusters that meet this condition will be fed into the subsequent probabilistic mixture model for fitting, in order to save computational resources and improve efficiency.

[0039] For clusters that are determined to satisfy the stretching or irregularity characteristics This invention employs a probabilistic mixture model for multi-component fitting. We use the Gaussian Mixture Model (GMM) as the probabilistic mixture model. The GMM assumes that the data points are generated by a mixture of a finite number of Gaussian distributions, each representing a "component" or subgroup.

[0040] The fitting process typically employs the Expectation-Maximization (EM) algorithm. The EM algorithm is an iterative optimization method used to estimate probabilistic model parameters in the presence of latent variables. In a Gaussian Model (GMM), the latent variables are the Gaussian components to which each data point belongs.

[0041] We are clusters Perform GMM fitting for single-component M=1 and two-component M=2 respectively: Single-component fitting: Hypothetical cluster All points in the dataset follow a single Gaussian distribution. We estimate the mean of this Gaussian distribution. Covariance Matrix After fitting, the maximum log-likelihood function value of the model with respect to the data can be calculated. , ,in, The mean is The covariance matrix is The Gaussian probability density function represents the probability density of point p under this Gaussian distribution; bicomponent fitting: assuming cluster... The points in the dataset are derived from a mixture of two Gaussian distributions. We need to estimate the means of each of the two Gaussian distributions. Covariance Matrix and their mixing coefficients ,in The mixing coefficient represents the prior probability of which Gaussian component the data point originates from. After fitting, the maximum log-likelihood function value of the model with respect to the data is calculated. , To evaluate whether one-component or two-component fitting better interprets the data, we use the Bayesian Information Criterion (BIC). BIC is a commonly used model selection criterion that penalizes model complexity while evaluating goodness of fit. A smaller BIC value indicates a better balance between fitting the data and model complexity. Where L is the maximum log-likelihood function value of the fitted model, and K is the number of model parameters. For a two-dimensional Gaussian distribution, a component contains a mean vector (e.g., 2 parameters), a covariance matrix (e.g., 3 non-redundant parameters), and a mixing coefficient (e.g., 1 parameter), for a total of 2 + 3 + 1 = 6 parameters. However, since the sum of the mixing coefficients is 1, M components have M−1 independent mixing coefficients. Therefore, a single-component model... One parameter. Two-component model. One parameter, It is a cluster The number of midpoints.

[0042] We calculate the one-component fitting separately. and bicomponent fitting .Compare and : If and only if This indicates that the model evaluation results of the two-component fitting are significantly better than those of the one-component fitting, meaning that using two Gaussian distributions to describe the data is more reasonable than using one Gaussian distribution. In this case, we determine the initial point cloud cluster... It is indeed composed of two independent subgroups, and can be further divided into two independent subclusters. and Splitting typically involves assigning points to corresponding subclusters based on the posterior probability of each point belonging to each Gaussian component, obtained from the EM algorithm.

[0043] if If the result is negative, it indicates that the advantage of two-component fitting is not significant, or that introducing more parameters actually leads to a decrease in model performance. In this case, we consider the cluster to be essentially a single entity and do not split it; the original cluster remains unchanged. As a refined cluster.

[0044] These split or unsplit clusters constitute the refined point cloud cluster set. This step effectively solved the problem of undersegmentation caused by bamboo sticking together, providing more accurate individual units for subsequent single-tree identification.

[0045] In this embodiment, step S4 includes merging cross-slice clusters into single-tree candidate objects, and the specific implementation is as follows: The preceding steps identified and refined point cloud clusters within each horizontal slice. The goal of this step is to vertically correlate and merge multiple refined point cloud clusters from different horizontal slices that are determined to originate from the same physical entity, forming preliminary candidate single-tree objects. This is based on two core criteria: spatial proximity and geometric similarity.

[0046] For two adjacent horizontal point cloud slices and and its corresponding set of refined point cloud clusters and We need to determine which clusters are continuous in the vertical direction, that is, they may belong to the same bamboo.

[0047] We calculate Each cluster in and Each cluster in The geometric center coordinates. It is the mean of the coordinates of all points within the cluster. For example, for the cluster... Its geometric center The calculation is as follows: ,in It is a cluster The number of points in the middle.

[0048] We iterate through all possible cluster pairs Calculate the Euclidean distance between their geometric centers. This distance quantifies how close two clusters are in three-dimensional space. When the Euclidean distance is less than the set spatial displacement threshold When the corresponding cluster satisfies the spatial proximity criterion, it is determined that the cluster satisfies the spatial proximity criterion. This is a key parameter that sets the maximum lateral and longitudinal offset between two clusters in the vertical direction that can be considered part of the same bamboo stalk. For example, it can be set to 0.15 meters, which allows for a certain degree of bending and tilting of the bamboo.

[0049] Spatial proximity alone is insufficient for accurate cluster association, as multiple adjacent bamboo trees may also be spatially close. Therefore, we introduce a geometric similarity criterion, particularly the similarity of cross-sectional shapes, for further screening. We utilize this geometric characteristic, considering that bamboo trunks are typically approximately circular in horizontal cross-section.

[0050] For cluster pairs that satisfy the spatial proximity criterion We perform a two-dimensional circle fitting on the points of each cluster on the horizontal plane. This can be achieved using the least squares method to find an optimal center and radius that minimizes the sum of the squared distances from all points within the cluster to the circumference of that circle. Through circle fitting, we can obtain an estimate of the cross-sectional radius of each cluster. and .

[0051] Then, we calculate the absolute value of the difference in cross-sectional radius between corresponding clusters of adjacent horizontal point cloud slices. If this difference is small, it indicates that their cross-sectional sizes are very similar, further supporting the hypothesis that they belong to the same bamboo plant. When the absolute value is less than a set radius tolerance threshold... When the corresponding cluster satisfies the geometric similarity criterion, it is determined that the cluster satisfies the geometric similarity criterion. This reflects our tolerance for variations in the thickness of bamboo stalks in the vertical direction. For example, it could be set to 0.03 meters.

[0052] Only cluster pairs that simultaneously satisfy both spatial proximity and geometric similarity criteria will be merged. The merging operation involves combining the points of these clusters, recording their hierarchical association information, and ultimately forming a single candidate tree. Each single-tree candidate This represents a cluster of point clouds that are continuous and geometrically similar across multiple slices, thus forming a potential individual bamboo.

[0053] In this embodiment, step S5 includes multi-peak verification and deduplication output, and the specific implementation method is as follows: In step S4, although we have performed meticulous merging, the complexity of the affected bamboo forest may still lead to some mismerging. For example, two very closely spaced or even intertwined bamboo trees may exhibit similar geometric features on multiple slices and be mistakenly identified as the same tree. This step is key to solving this residual mismerging problem in this invention, by examining the distribution pattern of candidate objects on the horizontal projection to finally confirm the individual tree.

[0054] For multiple candidate single trees that are spatially adjacent First, we need to check if they have multiple "cores" on the horizontal plane. To do this, we uniformly project all the original 3D data points contained in these candidate objects, rather than downsampled or clustered points, onto the 2D horizontal plane. The original point cloud is used here to preserve the finest spatial distribution information.

[0055] Let the original set of points associated with these candidate objects be . Its horizontal projection point set is Next, we use the kernel density estimation (KDE) method to construct the projection point set. Two-dimensional probability density function surface KDE is a nonparametric method for estimating the probability density function of a random variable. It works by placing a kernel function, such as a Gaussian kernel, at each data point and then summing these kernel functions to obtain a smooth density estimate. The general form of KDE is: ,in, It is the number of projection points; These are bandwidth parameters, which control the smoothness of the kernel function and the extent of its local influence. The choice of bandwidth has a significant impact on KDE results and is typically optimized through cross-validation or other adaptive methods. This refers to the kernel function. This invention typically uses the Gaussian kernel function because of its good smoothness and mathematical properties. The form of the two-dimensional Gaussian kernel function is: Through KDE, we obtained a continuous surface that reflects the distribution density of the original point cloud on the horizontal plane.

[0056] Obtaining the two-dimensional probability density function surface Then, we detect the number of local maxima on the surface. Local maxima correspond to density peaks on the horizontal projection of the point cloud, and each peak may represent an individual bamboo stalk.

[0057] if If the distribution pattern exhibits a significant multimodal characteristic, it is determined that the previously merged candidate objects actually contain multiple independent bamboo individuals. In this case, the present invention will suppress any previous merging. Specifically, based on the location of local maxima, overlapping candidate objects are split into multiple independent individual bamboos. These point clouds can be separated into independent individual bamboos using the valley values ​​of the density surface through watershed algorithms or other cluster separation techniques. This mechanism is key to preventing over-segmentation in the present invention.

[0058] Conversely, if If the surface exhibits a single peak or no obvious peak, it indicates that these candidate objects likely belong to the same bamboo stalk, or their overlap is insufficient to form multiple distinguishable peaks. In this case, spatially overlapping candidate objects are ultimately merged or deduplicated. For multiple candidate objects that do not exhibit multi-peak characteristics but overlap spatially, such as having multiple bounding boxes at the same location, this may be due to noise or redundancy within the algorithm. We calculate the bounding box volume of each candidate object. With point cloud density To ensure that the identified individual trees are solid entities and as complete as possible, the candidate objects with the highest point cloud density are typically retained, while other redundant objects with overlapping and lower density are discarded. Candidate objects with higher point cloud density usually contain more information about the main body of the bamboo, while low-density objects may just be edge points or noise.

[0059] For each remaining candidate bamboo stalk, a 3D cylindrical parametric fitting is performed. The geometry of a bamboo stalk approximates a cylinder, making cylindrical fitting an effective method for extracting its geometric parameters. The fitting process is typically implemented using iterative optimization algorithms, such as RANSAC or least squares, to find an optimal cylindrical model whose centerline, radius, and height best fit all points in the candidate bamboo stalk, minimizing the sum of the squared distances from each point to the cylinder surface.

[0060] The spatial three-dimensional coordinates and diameter at breast height (DBH) parameters of the corresponding damaged bamboo culms are extracted and output from the fitting results. The spatial three-dimensional coordinates can typically be defined as the coordinates of the bottom center or the centroid of the fitted cylinder, providing the precise location of the bamboo in three-dimensional space. The DBH parameter is obtained by multiplying the radius of the fitted cylinder by 2, a key indicator in forestry surveys. These high-precision geometric parameters provide reliable data support for forestry resource surveys, disaster loss assessment, the development of precise restoration plans, and the monitoring of stand recovery.

[0061] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0062] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for identifying damaged bamboo culms using multi-slice adaptive clustering and cluster splitting and merging, characterized in that, Includes the following steps: Step S1: Extract multiple overlapping horizontal point cloud slices from the 3D point cloud data along the vertical direction to construct multi-layer spatial sampling data; Step S2: For each horizontal point cloud slice, the spatial distance between each point in the point cloud and a group of its nearest neighbors is calculated to form a distance distribution. A preset statistical quantity is extracted from the distance distribution as an adaptive neighborhood radius parameter. Then, density clustering is performed on the slice point cloud to generate an initial point cloud cluster. Step S3: Perform geometric morphology analysis on each initial point cloud cluster. When it is determined that its shape meets the preset stretching or irregular features, use a probabilistic mixture model to fit it to multiple components. Evaluate the model quality before and after splitting according to the Bayesian information criterion. If and only if the evaluation result of the multi-component fitting model is better than that of the single component, split the initial point cloud cluster into multiple independent sub-clusters. Step S4: Based on the spatial proximity and geometric similarity criteria, multiple initial point cloud clusters that are determined to originate from the same physical entity in different horizontal point cloud slices are merged into a single tree candidate object. Step S5: For multiple candidate trees that are spatially adjacent, examine the distribution pattern of their corresponding original point clouds on the horizontal projection plane; if the distribution pattern shows significant multi-peak characteristics, it is determined to be multiple independent trees and merging is suppressed; otherwise, the overlapping candidate trees are finally merged or deduplicated, and the identified trees are output.

2. The method for identifying damaged bamboo stalks by multi-slice adaptive clustering and cluster splitting and merging according to claim 1, characterized in that, In step S1, the extraction of multiple overlapping horizontal point cloud slices along the vertical direction to construct multi-layer spatial sampling data specifically includes: Within a preset range of breast height, multiple consecutive horizontal point cloud slices are extracted with a fixed slice thickness and a fixed step distance. Adjacent horizontal point cloud slices have some overlapping areas in vertical space. Each extracted horizontal point cloud slice is spatially downsampled using a voxel grid filter, and the data points in each voxel grid are averaged into a single representative point to obtain the multi-layer spatial sampling data.

3. The method for identifying damaged bamboo stalks by multi-slice adaptive clustering and cluster splitting and merging according to claim 1, characterized in that, In step S2, the spatial distance between each point in the point cloud and a set of its nearest neighbors is calculated to form a distance distribution, and a preset statistic is extracted from this distance distribution as an adaptive neighborhood radius parameter, specifically: The Euclidean distance between each data point in each horizontal point cloud slice and its Kth nearest neighbor is calculated using a K-dimensional tree data structure. The Euclidean distances of all data points are arranged in ascending order to construct a one-dimensional distance distribution sequence. The distance values ​​at preset percentile positions in the distance distribution sequence are extracted and set as the adaptive neighborhood radius parameter of the horizontal point cloud slice.

4. The method for identifying damaged bamboo stalks by multi-slice adaptive clustering and cluster splitting and merging according to claim 1, characterized in that, In step S2, density clustering is performed on the sliced ​​point cloud to generate an initial point cloud cluster, specifically including: The adaptive neighborhood radius parameter and a set minimum number of core points are input into the density-based spatial clustering algorithm with accompanying noise. The algorithm iterates through all data points in the corresponding horizontal point cloud slice, divides the density-connected data points into the same set, removes outlier noise points that have not been divided, and outputs each independent set as the initial point cloud cluster.

5. The method for identifying damaged bamboo stalks by multi-slice adaptive clustering and cluster splitting and merging according to claim 1, characterized in that, In step S3, the geometric morphology analysis of each initial point cloud cluster is performed to determine whether its shape satisfies preset stretching or irregular features, specifically as follows: A covariance matrix is ​​constructed for the coordinates of all three-dimensional points in the initial point cloud cluster, and principal component analysis is performed to extract the first principal component eigenvalue and the second principal component eigenvalue of the covariance matrix. The ratio of the first principal component eigenvalue to the second principal component eigenvalue is calculated. When the ratio is greater than the set aspect ratio threshold, the initial point cloud cluster is determined to satisfy the stretching or irregular feature.

6. The method for identifying damaged bamboo stalks by multi-slice adaptive clustering and cluster splitting and merging according to claim 1, characterized in that, In step S3, the multi-component fitting of the probabilistic mixture model and the evaluation of the model's performance before and after the split based on the Bayesian information criterion specifically include: Using a Gaussian mixture model as the probabilistic mixture model, the number of components in the Gaussian mixture model is set to one and two, respectively. Two-step maximum likelihood estimation fitting is performed on the initial point cloud clusters that satisfy the stretching or irregular characteristics. The Bayesian information criterion values ​​are calculated for single-component fitting and two-component fitting, respectively. The Bayesian information criterion values ​​are jointly determined by the log-likelihood function value and the model parameter penalty term.

7. The method for identifying damaged bamboo stalks by multi-slice adaptive clustering and cluster splitting and merging according to claim 1, characterized in that, In step S4, the process of merging candidates into a single tree based on spatial proximity and geometric similarity criteria includes determining spatial proximity. The specific process is as follows: Calculate the geometric center coordinates of each initial point cloud cluster and the split independent sub-clusters in different horizontal point cloud slices; traverse the geometric center coordinates of each cluster between adjacent horizontal point cloud slices and calculate the Euclidean distance. When the Euclidean distance is less than the set spatial displacement threshold, the corresponding cluster is determined to satisfy the spatial proximity criterion.

8. The method for identifying damaged bamboo stalks by multi-slice adaptive clustering and cluster splitting and merging according to claim 1, characterized in that, In step S4, the process of merging candidates into a single tree based on spatial proximity and geometric similarity criteria includes determining geometric similarity. The specific process is as follows: Two-dimensional circle fitting is performed on clusters that satisfy the spatial proximity criterion using the least squares method to obtain the estimated cross-sectional radius of each cluster. Calculate the absolute value of the difference in cross-sectional radius between corresponding clusters of adjacent horizontal point cloud slices. When the absolute value is less than the set radius tolerance threshold, determine that the corresponding clusters meet the geometric similarity criterion and perform a merging operation.

9. The method for identifying damaged bamboo stalks by multi-slice adaptive clustering and cluster splitting and merging according to claim 1, characterized in that, In step S5, the step of examining the distribution pattern of the corresponding original point cloud on the horizontal projection plane and determining that the distribution pattern exhibits significant multi-peak characteristics specifically involves: The three-dimensional data points contained in multiple spatially adjacent single-tree candidate objects are uniformly projected vertically onto a two-dimensional horizontal plane; a two-dimensional probability density function surface of the projected point set is constructed using the kernel density estimation method; the number of local maxima on the two-dimensional probability density function surface is detected, and when the number of local maxima is greater than one, it is determined that the distribution pattern exhibits significant multimodal characteristics.

10. The method for identifying damaged bamboo culms by multi-slice adaptive clustering and cluster splitting and merging according to claim 1, characterized in that, In step S5, the final merging or deduplication of overlapping candidate objects, and the output of the identified individual trees, specifically includes: For multiple candidate objects that do not exhibit multi-peak characteristics and overlap in space, the bounding box volume and point cloud density of each candidate object are calculated. The candidate object with the highest point cloud density is retained and redundant objects are removed. The final retained single-tree candidate objects are subjected to three-dimensional cylindrical parametric fitting, and the spatial three-dimensional coordinates and diameter at breast height parameters of the corresponding damaged moso bamboo single tree are extracted and output.

Images