Tree point cloud instance segmentation method and system

Abstract

The application discloses a tree point cloud instance segmentation method and system, wherein the method comprises the following steps: performing feature extraction and multi-anchor point prediction on tree point cloud data to obtain semantic classification results and multi-anchor point offset information; extracting a tree point cloud subset according to the semantic classification results, and performing height adaptive processing on the tree point cloud subset to construct a two-dimensional horizontal cross-section projection image; performing trunk base detection and coordinate inverse transformation on the two-dimensional horizontal cross-section projection image to obtain three-dimensional base position prior information of a single tree instance; performing fusion processing on offset information of a plurality of preset height level anchors corresponding to each sampling point to obtain a fused instance center prediction value; and constructing a joint seed point set based on the three-dimensional base position prior information and the fused instance center prediction value, and sequentially performing joint clustering and instance allocation on the joint seed point set to obtain an instance segmentation result of the tree point cloud data. The method improves the accuracy of single tree segmentation in a complex forest scene.

Description

Technical Field

[0001] This invention relates to the fields of intelligent vegetation monitoring and computer vision technology, and in particular to a method and system for segmenting tree point cloud instances. Background Technology

[0002] Individual tree instance segmentation is a key step in the accurate survey of forest resources and the inversion of ecological parameters. Its core objective is to accurately separate each individual tree from complex 3D laser point cloud data.

[0003] In existing technologies, instance differentiation often relies on single centroid regression and local feature clustering. For example, existing point cloud instance segmentation networks typically predict a single offset vector pointing each point to the geometric center of its respective instance, and use this as the basis for clustering and instance assignment.

[0004] However, the morphological differences between the base and crown of trees are significant. A single center point is difficult to simultaneously represent the clear separation characteristics of the trunk base and the complex interwoven structure of the crown top. This can easily lead to problems such as adjacent trees sticking together, misattribution of the top of tall trees, or missed detection of weak individuals in dense forest areas due to center prediction drift or confidence decay. This seriously restricts the reliability and generalization ability of the segmentation results in real forest scenarios. Summary of the Invention

[0005] This invention provides a method and system for segmenting tree point cloud instances, thereby improving the accuracy of single-tree segmentation in complex forest scenes.

[0006] To address the aforementioned technical problems, this invention provides a method and system for segmenting tree point cloud instances, comprising: Tree point cloud data of the area to be processed is acquired, and feature extraction and multi-anchor point prediction are performed on the tree point cloud data respectively to obtain the semantic classification result and multi-anchor point offset information of each sampling point; wherein, the multi-anchor point offset information is configured to characterize the spatial relationship of each sampling point relative to multiple preset height level anchor points of its respective single tree instance through weight parameters. Based on the semantic classification results, a subset of tree point clouds is extracted from the tree point cloud data, and the tree point cloud subset is subjected to height adaptive processing to construct a two-dimensional horizontal cross-sectional projection image. Based on the preset visual basic model and text prompt conditions, the tree trunk base is detected in the two-dimensional horizontal cross-section projection image to obtain the two-dimensional tree trunk base mask and instance count prior. Perform an inverse coordinate transformation on the two-dimensional tree trunk base mask to obtain three-dimensional tree trunk base candidate points, and assign initial instance identifiers to the three-dimensional tree trunk base candidate points based on the instance counting prior to obtain the three-dimensional base position prior information of a single tree instance; Based on the weight parameters, the offset information of multiple preset height level anchor points corresponding to each sampling point is fused to obtain the fused instance center prediction value of each sampling point. Based on the prior information of the three-dimensional base location and the predicted value of the fused instance center, a joint seed point set is constructed, and joint clustering and instance assignment are performed on the joint seed point set in sequence to obtain the instance segmentation result of the tree point cloud data.

[0007] As one preferred embodiment, the step of performing feature extraction and multi-anchor point prediction on the tree point cloud data to obtain the semantic classification result and multi-anchor point offset information for each sampling point includes: The tree point cloud data is input into a pre-trained point cloud feature extraction network to obtain the high-dimensional geometric features of each sampling point; Based on the high-dimensional geometric features, the semantic segmentation head outputs the probability distribution of each sampling point belonging to a preset forest component category, which serves as the semantic classification result for each sampling point. Based on the high-dimensional geometric features, the spatial offset vector of each sampling point relative to its respective single tree instance is predicted by a multi-anchor point regression head at multiple preset height levels; wherein, the multiple height levels are distributed at intervals along the vertical direction, and the spatial offset vector represents the three-dimensional coordinate difference between the sampling point and the corresponding height level anchor point. Based on learnable weight parameters, the spatial offset vectors of each height level are weighted and aggregated to generate multi-anchor offset information for each sampling point.

[0008] As one preferred embodiment, the step of extracting a subset of tree point clouds from the tree point cloud data based on the semantic classification result, and performing height adaptive processing on the subset of tree point clouds to construct a two-dimensional horizontal cross-sectional projection image, includes: The sampling points in the semantic classification results that are classified as "tree trunk" are selected to form an initial tree trunk point cloud set; Based on the vertical distribution characteristics of the initial trunk point cloud set, determine the target height slice range that includes the trunk base features; The tree trunk sampling points located within the target height slice range are projected onto a horizontal plane, and a density value mapping matrix is ​​generated through rasterization. Normalization and size standardization operations are performed sequentially on the density value mapping matrix to obtain a two-dimensional horizontal cross-sectional projection image.

[0009] As one preferred embodiment, determining the target height slice range containing the trunk base features based on the vertical distribution characteristics of the initial trunk point cloud set includes: A density analysis was performed on the initial tree trunk point cloud set in the vertical direction to obtain the distribution curve of point cloud density as a function of height. Based on the distribution curve, identify the density abrupt change interval or peak interval corresponding to the base of the tree trunk; The height range corresponding to the density abrupt change interval or peak interval is determined as the target height slice range.

[0010] As one preferred embodiment, the step of detecting the tree trunk base of the two-dimensional horizontal cross-sectional projection image based on a preset visual model and text prompt conditions to obtain a two-dimensional tree trunk base mask and instance count prior includes: Construct a sequence of text prompt words containing semantic descriptions of the tree trunk base as text prompt conditions; The two-dimensional horizontal cross-sectional projection image and the text prompt conditions are jointly input into the visual basic model, and the pixel-level prediction result that semantically matches the text prompt word sequence is output. Extract the regions with confidence scores higher than a preset threshold from the pixel-level prediction results and use them as the two-dimensional trunk base mask; Connectivity analysis is performed on the two-dimensional trunk base mask, and the number of connected components obtained is used as the prior for instance counting.

[0011] As one preferred embodiment, the inverse coordinate transformation is performed on the two-dimensional trunk base mask to obtain three-dimensional trunk base candidate points, and initial instance identifiers are assigned to the three-dimensional trunk base candidate points based on the instance counting prior, thereby obtaining the three-dimensional base position prior information of a single tree instance, including: Establish a mapping relationship between the pixel coordinate system of the two-dimensional horizontal cross-section projection image and the three-dimensional world coordinate system of the tree point cloud data; Based on the mapping relationship, the effective pixels within the two-dimensional trunk base mask are back-projected into three-dimensional space, and the Z-axis coordinates are recovered according to the target height slice range to obtain three-dimensional trunk base candidate points. Based on the number of instances counted prior indications, connectivity verification is performed on the candidate points for the base of the three-dimensional trunk to obtain the target three-dimensional trunk base candidate points; A unique instance label is assigned to each cluster of target 3D trunk base candidate points, and the set of spatial coordinates associated with the instance label is used as prior information for the 3D base position.

[0012] As one preferred embodiment, the step of fusing the offset information of multiple preset height-level anchor points corresponding to each sampling point according to the weight parameters to obtain the fused instance center prediction value of each sampling point includes: Obtain the predicted spatial offset vector and corresponding weight parameters at each preset height level of each sampling point; Based on the weight parameters, the spatial offset vectors of each height level are weighted and summed to obtain the comprehensive offset vector. The composite offset vector is superimposed onto the original three-dimensional coordinates of the sampling point to calculate the predicted position of the sampling point pointing to the geometric center of its respective single-tree instance, which is used as the predicted value of the fused instance center.

[0013] As one preferred embodiment, the step of constructing a joint seed point set based on the prior information of the three-dimensional base location and the predicted value of the fused instance center, and sequentially performing joint clustering and instance assignment on the joint seed point set to obtain the instance segmentation result of the tree point cloud data includes: The candidate points of the 3D trunk base with initial instance identification and the sampling points whose confidence scores of the fused instance center prediction values ​​meet the preset threshold conditions are integrated to obtain a joint seed point set. Starting from the seed points in the joint seed point set, region growing clustering is performed on the tree point cloud data based on geometric proximity constraints and feature similarity measures to obtain preliminary instance clusters; Calculate the spatial distance and semantic consistency score between the remaining sampling points and each of the preliminary instance clusters, and assign the remaining sampling points to the preliminary instance cluster with the highest overall score; wherein, the remaining sampling points include remaining sampling points that do not belong to any of the preliminary instance clusters. Boundary smoothing and outlier filtering are performed on each initially assigned instance cluster, and the single-tree instance segmentation mask and the corresponding instance identifier are output as the instance segmentation result of the tree point cloud data.

[0014] Another aspect of the present invention provides a tree point cloud instance segmentation system, comprising: The acquisition module is used to acquire tree point cloud data of the area to be processed, and to perform feature extraction and multi-anchor point prediction on the tree point cloud data to obtain the semantic classification result and multi-anchor point offset information of each sampling point; wherein, the multi-anchor point offset information is configured to characterize the spatial relationship of each sampling point relative to multiple preset height level anchor points of its respective single tree instance through weight parameters. The extraction module is used to extract a subset of tree point clouds from the tree point cloud data based on the semantic classification results, and to perform height adaptive processing on the subset of tree point clouds to construct a two-dimensional horizontal cross-sectional projection image. The detection module is used to detect the base of the tree trunk in the two-dimensional horizontal cross-section projection image based on a preset visual basic model and text prompt conditions, and to obtain a two-dimensional tree trunk base mask and instance count prior. The transformation module is used to perform inverse coordinate transformation on the two-dimensional tree trunk base mask to obtain three-dimensional tree trunk base candidate points, and assign initial instance identifiers to the three-dimensional tree trunk base candidate points based on the instance counting prior to obtain the three-dimensional base position prior information of a single tree instance. The fusion module is used to fuse the offset information of multiple preset height level anchor points corresponding to each sampling point according to the weight parameters, so as to obtain the fusion instance center prediction value of each sampling point. The segmentation module is used to construct a joint seed point set based on the prior information of the three-dimensional base position and the predicted value of the fused instance center, and to perform joint clustering and instance assignment on the joint seed point set in sequence to obtain the instance segmentation result of the tree point cloud data.

[0015] As one preferred embodiment, the acquisition module is specifically used for: The tree point cloud data is input into a pre-trained point cloud feature extraction network to obtain the high-dimensional geometric features of each sampling point; Based on the high-dimensional geometric features, the semantic segmentation head outputs the probability distribution of each sampling point belonging to a preset forest component category, which serves as the semantic classification result for each sampling point. Based on the high-dimensional geometric features, the spatial offset vector of each sampling point relative to its respective single tree instance is predicted by a multi-anchor point regression head at multiple preset height levels; wherein, the multiple height levels are distributed at intervals along the vertical direction, and the spatial offset vector represents the three-dimensional coordinate difference between the sampling point and the corresponding height level anchor point. Based on learnable weight parameters, the spatial offset vectors of each height level are weighted and aggregated to generate multi-anchor offset information for each sampling point.

[0016] Compared with the prior art, the beneficial effects of the present invention are at least one of the following: 1) This invention effectively overcomes the shortcomings of traditional single-center regression methods in accurately locating trees in complex forest structures by introducing a multi-anchor prediction mechanism and a weight fusion strategy. Multiple anchor points at preset height levels represent the spatial relationship between sampling points and individual tree instances, and the offset information is adaptively fused using weight parameters, fully capturing the continuous structural features of trees along the vertical direction. This design not only enhances the model's robustness in center prediction for tall, tilted, or irregularly shaped individuals but also significantly reduces the risk of instance sticking and over-segmentation caused by foliage occlusion or uneven point cloud density, thereby improving overall segmentation accuracy.

[0017] 2) This invention innovatively introduces a visual foundation model into the 3D point cloud instance segmentation process. By constructing a 2D horizontal cross-sectional image through height adaptive projection and combining it with textual prompts for tree trunk base detection, it successfully achieves effective transfer of cross-modal prior knowledge. This strategy leverages the powerful zero-shot perception capability of the visual foundation model to obtain accurate 2D masks and instance count priors, and then generates reliable 3D base position priors through inverse coordinate transformation, providing a high-quality seed point set for subsequent joint clustering. This not only significantly reduces the dependence on large-scale 3D labeled data but also significantly enhances the algorithm's generalization ability and instance discrimination under dense forest areas and complex terrain conditions, balancing segmentation efficiency and accuracy. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating a tree point cloud instance segmentation method in one embodiment of the present invention. Figure 2 This is a structural block diagram of a tree point cloud instance segmentation system in one embodiment of the present invention; Figure label: The module consists of: 11. Acquisition module; 12. Extraction module; 13. Detection module; 14. Conversion module; 15. Fusion module; and 16. Segmentation module. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive. 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.

[0020] In the description of this invention, the terms "first," "second," "third," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first," "second," "third," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0021] In the description of this invention, it should be noted that, unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this specification is for the purpose of describing specific embodiments only and is not intended to limit the invention. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0022] One embodiment of the present invention provides a method for segmenting tree point cloud instances. For details, please refer to [link to relevant documentation]. Figure 1 , Figure 1 The diagram shown is a flowchart of a tree point cloud instance segmentation method according to one embodiment of the present invention, which includes steps S1-S6: S1: Obtain tree point cloud data of the area to be processed, and perform feature extraction and multi-anchor prediction on the tree point cloud data respectively to obtain the semantic classification result and multi-anchor offset information of each sampling point; wherein, the multi-anchor offset information is configured to characterize the spatial relationship of each sampling point relative to multiple preset height level anchor points of its respective single tree instance through weight parameters.

[0023] In the application of 3D digitization of forest resources and precision forestry management, high-precision semantic understanding and instance-level structural parsing of tree point cloud data are prerequisites for the automated extraction of individual tree parameters. However, due to the complex morphology of natural trees, the interlacing of branches and leaves, and the sparse and uneven distribution of point clouds, traditional methods struggle to simultaneously consider both local geometric feature representation and global vertical structural correlation.

[0024] To address this, this embodiment proposes a point cloud feature extraction and prediction mechanism that integrates multiple anchor point spatial constraints. The aim is to enhance the model's perception of the overall tree structure by explicitly modeling the spatial relationship between sampling points and single tree instances at multiple height levels, thereby providing an intermediate representation with both semantic accuracy and structural consistency for subsequent instance segmentation.

[0025] In this embodiment, after acquiring the tree point cloud data of the area to be processed, it is first input into a pre-trained point cloud feature extraction network to obtain the high-dimensional geometric features of each sampling point. The point cloud feature extraction network can adopt a deep learning architecture suitable for unordered point clouds, such as PointNet++, DGCNN, RandLA-Net, or Stratified Transformer, and this invention is not limited thereto.

[0026] Preferably, the network is pre-trained on a large-scale forest point cloud dataset, which can effectively capture low-level structural information such as geometric topology, curvature changes, and density distribution within the local neighborhood, and encode them into high-dimensional feature vectors of 128 or 256 dimensions. This high-dimensional geometric feature not only preserves the spatial coordinate information of the original points, but also integrates contextual semantic cues, providing a shared feature basis for subsequent multi-task heads.

[0027] Based on the aforementioned high-dimensional geometric features, the semantic segmentation head outputs a probability distribution representing the category of each sampling point belonging to a preset forest component, serving as the semantic classification result for each sampling point. The preset forest component categories include, but are not limited to, tree trunks, branches, leaves, and ground vegetation; the specific number and definition of categories can be flexibly configured according to the actual application scenario. The semantic segmentation head typically consists of several fully connected layers and a Softmax activation function, which independently performs classification prediction for each sampling point, outputting normalized probability values ​​for each category. This semantic classification result is not only used for subsequent extraction of tree point cloud subsets to construct a two-dimensional projection image, but also serves as an auxiliary supervisory signal in the training of a multi-anchor regression head, achieving collaborative optimization of semantic and structural information.

[0028] Furthermore, based on the same high-dimensional geometric features, a multi-anchor regression head predicts the spatial offset vector of each sampling point relative to its respective tree instance at multiple preset height levels. These multiple height levels are distributed at intervals along the vertical direction (usually the Z-axis), and can be set to key anatomical locations such as 0.5m, 1.3m, 3.0m above the ground, and the top of the canopy, etc., which are not limited in this invention. Each height level corresponds to an anchor point, and the spatial offset vector is a three-dimensional coordinate difference (Δx, Δy, Δz), representing the displacement from the current sampling point to the anchor point of that level. The multi-anchor regression head adopts a multi-branch parallel structure, with each branch responsible for the offset prediction of one height level, thereby explicitly encoding the structural prior of the tree on different vertical profiles. This design avoids the prediction drift problem caused by local occlusion in tall or sloping trees in single-center point regression.

[0029] It should be noted that, to enable adaptive aggregation of multi-anchor offset information, this embodiment introduces learnable weight parameters to perform weighted aggregation of spatial offset vectors at each height level, generating multi-anchor offset information for each sampling point. These weight parameters can be automatically learned through end-to-end training, and their values ​​reflect the contribution of each height level anchor point to the attribution determination of the current sampling point. For example, for sampling points near the ground, lower-level anchor points may have higher weights; while for points in the canopy region, higher-level anchor points are more discriminative. The weighted aggregation operation can be represented as a weighted sum, or the weights can be dynamically calculated using an attention mechanism; this invention does not limit the specific approach. The fused multi-anchor offset information retains the details of the multi-level structure while achieving information compression and noise suppression, laying the foundation for the subsequent calculation of the predicted value of the fused instance center.

[0030] In this embodiment, the semantic segmentation head and the multi-anchor regression head share the same feature extraction backbone network, but each has its own independent parameter space to achieve task decoupling and gradient stabilization. During training, a multi-task loss function can be used, including cross-entropy loss for semantic segmentation and Smooth L1 loss for multi-anchor regression, and the optimization difficulty of the two tasks can be balanced through uncertainty weighting or dynamic weighting strategies. During inference, the semantic classification result and multi-anchor offset information are output synchronously, allowing for direct use by downstream modules without additional post-processing.

[0031] Preferably, to improve inference efficiency, voxel downsampling or farthest point sampling can be performed on the point cloud to ensure uniform sampling point density and controllable computational load, while ensuring that points in key structural regions (such as the base of a tree trunk) are not excessively sparsified.

[0032] S2: Extract a subset of tree point clouds from the tree point cloud data based on the semantic classification results, and perform height adaptive processing on the subset of tree point clouds to construct a two-dimensional horizontal cross-sectional projection image.

[0033] In tree instance segmentation tasks based on 3D point clouds, directly applying the visual base model to the original point cloud results in modal mismatch, and a large number of non-trunk points (such as branches, leaves, and ground vegetation) in the original point cloud will seriously interfere with the detection accuracy of the tree trunk base.

[0034] To address the aforementioned issues, this embodiment proposes a semantically guided height-adaptive 2D projection construction method. By selecting point clouds of specific semantic categories and focusing on key height intervals, the 3D spatial structure is transformed into a regular 2D image representation. This method not only effectively filters out background noise but also preserves the most discriminative geometric features of the tree trunk base, providing high-quality input for subsequent zero-shot or few-shot detection using visual baseline models.

[0035] In this embodiment, firstly, based on the semantic classification results output in step S1, sampling points labeled "tree trunk" are selected from the original tree point cloud data to form an initial tree trunk point cloud set. This selection operation is based on point-by-point probability distribution or hard label determination, retaining only tree trunk points with high confidence, thereby significantly reducing the amount of data and computational complexity in subsequent processing. Preferably, a probability threshold (e.g., 0.7) can be set to further filter low-confidence points to improve the purity of the set.

[0036] It should be noted that although this embodiment uses "tree trunk" as the target category, in other application scenarios, it can be extended to composite semantic combinations such as "main trunk + first-level branches," and this invention does not limit it here. The initial tree trunk point cloud set serves as the input for height adaptive processing, ensuring that subsequent projection content focuses on the core supporting structure of the single tree instance.

[0037] Furthermore, a vertical density analysis is performed on the initial trunk point cloud set to obtain the distribution curve of point cloud density as a function of height. Specifically, the Z-coordinate of the trunk points is divided into bins with a fixed step size (e.g., 0.05m), and the number of points in each height interval is counted to form a one-dimensional density histogram. To eliminate the influence of local fluctuations, a moving average or Gaussian smoothing filter can be applied to the histogram to obtain a continuous density-height distribution curve. This curve reflects the clustering characteristics of the trunk points in the vertical dimension, typically showing a significant peak in the trunk base region or a density abrupt change due to scan occlusion. In this embodiment, the density analysis is performed only on the initial trunk point cloud set, avoiding interference from non-trunk points on the distribution pattern, thereby ensuring the specificity and robustness of feature extraction.

[0038] Based on the aforementioned distribution curves, the system automatically identifies density abrupt change intervals or peak intervals corresponding to the base of the tree trunk and determines the corresponding height range as the target height slice range. Specifically, algorithms such as first-order derivative extremum detection, local maximum search, or adaptive threshold segmentation can be used to locate continuous intervals in the curve that are significantly higher than the background level. For example, when the density value of a certain height segment exceeds twice the standard deviation of the global mean and its sustained width is greater than the number of bins corresponding to the preset minimum trunk diameter, it is determined to be a valid base interval. Preferably, if multiple candidate intervals are detected, the interval with the lowest position is selected as the target to conform to the physical definition of the tree trunk base. It should be noted that this adaptive mechanism is compatible with data variations under different terrain undulations, scanning angles, and stand densities, eliminating the need for manually setting a uniform height window and significantly enhancing the generalization ability of the method.

[0039] After determining the target height slice range, the tree trunk sampling points within this range are orthogonally projected onto a horizontal plane (XY plane), and a density value mapping matrix is ​​generated through rasterization. Rasterization involves dividing the projected area into two-dimensional grid cells of fixed resolution, and counting the number of points falling within each cell as the original density value of the corresponding pixel. This density value mapping matrix is ​​essentially a bird's-eye view (BEV) representation; its grayscale intensity intuitively reflects the spatial occupancy pattern of the tree trunk cross-section, with circular or elliptical high-density areas corresponding to potential single-tree base locations. Preferably, the raster resolution can be dynamically adjusted based on the average spacing of the point cloud, for example, set between 0.02m and 0.05m, to balance detail preservation and computational efficiency. For empty grids with no points falling within them, their density value is set to zero to ensure matrix integrity.

[0040] Finally, normalization and size standardization operations are performed sequentially on the density value mapping matrix to obtain the final two-dimensional horizontal cross-sectional projection image. The normalization operation linearly maps the original density values ​​to the integer range of [0, 255] or the floating-point range of [0, 1], eliminating the dimensional influence of absolute point counts caused by differences in scanning distance or vegetation density. Size standardization uniformly scales or fills the image to the fixed input size required by the visual base model (e.g., 1024×1024 pixels), while recording the scaling factor and translation offset for subsequent coordinate inverse transformation.

[0041] In this embodiment, the standardization process can employ bilinear interpolation to maintain edge continuity and avoid introducing artificial jagged edges. The resulting two-dimensional horizontal cross-sectional projection image not only preserves the topological structure and relative spatial relationships of the tree trunk base but also possesses texture statistical characteristics similar to natural images, enabling seamless integration into visual segmentation processes based on text prompts and achieving effective cross-modal knowledge transfer.

[0042] S3: Based on the preset visual base model and text prompt conditions, perform tree trunk base detection on the two-dimensional horizontal cross-section projection image to obtain the two-dimensional tree trunk base mask and instance count prior.

[0043] In the interdisciplinary application of forestry remote sensing and computer vision, traditional methods for detecting tree trunk bases heavily rely on training dedicated models with large amounts of labeled data, making it difficult to adapt to the diversity of different forest stand types, scanning equipment, and environmental conditions. In recent years, large-scale pre-trained visual basic models have demonstrated strong zero-shot generalization capabilities, but when directly applied to two-dimensional point cloud projections of non-natural images, they often fail to detect due to semantic gaps.

[0044] To address this, this embodiment proposes a visual basic model-driven tree trunk base detection mechanism based on text prompts. By encoding domain knowledge into natural language descriptions, the model's ability to understand specific structures is activated, thereby achieving high-precision mask generation and instance counting without additional fine-tuning.

[0045] In this embodiment, a sequence of text prompt words containing semantic descriptions of the tree trunk base is first constructed as text prompt conditions. This sequence of text prompt words is not a simple label, but a structured natural language description that integrates morphological, positional, and contextual information. For example, "a dense cluster of circular or elliptical dots located in the central region of the image, representing the root cross-section of a single tree" or "a high-density region distributed in a ring on a horizontal cross-section, corresponding to the tree trunk base." Preferably, multiple candidate prompt words can be constructed and robustness improved through an integration strategy; alternatively, synonym expansion can be introduced to cover variations in expression.

[0046] It should be noted that the specific form of the text prompts can be adjusted according to the prompting engineering specifications of the underlying visual model used, and this invention does not impose any limitations on it. These text prompts, acting as semantic anchors, effectively bridge the modal differences between point cloud density maps and natural images, guiding the model to focus on the target structure rather than background noise.

[0047] Furthermore, the two-dimensional horizontal cross-sectional projection image and the aforementioned text prompt conditions are jointly input into a preset visual foundation model, outputting pixel-level prediction results that semantically match the text prompt word sequence. The visual foundation model can be CLIP-Seg, Grounded SAM, OpenSeeD, or other multimodal architectures that support open vocabulary segmentation, which internally achieves the alignment and fusion of image features and text embeddings through a cross-attention mechanism.

[0048] In this embodiment, the pixel-level prediction results output by the model are a probability map or similarity heatmap of each pixel belonging to the "tree trunk base" category, and the high-response areas represent spatial locations consistent with the semantics of the text description. Preferably, the input image can be processed by multi-scale cropping or sliding window processing to balance the detection accuracy of small targets with the coverage of large scenes; however, this invention does not impose any limitations on this. This process does not require any forest-specific annotation data and relies entirely on pre-trained knowledge transfer, significantly reducing the deployment threshold.

[0049] Based on pixel-level prediction results, regions with confidence levels higher than a preset threshold are extracted and used as two-dimensional trunk base masks. The preset threshold can be dynamically set according to the performance of the validation set, or it can be determined using adaptive methods such as the Otsu algorithm or based on percentiles of the predicted value distribution, to balance recall and false positive rates. The extracted mask is a binary image, where foreground pixels identify potential trunk base locations.

[0050] It should be noted that before thresholding, morphological opening and closing operations or nonmaximum suppression can be applied to the original prediction map to eliminate fragmented responses and merge adjacent high-confidence regions, thereby improving the connectivity and geometric rationality of the mask. The resulting two-dimensional trunk base mask not only provides accurate spatial positioning information but also implicitly contains the boundary contours of individual trees, laying a two-dimensional prior foundation for subsequent three-dimensional seed point generation.

[0051] Finally, connected component analysis is performed on the 2D trunk base mask, and the number of connected components obtained is used as the instance count prior. Specifically, the binary mask is labeled using a four-neighbor or eight-neighbor connectivity criterion, with each independent connected region corresponding to a potential single tree instance. This instance count prior is not the final result, but rather serves as a strong constraint signal to guide seed point selection and anomaly filtering in the subsequent 3D clustering process. For example, if candidate clusters far exceeding this count are detected in 3D space, oversegmentation can be determined; conversely, it may result in missed detections. Preferably, auxiliary attributes such as the area, centroid coordinates, and shape factor of each connected component can be recorded simultaneously for weighted use by downstream modules. In this embodiment, this counting prior originates from the semantic understanding capability of the visual base model, exhibiting stronger semantic consistency and noise resistance compared to pure geometric clustering methods.

[0052] S4: Perform inverse coordinate transformation on the 2D tree trunk base mask to obtain 3D tree trunk base candidate points, and assign initial instance identifiers to the 3D tree trunk base candidate points based on instance count priors to obtain the 3D base position prior information of a single tree instance.

[0053] In the tree instance segmentation process based on a visual fundamental model, although 2D detection results can provide high-precision trunk base masks and instance count priors, they are essentially limited to the image plane and cannot be directly used for instance segmentation in 3D point cloud space. To achieve cross-modal connection from 2D perception to 3D reconstruction, the mask information must be accurately mapped back to the original point cloud coordinate system, and spatial verification must be performed in conjunction with instance count constraints. This embodiment proposes a mechanism for joint assignment of coordinate inverse transformation and instance identifiers, aiming to generate 3D base position prior information with both geometric accuracy and semantic consistency, providing reliable seed point guidance for subsequent 3D clustering.

[0054] In this embodiment, a mapping relationship is first established between the pixel coordinate system of the two-dimensional horizontal cross-section projection image and the three-dimensional world coordinate system of the tree point cloud data. This mapping relationship is defined by the resolution, origin offset, and size normalization parameters determined by the rasterization process in step S2, and can usually be expressed as an affine transformation matrix or an explicit coordinate transformation formula.

[0055] Specifically, the 3D world coordinates (X, Y) corresponding to the pixel coordinates (u, v) can be calculated through inverse scaling and translation, while the Z-axis coordinate is recovered based on the target height slice range determined in step S2, for example, by taking the median or weighted centroid height of that range. Preferably, if non-uniform sampling or adaptive raster is used during projection, the mapping relationship should include corresponding inverse transformation logic to ensure coordinate accuracy. This invention does not limit the specific mathematical expression form, as long as it can achieve bidirectional lossless or near-lossless coordinate transformation.

[0056] Based on the above mapping relationship, the effective pixels (i.e., foreground pixels) within the two-dimensional tree trunk base mask are back-projected into three-dimensional space, and the Z-axis coordinates are recovered according to the target height slice range to obtain three-dimensional tree trunk base candidate points. The effective pixels refer to pixels in the mask with a value of 1 or a confidence level higher than the threshold. The back-projection process is essentially "stretching" the two-dimensional detection area into a columnar candidate in three-dimensional space.

[0057] In this embodiment, each valid pixel corresponds to one or more 3D candidate points. If the original point cloud contains multiple trunk points within the pixel grid, all of them can be retained as candidates. If there are no original points, virtual candidate points can be generated by interpolation based on neighboring points to avoid information loss. It should be noted that the 3D trunk base candidate points obtained by backprojection are not the final instance representatives, but rather an initial set containing potential noise and redundancy. Further screening and verification are required by combining instance count priors.

[0058] Furthermore, based on the instance count prior indication, connectivity verification is performed on the candidate points for the 3D trunk base to obtain the target 3D trunk base candidate points. Specifically, using the instance count prior as a strong constraint on the number of clusters, density-based clustering algorithms (such as DBSCAN) or graph connected component analysis are used to group the distribution of candidate points in 3D space. If the number of clusters obtained from the initial clustering is consistent with the instance count prior, it is directly adopted; if there is a deviation, iterative optimization is performed by adjusting parameters such as neighborhood radius and minimum number of points until the quantity constraint is met or a preset tolerance range is reached.

[0059] Preferably, the multi-anchor point offset information from step S1 can be introduced as an auxiliary criterion to eliminate outlier candidate points that do not match the characteristics of the tree trunk base in the vertical structure. This connectivity verification mechanism effectively suppresses the over-segmentation / under-segmentation problem caused by mask fragmentation or back-projection errors, ensuring that each connected cluster corresponds to a real single tree base.

[0060] Finally, a unique instance label is assigned to each candidate point at the base of the target 3D tree trunk, and the set of spatial coordinates associated with this instance label is used as prior information for the 3D base position of the individual tree instance. The instance label can be an integer ID, UUID, or other unique identifier, which remains unchanged throughout the segmentation process and is used to track the representation of the same tree at different processing stages. The set of spatial coordinates not only includes the XYZ coordinates of the candidate points but can also include metadata such as semantic classification probability, multi-anchor offset information, and 2D mask confidence, forming a rich prior description. In this embodiment, this prior information for the 3D base position will be used as the seed point input for subsequent 3D clustering, significantly reducing the dependence on pure geometric features and improving the segmentation robustness under complex forest stands.

[0061] S5: Based on the weight parameters, the offset information of multiple preset height level anchor points corresponding to each sampling point is fused to obtain the fused instance center prediction value of each sampling point.

[0062] In point cloud-based tree instance segmentation tasks, due to the complex morphology of trees, the interlacing of branches and leaves, and the presence of occlusion and noise in the scanned data, geometric features at a single height level are insufficient to stably represent the geometric center of a single tree instance. Relying solely on the offset prediction of a single anchor point is susceptible to local structural distortions, leading to center estimation bias and consequently causing oversegmentation or undersegmentation.

[0063] To address this, this embodiment proposes a weighted fusion mechanism for anchor point offset information across multiple height levels. By integrating spatial offset predictions and their confidence weights from multiple preset levels in the vertical dimension, robust fused instance center predictions are generated. This mechanism effectively suppresses the uncertainty of single-layer predictions, enhances the consistent representation of cross-height structures, and provides high-precision center guidance signals for subsequent 3D clustering.

[0064] In this embodiment, the spatial offset vector predicted for each sampling point at each preset height level and the corresponding weight parameters are first obtained. The preset height levels refer to several representative sections (such as 0.3m, 1.0m, 1.8m, etc.) pre-divided along the vertical direction of the tree trunk. Each level corresponds to an anchor point, and the model outputs a three-dimensional offset vector (Δx, Δy, Δz) from the current sampling point to the center of the instance at that level. Simultaneously, the model also outputs the reliability weight of this offset prediction, which is typically dynamically generated by semantic confidence, geometric consistency score, or attention mechanism.

[0065] Preferably, the weight parameters can be normalized using softmax to ensure that the sum of the weights of the same sampling point at each level is 1, thereby maintaining numerical stability. It should be noted that the number and location of the preset height levels can be adaptively configured according to the forest stand type; this invention does not limit this, as long as it covers the key structural regions of the tree trunk.

[0066] Furthermore, the spatial offset vectors of each height level are weighted and summed based on the obtained weight parameters to obtain the comprehensive offset vector. Specifically, for any sampling point p, its comprehensive offset vector is expressed as: in For the first Weights at each high level, This is the spatial offset vector corresponding to this level. This weighted fusion process is essentially a soft voting mechanism, where the offset contribution of high-confidence levels is greater, while the influence of low-confidence or anomalous levels is automatically suppressed.

[0067] In this embodiment, the weighted summation can be performed in parallel on the GPU to meet the real-time requirements of large-scale point cloud processing. Preferably, nonlinear activation or residual correction terms can also be introduced to compensate for the systematic bias that may be caused by linear superposition, but the core still maintains the basic paradigm of weighted linear fusion. This comprehensive offset vector represents the optimal estimated direction and distance of the sampling point in three-dimensional space towards the geometric center of its respective tree instance.

[0068] Subsequently, the composite offset vector is superimposed onto the original 3D coordinates of the sampling point to calculate the predicted position of the sampling point pointing to the geometric center of its respective individual tree instance, which is then used as the predicted value of the fused instance center. That is: in, These are the original (x, y, z) coordinates of the sampling point. These are the fused instance center prediction coordinates. This prediction is no longer limited to a specific height plane, but rather incorporates the geometric consensus of the entire tree trunk's vertical structure, thus exhibiting greater robustness to local missing or deformed features. It should be noted that the resulting fused instance center predictions are continuous three-dimensional coordinates, which can be directly used for subsequent density-based clustering or voting space construction without additional discretization. In this embodiment, all sampling points independently calculate their fused center predictions, forming a dense candidate center field, providing a high-resolution prior for instance-level aggregation.

[0069] Preferably, the weight parameters can be dynamically modulated during the fusion process by incorporating the prior information on the three-dimensional basal location obtained in step S4. For example, if the distance between the predicted fusion center value of a sampling point and the nearest three-dimensional basal prior point exceeds a threshold, its overall confidence can be reduced or a reweighting strategy can be triggered to avoid false centers far from the real basal location interfering with the clustering results.

[0070] S6: Construct a joint seed point set based on the prior information of the three-dimensional base location and the predicted value of the fused instance center, and perform joint clustering and instance assignment on the joint seed point set in sequence to obtain the instance segmentation result of the tree point cloud data.

[0071] In forest point cloud instance segmentation tasks, relying solely on geometric clustering or pure semantic prediction often fails to balance the localization accuracy of the trunk base with the integrity of the canopy structure. Traditional region growing algorithms are sensitive to seed point selection and are prone to over-segmentation or under-segmentation due to initial position deviations; while center-voting methods can capture the global structure, they may generate false clusters in dense forest stands. To address this, this embodiment proposes a joint seed point construction and allocation mechanism that integrates 3D base location priors with instance center predictions. By integrating bottom-up structural anchor points and top-down center guidance signals, robust and accurate single-tree instance segmentation is achieved.

[0072] In this embodiment, the candidate points of the 3D trunk base with initial instance identifiers and the sampling points whose confidence levels of the fused instance center predictions meet a preset threshold are first integrated to obtain a joint seed point set. The candidate points of the 3D trunk base originate from step S4, carrying explicit instance labels and spatial coordinates, representing the physical roots of a single tree; while the high-confidence fused instance center prediction sampling points come from step S5, reflecting the model's strong response region to the geometric center of the instance.

[0073] Preferably, the two types of seed points can be deduplicated and merged. When the spatial distance between the two is less than a preset tolerance, the basal candidate points are retained to maintain instance identification consistency. This joint seed point set has both structural reliability and semantic saliency, effectively avoiding initialization bias caused by relying on only a single type of seed. This invention does not limit the specific integration strategy, as long as the complementary fusion of the two types of prior information is achieved.

[0074] Furthermore, starting with the seed points in the joint seed point set, region growing clustering is performed on the tree point cloud data based on geometric proximity constraints and feature similarity metrics to obtain preliminary instance clusters. The geometric proximity constraints can employ metrics such as Euclidean distance, normal vector angle, or curvature difference to ensure the growth process extends along a continuous surface. The feature similarity metrics combine semantic classification probability, color information, and the consistency of the center prediction direction obtained in step S5 to prevent erroneous merging across instances. In this embodiment, region growing uses a priority queue to dynamically update the neighborhood, prioritizing the expansion of high-similarity points to improve boundary fit.

[0075] It should be noted that a maximum radius or upper limit on the number of points can be set during the growth process to avoid excessive expansion caused by abnormal seeds. The resulting preliminary instance clusters cover most of the clearly assigned trunk and main branch points, but some edge points, sparse branch and leaf points, or shading area points are still not assigned, forming the remaining set of sampling points.

[0076] For the remaining sampling points, their spatial distance to each preliminary instance cluster and semantic consistency score are calculated, and they are assigned to the preliminary instance cluster with the highest overall score. Specifically, the spatial distance can be the Euclidean distance from the point to the cluster centroid or the nearest distance to the cluster boundary; the semantic consistency score is based on the degree of matching between the semantic category probability of the point and the dominant category of the target cluster, and the cosine of the angle between the center prediction vector and the cluster principal axis direction can also be introduced as an auxiliary criterion. The overall score is usually expressed as a weighted linear combination or product to balance geometric proximity and semantic compatibility.

[0077] Preferably, a minimum score threshold can be set, and remaining points below this threshold are not assigned temporarily, to be determined in subsequent post-processing stages to determine whether they are noise. This assignment mechanism makes full use of the structural context of the initial clusters, making the assignment decision of fuzzy regions based on evidence, and significantly improving the recall and boundary integrity of the segmentation results.

[0078] Finally, boundary smoothing and outlier filtering are performed on each initially assigned instance cluster, outputting a single-tree instance segmentation mask and its corresponding instance identifier as the final instance segmentation result of the tree point cloud data. Boundary smoothing can employ graph cut, conditional random fields, or morphological closing operations to eliminate jagged edges and enhance geometric continuity; outlier filtering removes misassigned points based on local density, normal consistency, or Mahalanobis distance from the cluster center. In this embodiment, the output instance segmentation mask is a point-by-point label array, with each point associated with a unique instance ID, which can be directly used to extract forestry parameters such as diameter at breast height (DBH), tree height, and crown width. It is important to emphasize that the entire joint clustering and assignment process is entirely driven by the prior information provided in steps S4 and S5, requiring no additional annotation or manual intervention, achieving end-to-end automation from perception to segmentation.

[0079] Another embodiment of the present invention provides a tree point cloud instance segmentation system. For details, please refer to [link to relevant documentation]. Figure 2 , Figure 2 The diagram shown illustrates a structural block diagram of a tree point cloud instance segmentation system according to one embodiment of the present invention, comprising: The acquisition module 11 is used to acquire tree point cloud data of the area to be processed, and to perform feature extraction and multi-anchor point prediction on the tree point cloud data to obtain the semantic classification result and multi-anchor point offset information of each sampling point; wherein, the multi-anchor point offset information is configured to characterize the spatial relationship of each sampling point relative to multiple preset height level anchor points of its respective single tree instance through weight parameters. Extraction module 12 is used to extract a subset of tree point clouds from tree point cloud data based on semantic classification results, and to perform height adaptive processing on the subset of tree point clouds to construct a two-dimensional horizontal cross-sectional projection image. The detection module 13 is used to detect the base of the tree trunk in a two-dimensional horizontal cross-section projection image based on a preset visual basic model and text prompt conditions, and to obtain a two-dimensional tree trunk base mask and instance count prior. The transformation module 14 is used to perform inverse coordinate transformation on the two-dimensional tree trunk base mask to obtain three-dimensional tree trunk base candidate points, and assign initial instance identifiers to the three-dimensional tree trunk base candidate points based on instance counting priors to obtain the three-dimensional base position prior information of a single tree instance. The fusion module 15 is used to fuse the offset information of multiple preset height level anchor points corresponding to each sampling point according to the weight parameters to obtain the fusion instance center prediction value of each sampling point. The segmentation module 16 is used to construct a joint seed point set based on the prior information of the three-dimensional base position and the predicted value of the fused instance center, and to perform joint clustering and instance assignment on the joint seed point set in sequence to obtain the instance segmentation result of the tree point cloud data.

[0080] Preferably, in one embodiment of the present invention, the acquisition module is specifically used for: Tree point cloud data is input into a pre-trained point cloud feature extraction network to obtain high-dimensional geometric features of each sampling point; Based on high-dimensional geometric features, the semantic segmentation head outputs the probability distribution of each sampling point belonging to a preset forest component category, which serves as the semantic classification result for each sampling point. Based on high-dimensional geometric features, the spatial offset vector of each sampling point relative to its respective single tree instance is predicted by a multi-anchor point regression head at multiple preset height levels. The multiple height levels are distributed at intervals along the vertical direction, and the spatial offset vector represents the three-dimensional coordinate difference between the sampling point and the corresponding height level anchor point. Based on learnable weight parameters, the spatial offset vectors of each height level are weighted and aggregated to generate multi-anchor offset information for each sampling point.

[0081] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for segmenting tree point cloud instances, characterized in that, include: Tree point cloud data of the area to be processed is acquired, and feature extraction and multi-anchor point prediction are performed on the tree point cloud data respectively to obtain the semantic classification result and multi-anchor point offset information of each sampling point; wherein, the multi-anchor point offset information is configured to characterize the spatial relationship of each sampling point relative to multiple preset height level anchor points of its respective single tree instance through weight parameters. Based on the semantic classification results, a subset of tree point clouds is extracted from the tree point cloud data, and the tree point cloud subset is subjected to height adaptive processing to construct a two-dimensional horizontal cross-sectional projection image. Based on the preset visual basic model and text prompt conditions, the tree trunk base is detected in the two-dimensional horizontal cross-section projection image to obtain the two-dimensional tree trunk base mask and instance count prior. Perform an inverse coordinate transformation on the two-dimensional tree trunk base mask to obtain three-dimensional tree trunk base candidate points, and assign initial instance identifiers to the three-dimensional tree trunk base candidate points based on the instance counting prior to obtain the three-dimensional base position prior information of a single tree instance; Based on the weight parameters, the offset information of multiple preset height level anchor points corresponding to each sampling point is fused to obtain the fused instance center prediction value of each sampling point. Based on the prior information of the three-dimensional base location and the predicted value of the fused instance center, a joint seed point set is constructed, and joint clustering and instance assignment are performed on the joint seed point set in sequence to obtain the instance segmentation result of the tree point cloud data.

2. The tree point cloud instance segmentation method as described in claim 1, characterized in that, The step of performing feature extraction and multi-anchor point prediction on the tree point cloud data to obtain the semantic classification result and multi-anchor point offset information for each sampling point includes: The tree point cloud data is input into a pre-trained point cloud feature extraction network to obtain the high-dimensional geometric features of each sampling point; Based on the high-dimensional geometric features, the semantic segmentation head outputs the probability distribution of each sampling point belonging to a preset forest component category, which serves as the semantic classification result for each sampling point. Based on the high-dimensional geometric features, the spatial offset vector of each sampling point relative to its respective single tree instance is predicted by a multi-anchor point regression head at multiple preset height levels; wherein, the multiple height levels are distributed at intervals along the vertical direction, and the spatial offset vector represents the three-dimensional coordinate difference between the sampling point and the corresponding height level anchor point. Based on learnable weight parameters, the spatial offset vectors of each height level are weighted and aggregated to generate multi-anchor offset information for each sampling point.

3. The tree point cloud instance segmentation method as described in claim 1, characterized in that, The step of extracting a subset of tree point clouds from the tree point cloud data based on the semantic classification result, and performing height adaptive processing on the subset of tree point clouds to construct a two-dimensional horizontal cross-sectional projection image, includes: The sampling points in the semantic classification results that are classified as "tree trunk" are selected to form an initial tree trunk point cloud set; Based on the vertical distribution characteristics of the initial trunk point cloud set, determine the target height slice range that includes the trunk base features; The tree trunk sampling points located within the target height slice range are projected onto a horizontal plane, and a density value mapping matrix is ​​generated through rasterization. Normalization and size standardization operations are performed sequentially on the density value mapping matrix to obtain a two-dimensional horizontal cross-sectional projection image.

4. The tree point cloud instance segmentation method as described in claim 3, characterized in that, The step of determining the target height slice range containing the trunk base features based on the vertical distribution characteristics of the initial trunk point cloud set includes: A density analysis was performed on the initial tree trunk point cloud set in the vertical direction to obtain the distribution curve of point cloud density as a function of height. Based on the distribution curve, identify the density abrupt change interval or peak interval corresponding to the base of the tree trunk; The height range corresponding to the density abrupt change interval or peak interval is determined as the target height slice range.

5. The tree point cloud instance segmentation method as described in claim 1, characterized in that, The process of detecting the tree trunk base in the two-dimensional horizontal cross-sectional projection image based on a preset visual model and text prompt conditions, and obtaining a two-dimensional tree trunk base mask and instance count prior, includes: Construct a sequence of text prompt words containing semantic descriptions of the tree trunk base as text prompt conditions; The two-dimensional horizontal cross-sectional projection image and the text prompt conditions are jointly input into the visual basic model, and the pixel-level prediction result that semantically matches the text prompt word sequence is output. Extract the regions with confidence scores higher than a preset threshold from the pixel-level prediction results and use them as the two-dimensional trunk base mask; Connectivity analysis is performed on the two-dimensional trunk base mask, and the number of connected components obtained is used as the prior for instance counting.

6. The tree point cloud instance segmentation method as described in claim 5, characterized in that, The inverse coordinate transformation is performed on the two-dimensional tree trunk base mask to obtain three-dimensional tree trunk base candidate points. Based on the instance counting prior, initial instance identifiers are assigned to the three-dimensional tree trunk base candidate points to obtain the three-dimensional base position prior information of a single tree instance, including: Establish a mapping relationship between the pixel coordinate system of the two-dimensional horizontal cross-section projection image and the three-dimensional world coordinate system of the tree point cloud data; Based on the mapping relationship, the effective pixels within the two-dimensional trunk base mask are back-projected into three-dimensional space, and the Z-axis coordinates are recovered according to the target height slice range to obtain three-dimensional trunk base candidate points. Based on the number of instances counted prior indications, connectivity verification is performed on the candidate points for the base of the three-dimensional trunk to obtain the target three-dimensional trunk base candidate points; A unique instance label is assigned to each cluster of target 3D trunk base candidate points, and the set of spatial coordinates associated with the instance label is used as prior information for the 3D base position.

7. The tree point cloud instance segmentation method according to claim 1, characterized in that, The step of fusing the offset information of multiple preset height-level anchor points corresponding to each sampling point according to the weight parameters to obtain the fused instance center prediction value of each sampling point includes: Obtain the predicted spatial offset vector and corresponding weight parameters at each preset height level of each sampling point; Based on the weight parameters, the spatial offset vectors of each height level are weighted and summed to obtain the comprehensive offset vector. The composite offset vector is superimposed onto the original three-dimensional coordinates of the sampling point to calculate the predicted position of the sampling point pointing to the geometric center of its respective single-tree instance, which is used as the predicted value of the fused instance center.

8. The tree point cloud instance segmentation method according to claim 6, characterized in that, The process involves constructing a joint seed point set based on the prior information of the three-dimensional base location and the predicted value of the fused instance center, and then sequentially performing joint clustering and instance assignment on the joint seed point set to obtain the instance segmentation result of the tree point cloud data, including: The candidate points of the 3D trunk base with initial instance identification and the sampling points whose confidence scores of the fused instance center prediction values ​​meet the preset threshold conditions are integrated to obtain a joint seed point set. Starting from the seed points in the joint seed point set, region growing clustering is performed on the tree point cloud data based on geometric proximity constraints and feature similarity measures to obtain preliminary instance clusters; Calculate the spatial distance and semantic consistency score between the remaining sampling points and each of the preliminary instance clusters, and assign the remaining sampling points to the preliminary instance cluster with the highest overall score; wherein, the remaining sampling points include remaining sampling points that do not belong to any of the preliminary instance clusters. Boundary smoothing and outlier filtering are performed on each initially assigned instance cluster, and the single-tree instance segmentation mask and the corresponding instance identifier are output as the instance segmentation result of the tree point cloud data.

9. A tree point cloud instance segmentation system, characterized in that, include: The acquisition module is used to acquire tree point cloud data of the area to be processed, and to perform feature extraction and multi-anchor point prediction on the tree point cloud data to obtain the semantic classification result and multi-anchor point offset information of each sampling point; wherein, the multi-anchor point offset information is configured to characterize the spatial relationship of each sampling point relative to multiple preset height level anchor points of its respective single tree instance through weight parameters. The extraction module is used to extract a subset of tree point clouds from the tree point cloud data based on the semantic classification results, and to perform height adaptive processing on the subset of tree point clouds to construct a two-dimensional horizontal cross-sectional projection image. The detection module is used to detect the base of the tree trunk in the two-dimensional horizontal cross-section projection image based on a preset visual basic model and text prompt conditions, and to obtain a two-dimensional tree trunk base mask and instance count prior. The transformation module is used to perform inverse coordinate transformation on the two-dimensional tree trunk base mask to obtain three-dimensional tree trunk base candidate points, and assign initial instance identifiers to the three-dimensional tree trunk base candidate points based on the instance counting prior to obtain the three-dimensional base position prior information of a single tree instance. The fusion module is used to fuse the offset information of multiple preset height level anchor points corresponding to each sampling point according to the weight parameters, so as to obtain the fusion instance center prediction value of each sampling point. The segmentation module is used to construct a joint seed point set based on the prior information of the three-dimensional base position and the predicted value of the fused instance center, and to perform joint clustering and instance assignment on the joint seed point set in sequence to obtain the instance segmentation result of the tree point cloud data.

10. The tree point cloud instance segmentation system as described in claim 9, characterized in that, The acquisition module is specifically used for: The tree point cloud data is input into a pre-trained point cloud feature extraction network to obtain the high-dimensional geometric features of each sampling point; Based on the high-dimensional geometric features, the semantic segmentation head outputs the probability distribution of each sampling point belonging to a preset forest component category, which serves as the semantic classification result for each sampling point. Based on the high-dimensional geometric features, the spatial offset vector of each sampling point relative to its respective single tree instance is predicted by a multi-anchor point regression head at multiple preset height levels; wherein, the multiple height levels are distributed at intervals along the vertical direction, and the spatial offset vector represents the three-dimensional coordinate difference between the sampling point and the corresponding height level anchor point. Based on learnable weight parameters, the spatial offset vectors of each height level are weighted and aggregated to generate multi-anchor offset information for each sampling point.

Images