Method and system for automatic annotation of feature points based on point cloud data

Abstract

The present invention relates to the field of image processing technology, and in particular to a method and system for automatic feature point annotation based on point cloud data. The method comprises the following steps: acquiring regional multi-source point cloud data, and performing multimodal data fusion to obtain point cloud fusion data; performing ground point cloud segmentation of shoreline structure based on the point cloud fusion data to obtain a ground point cloud segmentation data set; performing edge detection on the ground point cloud segmentation data set, and identifying geometrically significant points based on the edge detection results, thereby obtaining a feature point candidate set; performing ground feature type classification based on the feature point candidate set to obtain a target classification data set; assigning symbol styles to the target classification data set and binding point cloud feature attributes to obtain a symbolized data set; performing multi-level symbol dynamic adjustment and interactive annotation based on the symbolized data set to obtain a symbol dynamic view. The present invention helps to improve the accuracy and efficiency of the symbolic expression of feature points on the shore, and has strong interactivity.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a method and system for automatically labeling feature points based on point cloud data. Background Art

[0002] As a high-precision spatial data acquisition method, point cloud data has been widely used in various measurement and modeling fields. Point cloud data offers significant advantages, particularly in the symbolization of onshore feature points. By processing and analyzing point cloud data of terrain, buildings, and other natural or man-made features, complex landform structures can be accurately restored, providing important data support for environmental monitoring, urban planning, disaster assessment, and other applications. However, traditional methods for symbolizing onshore feature points suffer from numerous drawbacks when applying point cloud data. First, traditional methods rely primarily on manual feature point extraction, which is inefficient, costly, and susceptible to subjective factors, leading to inconsistent extraction results. Second, because point cloud data often contains significant noise and irregularities, traditional feature point extraction methods struggle to effectively remove this noise, significantly compromising the accuracy and reliability of the extracted results. Consequently, data processing and feature extraction during onshore feature point symbolization often require significant computational resources and time. Summary of the Invention

[0003] Based on this, it is necessary for the present invention to provide a method and system for automatic feature point annotation based on point cloud data to solve at least one of the above technical problems.

[0004] To achieve the above purpose, a method for automatically labeling feature points based on point cloud data includes the following steps:

[0005] Step S1: Acquire regional multi-source point cloud data, and perform multi-modal data fusion on the regional multi-modal point cloud data to obtain point cloud fusion data;

[0006] Step S2: Segment the shoreline structure ground point cloud based on the point cloud fusion data to obtain a ground point cloud segmentation dataset;

[0007] Step S3: performing edge detection on the ground point cloud segmentation dataset, and identifying geometrically significant points based on the edge detection results, thereby obtaining a candidate set of feature points;

[0008] Step S4: classify the feature types based on the feature point candidate set to obtain a target classification dataset; assign a symbol style to the target classification dataset and bind point cloud feature attributes in combination with a predefined symbol style library to obtain a symbolized dataset;

[0009] Step S5: Perform multi-level symbol dynamic adjustment and interactive annotation based on the symbolized dataset to obtain a symbol dynamic view.

[0010] By fusing multi-source point cloud data, the present invention improves data integrity and consistency, enabling point cloud data collected by different sensors to complement each other, thereby enhancing the ability to express shoreline features. Based on point cloud fusion, a high-precision shoreline structure ground point cloud segmentation method is employed to effectively distinguish between water and land areas, ensuring the accuracy of subsequent feature point extraction. Edge detection technology is used to automatically identify geometrically significant points on the shore, reducing manual intervention and improving the automation level of feature point extraction. It also reduces the impact of human subjective factors on the consistency of the results, making feature point positioning more stable and reliable. Furthermore, during the feature point classification process, the geometric and texture features of the target features are fully utilized to improve the accuracy of feature type classification, making the symbolization process more consistent with the actual feature attributes and avoiding the symbol mismatch problem caused by classification errors in traditional methods. During the symbolization process, a predefined symbol style library is combined to automatically match the appropriate symbol style for the classified target feature and bind the relevant point cloud features, achieving symbolic expression of the point cloud data, thereby improving the intuitiveness and readability of the data. Based on symbolic data, the system dynamically adjusts multi-level symbols to optimize their layout and hierarchical structure, making symbolic expression clearer and more organized. Combined with interactive annotation, this system enhances the user's perception of symbolic information from different viewing angles. Ultimately, the generation of dynamic symbol views ensures data visualization, making the symbolic representation of onshore feature points not only precise but also highly interactive, providing efficient and accurate symbolic point cloud data support for environmental monitoring and other related applications.

[0011] Optionally, step S1 specifically includes:

[0012] Step S11: acquiring regional multi-source point cloud data, and performing coordinate transformation and time stamp alignment on the regional multi-source point cloud data to obtain multi-source point cloud data to be processed;

[0013] Step S12: performing noise filtering and voxel downsampling on the multi-source point cloud data to be processed to obtain denoised point cloud data;

[0014] Step S13: performing point cloud registration and coordinate alignment on the denoised point cloud data to obtain aligned point cloud data;

[0015] Step S14: identifying geometric differences of multi-source point cloud data based on the aligned point cloud data, and performing local morphological reconstruction optimization on the geometric differences of the multi-source point cloud data to obtain standardized point cloud geometric data;

[0016] Step S15: performing point cloud data fusion on the aligned point cloud data according to the standardized point cloud geometry data to obtain point cloud fusion data.

[0017] The present invention performs coordinate transformation and timestamp alignment on regional multi-source point cloud data, ensuring the spatial and temporal consistency of point cloud data from different sources, solving the problem of data alignment errors caused by differences in coordinate systems and different acquisition times, and making subsequent processing more accurate and reliable. In the denoising and downsampling stages, a noise filtering method is used to effectively remove isolated points, outliers and other interfering data. At the same time, voxelized downsampling is used to reduce data redundancy, improve the efficiency of point cloud processing and storage optimization capabilities, make the data more refined and retain the main feature information. In the point cloud registration process, feature matching and optimization algorithms are combined to enable point clouds from different data sources to be aligned with high precision in a unified coordinate system, eliminating the impact of coordinate offset and error accumulation on the overall data quality, and enhancing the global consistency of the data. Based on the geometric difference calculation of aligned point cloud data, it is possible to detect and quantify the morphological differences between multi-source point clouds, and optimize and compensate for data missing areas through local morphological reconstruction, thereby improving the integrity and accuracy of the data and making the point cloud model closer to the real environment. Multi-source point cloud fusion is performed based on the optimized standardized point cloud data to achieve complementary integration of data from different sources, making the spatial expression of point cloud data more delicate and continuous, enhancing the restoration and adaptability of shore features, and laying a high-quality foundation for subsequent shoreline structure segmentation, feature point extraction and symbolic expression.

[0018] Optionally, step S12 is specifically as follows:

[0019] Step S121: performing outlier detection on the multi-source point cloud data to be processed, setting the local area point count threshold to [10, 100] and the mean distance standard deviation to [1.0, 3.0] to filter the point cloud subset and obtain a preliminary denoised point cloud dataset;

[0020] Step S122: performing density equalization processing on the preliminary denoised point cloud dataset, setting the spatial radius to [0.1, 2.0] to remove sparse noise points, and generating a density-optimized point cloud dataset;

[0021] Step S123: constructing a three-dimensional voxel grid based on the density-optimized point cloud dataset, setting the voxel size to [0.01, 0.5] for spatial discretization, and performing voxel center mapping on the point cloud within each voxel to obtain a voxelized downsampled point cloud dataset;

[0022] Step S124: adjusting the voxel spacing in the voxelized downsampled point cloud dataset to [0.5, 2.0], and performing edge detail optimization to obtain a uniformly distributed point cloud dataset;

[0023] Step S125: performing coordinate normalization processing on the uniformly distributed point cloud data set, and correcting the direction of the point cloud normal vector to generate denoised point cloud data.

[0024] Through outlier detection, the present invention effectively removes isolated and abnormal points from the data, improving the overall quality of the point cloud data and avoiding the impact of noise interference on subsequent processing. During density equalization, sparse noise points are removed by setting a spatial radius, making the distribution of the point cloud data more uniform, thereby improving the stability and consistency of the point cloud data and reducing processing errors caused by uneven point density. During the construction of the three-dimensional voxel grid, a voxel downsampling strategy is adopted to spatially discretize the point cloud data. Voxel center mapping is used to reduce data redundancy while retaining key structural information, improving storage and computational efficiency. During the voxel spacing adjustment and edge detail optimization stages, the distribution of the point cloud data is rebalanced to maintain a reasonable density in different areas, especially optimizing the edge detail areas to avoid the loss of important geometric information. Finally, during coordinate normalization and normal vector correction, the uniform scale of the point cloud data is ensured, while the geometric orientation characteristics of the point cloud are optimized, making it more suitable for subsequent analysis and feature extraction, thus providing a solid foundation for the accuracy and stability of the symbolization of onshore feature points.

[0025] Optionally, step S2 is specifically:

[0026] Step S21: performing elevation histogram statistical analysis on the point cloud fusion data, calculating the elevation frequency histogram to identify elevation anomaly areas, and performing outlier detection on the elevation anomaly areas to generate elevation normalized point cloud data;

[0027] Step S22: extracting ground points from the elevation-normalized point cloud data based on a cloth simulation filtering algorithm, performing cloth simulation iteration to approximate the terrain surface, separating ground points from non-ground points, and generating a preliminary ground point cloud dataset;

[0028] Step S23: performing plane regression analysis on the preliminary ground point cloud dataset, setting the plane inlier threshold to [10, 500] and the inlier residual threshold to [0.001, 0.05], removing floating points and abnormal points with local elevation mutations, and generating an optimized ground point cloud dataset;

[0029] Step S24: Calculate the local slope and curvature of the optimized ground point cloud dataset, and mark the partitions according to the local slope and curvature to generate a terrain feature enhanced point cloud dataset;

[0030] Step S25: Segment the shoreline structure point cloud based on the terrain feature enhanced point cloud dataset, set the neighborhood search radius to [0.5, 5.0] and the terrain similarity threshold to [0.5, 1.0] for structure segmentation, and generate a ground point cloud segmentation dataset.

[0031] This invention effectively identifies elevation anomalies in point cloud data through elevation histogram statistical analysis. Combined with outlier detection, it ensures the overall consistency of elevation data, thereby improving the standardization of point cloud data. During ground point extraction, a cloth-simulation filtering algorithm is employed to iteratively approximate the terrain surface, accurately distinguishing between ground and non-ground points. This improves the accuracy of terrain data extraction and reduces misclassification. Planar regression analysis effectively removes floating points and localized abrupt outliers, ensuring the overall continuity and rationality of the ground point cloud data, reducing noise interference and making the ground point data smoother and more stable. During the local slope and curvature calculation phase, regional marking is performed based on terrain characteristics, enabling the point cloud data to better reflect the actual terrain fluctuations and providing more discriminative feature information for subsequent feature classification and symbolization. Finally, during the shoreline structure point cloud segmentation phase, precise segmentation based on geometric features is achieved by setting a neighborhood search radius and terrain similarity threshold. This ensures the integrity of the shoreline structure and improves the reliability of shoreline extraction, thus providing high-quality basic data for symbolizing shoreline feature points in the point cloud data.

[0032] Optionally, step S22 is specifically as follows:

[0033] Step S221: Grid-segment the elevation normalized point cloud data, set the grid size to [0.1, 5.0], construct a grid point cloud distribution matrix, and calculate the lowest elevation point of each grid cell to generate an initial terrain skeleton point cloud dataset;

[0034] Step S222: Initialize the cloth simulation on the initial terrain skeleton point cloud dataset, set the cloth elastic coefficient to [0.1, 1.0] and the gravity factor to [9.0, 9.8], calculate the cloth contact points using an iterative approximation method, and generate terrain surface fitting point cloud data;

[0035] Step S223: Calculate the elevation deviation between the terrain surface fitting point cloud data and the elevation normalized point cloud data, set the ground fitting error threshold to [0.01, 0.5], remove non-ground points that exceed the error threshold, and obtain a candidate set of ground points;

[0036] Step S224: Density filtering is performed on the ground point candidate set, the density threshold is set to [1, 10], isolated points and sparse points are removed, and an optimized ground point cloud dataset is generated;

[0037] Step S225: Perform connected area detection on the optimized ground point cloud dataset, set the connectivity threshold to [0.5, 1.0], remove floating points and unconnected ground segments, and generate a preliminary ground point cloud dataset.

[0038] The present invention performs preliminary structural processing on the elevation normalized point cloud data through grid segmentation, so that the point cloud data can be statistically analyzed in a local range, and the lowest elevation point is extracted to construct the initial terrain skeleton point cloud data set, thereby improving the stability and computational efficiency of ground point recognition. The terrain surface fitting is performed in combination with the cloth simulation algorithm, and the dynamic adjustment of the elastic coefficient and gravity factor is used to make the cloth model more consistent with the actual terrain morphology, reducing the misclassification problem caused by the fixed threshold in the traditional method, thereby achieving more accurate terrain surface fitting. In the terrain fitting error evaluation stage, by calculating the elevation deviation and setting the error threshold, non-ground points that deviate from the terrain surface are eliminated, effectively improving the purity of the ground point cloud data and making the identification of ground points more reliable. In the density filtering process, the density threshold constraint is used to remove isolated points and sparse points, avoiding the misclassification caused by the sparsity of the point cloud, and making the distribution of ground points more uniform. In the final connected area detection stage, by setting the connectivity threshold, floating points and unconnected ground fragments were eliminated to ensure the spatial integrity of the ground point cloud data, thereby improving the structural consistency of the terrain data and providing high-quality basic data support for the subsequent shoreline structure point cloud segmentation and symbolization.

[0039] Optionally, the step S3 of identifying geometrically salient points specifically includes:

[0040] Calculate the normal vector based on the edge detection results, set the field search radius to [0.1, 2.0], calculate the normal vector of each edge point and the angle deviation between the edge point and the neighboring point, and obtain the normal vector deviation point cloud data;

[0041] Calculate the principal curvature and curvature gradient of the normal vector deviation point cloud data, set the curvature threshold to [0.001, 0.1] to filter out low curvature points, and generate curvature enhanced point cloud data;

[0042] Perform feature response analysis on the curvature-enhanced point cloud data. Based on the Harris 3D corner detection algorithm, set the response threshold to [0.01, 0.5], extract the local maximum feature point, and generate a preliminary salient point set.

[0043] Perform spatial distribution analysis on the preliminary salient point set, set the cluster radius to [0.05, 1.0] and the minimum number of neighborhood points to [3, 20], remove isolated points, and generate an optimized salient point set;

[0044] Based on the optimized salient point set, geometric consistency analysis is performed, the geometric structure consistency threshold is set to [0.5, 1.0], low consistency points are removed, and finally a candidate set of feature points is generated.

[0045] The present invention calculates the normal vector through the edge detection results, and calculates the normal vector of each edge point and its angle deviation with the neighboring points based on the set field search radius, which can effectively identify the edge features in the point cloud data. The beneficial effect of this process is that it can accurately extract points with significant surface changes in complex terrain or building structures, laying a solid foundation for subsequent feature extraction and classification. Using the normal vector deviation point cloud data to calculate the main curvature and curvature gradient can significantly enhance the area with larger curvature and filter out the low curvature points, thereby improving the detail performance of the ground object and making the identification of ground and non-ground features more accurate. At the same time, the Harris 3D corner detection algorithm is used for feature response analysis, which can effectively extract the local maximum feature point, thereby obtaining more stable and representative feature points, which is crucial for subsequent feature point matching and analysis. Through spatial distribution analysis and setting a reasonable clustering radius and neighborhood point number limit, isolated points can be removed to ensure the accuracy and consistency of the feature point set. Finally, low-consistency points are eliminated based on geometric consistency analysis to ensure the spatial structural consistency and geometric rationality of the generated feature point candidate set, thereby further improving the quality and reliability of the dataset and providing accurate input data for subsequent land feature classification and symbolization.

[0046] Optionally, the ground feature type classification described in step S4 is specifically:

[0047] Perform local geometric feature extraction on the feature point candidate set, extract the surface normal vector, curvature, point density and morphological characteristics of each feature point, and generate a ground feature geometric feature dataset;

[0048] Perform scale space analysis on the ground feature geometric feature dataset, calculate the feature change rate at different scales, and generate a multi-scale ground feature feature dataset;

[0049] Obtain a target feature type database, perform classification training based on the target feature type database, and build a feature classification model;

[0050] The multi-scale feature dataset is classified into target feature categories using the feature classification model, and high-confidence classification results are screened to obtain a preliminary feature classification dataset.

[0051] The classification error point categories of the preliminary object classification dataset are adjusted to generate the target classification dataset.

[0052] The present invention comprehensively characterizes the spatial geometric properties of each point by extracting local geometric features from a candidate set of feature points, including surface normals, curvature, point density, and morphological features. This allows for the extraction of detailed information from complex feature morphologies and the effective differentiation of features between different features. These geometric features provide key information for subsequent feature classification and symbolization, enabling the classification model to accurately identify features of different categories. Simultaneously, scale-space analysis and calculation of feature change rates at different scales capture multi-layered feature characteristics and enhance the ability to identify details of various features across a large area. This helps adapt to the variability and complexity of different features at different scales, thereby providing richer multi-scale feature data for refined feature classification and symbolization. By acquiring a database of target feature type features and performing classification training based on them, an efficient and accurate feature classification model can be constructed. This model continuously optimizes classification accuracy based on training data, resulting in more reliable classification results in complex scenarios. By using the feature classification model to classify target feature categories within a multi-scale feature dataset, high-confidence classification results can be efficiently screened, reducing classification error and improving classification accuracy. At the same time, adjusting the categories of classification error points in the preliminary land feature classification dataset can further optimize the classification results, reduce misclassification or missed classification, and ensure that the final generated target classification dataset has high reliability and accuracy.

[0053] Optionally, the allocation symbol style and binding point cloud feature attributes described in step S4 are specifically:

[0054] Based on the target classification data set, calculate the global and local statistical characteristics of each category of ground objects and generate a statistical table of ground object characteristics;

[0055] Based on the predefined symbol style library and the statistical table of land feature characteristics, the symbol style of the target classification dataset is matched and mapped to obtain a preliminary symbol mapping dataset;

[0056] Adaptively adjust the preliminary symbol mapping dataset based on the point cloud feature attributes in the feature statistics table, set the symbol scale adjustment factor to [0.5, 2.0] and the color mapping parameter to [0, 255], adjust the symbol shape according to the height, density and classification confidence of the feature, and generate an optimized symbol mapping dataset;

[0057] Bind the optimized symbol mapping dataset to the target classification dataset, set the point cloud and symbol matching constraint to [0.1, 1.0], and generate a symbolic point cloud dataset;

[0058] Perform consistency checks on the symbolized point cloud dataset, detect and correct abnormal mapping points, and finally output the symbolized dataset.

[0059] By calculating global and local statistical features for each feature category and generating a feature feature statistical table, the present invention provides precise feature descriptions for subsequent symbolic mapping. This step not only facilitates a comprehensive analysis of the geometry and attribute information of each feature category, but also assigns the most appropriate symbol style to each category during the symbolization process, thereby improving the accuracy and consistency of symbol mapping. By matching and mapping symbol styles based on a predefined symbol style library and the feature feature statistical table, a preliminary symbolic representation can be provided for the target classification dataset. This automated process significantly improves symbolization efficiency and reduces the impact of manual intervention and subjective bias. Furthermore, adaptive adjustments are made to the preliminary symbol mapping dataset. By combining parameters such as feature height, density, and classification confidence, symbol scale adjustment factors and color mapping parameters are set to customize symbol adjustments for different feature categories. This provides visually richer and easier-to-distinguish symbol styles, enhancing data visualization and clarity. Binding the optimized symbol mapping dataset to the target classification dataset and setting point cloud and symbol matching constraints effectively ensures the rationality and accuracy of the symbolized point cloud data and avoids mismatches or errors between symbols and point clouds. Finally, consistency checking of the symbolized point cloud dataset can timely detect and correct abnormal mapping points, ensuring that the generated symbolized dataset meets the expected land feature expression standards and providing reliable and accurate symbolic data support for subsequent spatial analysis and applications.

[0060] Optionally, step S5 is specifically:

[0061] Step S51: Organize the symbolized data set into multiple levels of symbols, construct a symbol hierarchy tree, and obtain a preliminary symbol hierarchy structure;

[0062] Step S52: performing a hierarchical integrity check on the preliminary symbol hierarchy structure to obtain an optimized symbol hierarchy data set;

[0063] Step S53: obtaining a user interaction parameter set, and performing a user interaction perspective perception simulation on the user interaction parameter set and the optimized symbol hierarchy data set to obtain observation perspective character perception data;

[0064] Step S54: performing symbol visibility evaluation based on the observation perspective character perception data, and optimizing the symbol layout in the symbol hierarchy dataset according to the symbol visibility to obtain a symbol dynamic adjustment dataset;

[0065] Step S55: setting an interactive annotation mode between the symbol dynamic adjustment data set and the user interaction parameter set, constructing an interactive annotation model, and performing the symbol annotation task based on the interactive annotation model, thereby obtaining an interactive annotation data set;

[0066] Step S56: Perform dynamic visualization rendering on the interactive annotation dataset and set rendering parameters to obtain a dynamic view of the symbol.

[0067] By organizing a symbolized dataset into multiple levels of symbols and constructing a symbol hierarchy tree, the present invention enables hierarchical management and dynamic optimization of symbols, thereby improving the organization and clarity of symbol representation. The establishment of a preliminary symbol hierarchy lays the foundation for subsequent symbol optimization and adjustment, effectively supporting the presentation and visual expression of symbols at different levels. A hierarchical integrity check is performed on the symbol hierarchy to ensure the rationality and completeness of the hierarchical relationships of symbols, eliminate potential symbol mismatches or hierarchical logic errors, and generate an optimized symbol hierarchy dataset, further improving the accuracy of symbol organization. Acquiring a user interaction parameter set and performing perspective perception simulation provides more personalized perspectives for symbol visualization, enabling the symbolized dataset to better meet user needs and practical application scenarios, enhancing user experience and interactivity. Symbol visibility assessment based on observation perspective character perception data optimizes and adjusts symbol layout based on actual observation perspectives, ensuring symbol visibility and comprehensibility under different display conditions and enhancing the usability and functionality of the symbol system. By setting an interactive annotation mode between the symbol dynamic adjustment dataset and the user interaction parameter set and constructing an interactive annotation model, the accuracy and efficiency of symbol annotation are effectively improved, ensuring dynamic updating and real-time reflection of symbol information. Furthermore, by performing symbol annotation tasks and obtaining interactive annotation datasets, we provide data support for subsequent symbol operations, ensuring the accuracy and consistency of the annotation results. Finally, through dynamic visualization rendering, the interactive annotation dataset is displayed in real time, further enhancing the visibility and interactivity of the symbol data. This makes the final dynamic symbol view more intuitive and clear in practical applications, enhancing the operability and value of the data.

[0068] Optionally, this specification further provides a feature point automatic labeling system based on point cloud data, which is used to execute the feature point automatic labeling method based on point cloud data as described above. The feature point automatic labeling system based on point cloud data includes:

[0069] Point cloud fusion module, used to obtain regional multi-source point cloud data and perform multimodal data fusion on regional multimodal point cloud data to obtain point cloud fusion data;

[0070] The ground point cloud segmentation module is used to segment the shoreline structure ground point cloud based on the point cloud fusion data to obtain a ground point cloud segmentation dataset;

[0071] The salient point recognition module is used to perform edge detection on the ground point cloud segmentation dataset and identify geometric salient points based on the edge detection results to obtain a candidate set of feature points;

[0072] The feature point symbolization module is used to classify the ground object type based on the feature point candidate set to obtain the target classification data set; it assigns symbol styles to the target classification data set and binds point cloud feature attributes in combination with the predefined symbol style library to obtain the symbolized data set;

[0073] The dynamic interactive annotation module is used to dynamically adjust and interactively annotate multi-level symbols based on symbolic datasets to obtain a dynamic view of the symbols.

[0074] The automatic feature point labeling system based on point cloud data of the present invention can implement any one of the automatic feature point labeling methods based on point cloud data of the present invention, and is used to combine the operations and signal transmission media between various modules to complete the automatic feature point labeling method based on point cloud data. The internal modules of the system cooperate with each other, so that the symbolic expression of the onshore feature points is not only accurate but also highly interactive. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Other features, objects and advantages of the present invention will become more apparent upon reading the detailed description of non-limiting embodiments thereof made with reference to the following drawings:

[0076] Figure 1 Schematic diagram of the steps of the method for automatically labeling feature points based on point cloud data of the present invention;

[0077] Figure 2 Detailed step flow diagram of step S1 in the present invention;

[0078] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0079] The following is a clear and complete description of the technical method of the present invention in conjunction with the accompanying drawings. It is obvious that the embodiments described are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts are within the scope of protection of the present invention.

[0080] In addition, the accompanying drawings are merely schematic illustrations of the present invention and are not necessarily drawn to scale. Identical reference numerals in the figures denote identical or similar parts, and thus repetitive descriptions thereof will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities that do not necessarily correspond to physically or logically separate entities. These functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor and / or microcontroller approaches.

[0081] It should be understood that although the terms "first," "second," and the like may be used herein to describe various elements, these elements should not be limited by these terms. These terms are used solely to distinguish one element from another. For example, a first element may be referred to as a second element, and similarly, a second element may be referred to as a first element, without departing from the scope of the exemplary embodiments. The term "and / or" as used herein includes any and all combinations of one or more of the listed associated items.

[0082] To achieve this, please refer to Figures 1 to 2 The present invention provides a method for automatically labeling feature points based on point cloud data, the method comprising the following steps:

[0083] Step S1: Acquire regional multi-source point cloud data, and perform multi-modal data fusion on the regional multi-modal point cloud data to obtain point cloud fusion data;

[0084] In this embodiment, a variety of sensors (laser scanners, drones, satellite images, etc.) are used to obtain regional multi-source point cloud data, and coordinate alignment and data format standardization are performed. Set the point cloud coordinate transformation matrix Perform multi-view ICP (Iterative Closest Point) registration to align point clouds from different sources in the global coordinate system. ICP convergence threshold m, with a maximum number of iterations of 50. The aligned point cloud data is subjected to noise filtering, and outliers are removed using the Statistical Outlier Filter (SOR) method. The local neighborhood point count threshold N = 50 and the standard deviation filter threshold a = 2.0 can be set. Density equalization filtering is used to adjust the point cloud density, with a spatial radius r = 0.2m for homogenization. Finally, a density clustering-based fusion method (such as DBSCAN, with a density threshold D = 10) is used to remove redundant data, complete multimodal point cloud fusion, and generate fused point cloud data.

[0085] Step S2: Segment the shoreline structure ground point cloud based on the point cloud fusion data to obtain a ground point cloud segmentation dataset;

[0086] In this embodiment, elevation normalization is performed based on the point cloud fusion data. An elevation histogram of the point cloud data is calculated, and an elevation anomaly threshold is set to 2.0m to remove floating points and local outliers. Cloth simulation filtering (CSF) is used to extract ground points. The cloth relaxation coefficient can be set to 0.5 and the gravity factor to 9.81m / s² to ensure that the cloth simulation process closely matches the actual terrain. This generates a preliminary ground point cloud dataset. Planar regression analysis is then performed, using RANSAC to fit the ground plane. The in-plane point threshold is set to 300 and the residual threshold to 0.05m to remove local floating points and sudden changes. This ultimately results in a ground point cloud segmentation dataset.

[0087] Step S3: performing edge detection on the ground point cloud segmentation dataset, and identifying geometrically significant points based on the edge detection results, thereby obtaining a candidate set of feature points;

[0088] In this embodiment, edge detection is performed on the ground point cloud segmentation dataset. Using the Canny edge detection algorithm, the gradient magnitude and direction are calculated on the 3D point cloud. A low threshold of 0.02 and a high threshold of 0.08 are set to filter edge points, generating an edge point cloud dataset. The normal vector of each edge point is then calculated. A field search radius of 0.3m is set, and the normal vector deviation is calculated. Low-feature points are filtered using a deviation threshold of 15∘ to generate normal deviation point cloud data. The principal curvature and curvature gradient are calculated, and a curvature threshold of 0.1 is set to filter low-curvature points, generating curvature-enhanced point cloud data. Based on the Harris 3D corner detection algorithm, a feature response threshold of 0.01 is set to extract the local maximum feature point, generating a preliminary set of salient points. Finally, density clustering (e.g., DBSCAN, with a radius of 0.5m and a minimum number of neighborhood points of 10) is performed on the preliminary set of salient points to remove isolated points. A geometric consistency threshold of 0.8 is set to eliminate low-consistency points, generating a candidate set of feature points.

[0089] Step S4: classify the feature types based on the feature point candidate set to obtain a target classification dataset; assign a symbol style to the target classification dataset and bind point cloud feature attributes in combination with a predefined symbol style library to obtain a symbolized dataset;

[0090] In this embodiment, the object classification is performed based on the obtained feature point candidate set, and the global and local features of the object are extracted, including elevation distribution, normal vector angle, point cloud density and surface roughness, to generate a statistical table of object features. The support vector machine (SVM) classifier is used for classification, and the RBF kernel function is adopted. The regularization parameter is set to 10 and the kernel width is set to 0.1, and the target classification data set is finally obtained. The symbol style matching of the target classification data set is performed in combination with the predefined symbol style library (containing 10 different colors and 5 shapes), and adaptive adjustment is performed based on the point cloud attributes. The symbol scale adjustment factor is set to [0.5, 2.0] and the color mapping parameter is set to [0, 255] (RGB color space). The symbol shape is adjusted according to the height, density and classification confidence of the object to generate an optimized symbol mapping data set. The point cloud and symbol matching constraint is set to 0.05m to bind the point cloud and the symbol, and finally a symbolized data set is generated.

[0091] Step S5: Perform multi-level symbol dynamic adjustment and interactive annotation based on the symbolized dataset to obtain a symbol dynamic view.

[0092] In this embodiment, multi-level symbol dynamic adjustment is performed based on the obtained symbolized data set. The symbol transparency is adjusted according to the confidence of the feature classification, and the transparency mapping parameter is set to [0.3, 1.0] (where 0.3 represents low confidence and 1.0 represents high confidence). Symbol adaptive adjustment is performed based on user interaction feedback, and the interaction update step is set to 0.1 to gradually optimize the symbol position, size and shape. A gradient descent-based optimization method is used to minimize the interactive annotation error, and the error convergence threshold is set to 0.01 to improve the accuracy of the interactive annotation. Finally, a dynamic view of the symbol is generated, which can be used for dynamic interaction and multi-level analysis in a GIS or point cloud visualization platform.

[0093] Optionally, step S1 specifically includes:

[0094] Step S11: acquiring regional multi-source point cloud data, and performing coordinate transformation and time stamp alignment on the regional multi-source point cloud data to obtain multi-source point cloud data to be processed;

[0095] In this embodiment, a variety of sensors (laser scanners, drones, satellite images, etc.) are used to obtain regional multi-source point cloud data, and coordinate conversion and time stamp alignment are performed on them. First, the various point cloud data are converted to a unified geographic coordinate system (such as WGS84 or UTM) using the quaternion transformation matrix. Rotation correction is performed to ensure consistency of the point cloud in the global coordinate system. For timestamp alignment, linear interpolation is used for time synchronization. A time interval of Δt = 0.02s is set, and linear time interpolation is performed on the point cloud data to eliminate drift errors caused by acquisition times of different devices. Time synchronization errors are controlled within ±5ms to ensure temporal consistency of the data. Ultimately, the multi-source point cloud data to be processed is obtained, where the point cloud data format is converted to PCD (Point Cloud Data) or LAS format, and the location information (x, y, z), reflection intensity (Intensity), and timestamp (Timestamp) of each point are recorded.

[0096] Step S12: performing noise filtering and voxel downsampling on the multi-source point cloud data to be processed to obtain denoised point cloud data;

[0097] In this embodiment, noise filtering and voxel downsampling are performed on the multi-source point cloud data to be processed to reduce data redundancy and improve the efficiency of subsequent processing. First, the Statistical Outlier Removal (SOR) method is used to remove outliers, and the neighborhood point threshold k=50 and the standard deviation multiplier α=1.5 are set to ensure that only high-confidence point cloud data are retained. Then, the voxel grid downsampling method is used to sparse the point cloud data, and the voxel size v=0.1m is set to ensure that the key geometric information can still be maintained after downsampling. The voxel center point is calculated using the mean method, that is, for all points in the voxel , calculate the center point for: ;in is the number of point clouds within a voxel. Finally, the denoised point cloud data is obtained, and its point cloud density is kept above 1000 points / m² to ensure sufficient ground object details.

[0098] Step S13: performing point cloud registration and coordinate alignment on the denoised point cloud data to obtain aligned point cloud data;

[0099] In this embodiment, point cloud registration is performed based on denoised point cloud data to eliminate the posture deviation collected by different sensors. First, a coarse registration method is used to perform preliminary registration using a 3D RANSAC algorithm based on feature points. The minimum number of matching points M = 500 and the number of iterations I = 1000 are set, and the FPFH (Fast Point Feature Histogram) descriptor is used for feature matching to generate a preliminary registration transformation matrix. Then, a fine registration method is used to optimize the preliminary registration transformation matrix using the ICP (Iterative Closest Point) algorithm. The maximum number of iterations is set to 50 and the convergence threshold is set to 100. m and the maximum corresponding point error is 0.05m to ensure accurate alignment. Finally, the optimized transformation matrix is ​​expressed as: ;in, is the rotation matrix, is the translation vector. Finally, the aligned point cloud data is obtained, and its root mean square error (RMSE) is controlled within 0.01m to ensure the registration accuracy.

[0100] Step S14: identifying geometric differences of multi-source point cloud data based on the aligned point cloud data, and performing local morphological reconstruction optimization on the geometric differences of the multi-source point cloud data to obtain standardized point cloud geometric data;

[0101] In this embodiment, based on the aligned point cloud data, the geometric differences of multi-source point cloud data are identified and local morphological reconstruction optimization is performed. First, the Euclidean distance between different point cloud data sets is calculated and the point cloud geometric difference matrix is ​​constructed: ;in, and Represent the coordinate vectors of corresponding points in the two point cloud datasets. The geometric difference threshold is set to 0.05m to filter out abnormal areas. For the identified geometric difference areas, the local surface fitting method is used to optimize the morphology. Using quadratic surface fitting (Quadratic Surface Fitting), the local surface equation is constructed: ; Among them, the fitting coefficient The least squares estimation method is used to set the fitting window size w = 10 neighborhood points to ensure that the surface curvature remains smooth after optimization. Finally, the standardized point cloud geometry data is generated, and its local error is controlled within 0.02m.

[0102] Step S15: performing point cloud data fusion on the aligned point cloud data according to the standardized point cloud geometry data to obtain point cloud fusion data.

[0103] In this example, data fusion is performed on the aligned point cloud data based on standardized point cloud geometry. First, the KNN (K-Nearest Neighbors) method is used to adjust the point cloud density, setting the nearest neighbor search radius r = 0.2m to complete the data points in sparse areas. Then, the Poisson surface reconstruction algorithm is used to optimize the point cloud as a whole, constructing the Poisson equation: ;in, is the point cloud normal vector, is the density field. By solving the Poisson equation, a continuous 3D surface is generated to remove boundary noise during the point cloud fusion process. The Poisson reconstruction depth is set to d = 10, and the smoothing factor is set to s = 0.5 to ensure that the fused point cloud surface is smooth and retains detail. Finally, a KD-Tree is used to optimize the point cloud index to accelerate data retrieval, and the fused point cloud data is stored in LAS or PLY format for subsequent analysis and visualization. The generated point cloud fusion data has a high density (greater than 5000 points / m²), and the fusion error is controlled within 0.01m to ensure high-precision spatial information consistency.

[0104] Optionally, step S12 is specifically as follows:

[0105] Step S121: performing outlier detection on the multi-source point cloud data to be processed, setting the local area point count threshold to [10, 100] and the mean distance standard deviation to [1.0, 3.0] to filter the point cloud subset and obtain a preliminary denoised point cloud dataset;

[0106] In this embodiment, outlier detection is performed on the multi-source point cloud data to be processed to remove isolated points and abnormal points introduced by measurement errors. The statistical outlier removal (SOR) method is used to calculate the mean distance of points in the local area and remove abnormal points based on the standard deviation threshold. Set the neighborhood point count threshold , that is, for each point, calculate its nearest The mean Euclidean distance between neighbors , and set the mean distance standard deviation multiplier . Eliminate points that meet the following conditions: ;in, From a point in the point cloud to its The mean distance between neighbors, is the global standard deviation. This method ensures that the data in high-density areas are retained while effectively removing isolated noise points. For example, in the UAV-LS (UAV LiDAR) dataset, , , the removal rate reaches 5%~10%, and finally a preliminary denoised point cloud dataset is obtained, with the proportion of noise points reduced to less than 2%.

[0107] Step S122: performing density equalization processing on the preliminary denoised point cloud dataset, setting the spatial radius to [0.1, 2.0] to remove sparse noise points, and generating a density-optimized point cloud dataset;

[0108] In this embodiment, the generated preliminary denoised point cloud dataset is subjected to density balancing to balance the data distribution of high-density and low-density areas and reduce the deviation caused by uneven sampling. Adaptive Mean Filter (AMF) is used to calculate the local density of the point cloud and set the spatial radius. Perform density redistribution. The specific method is as follows: for each point , calculate its radius The number of neighboring points within And set the global target density .like , then randomly remove points to reduce local density; if , then use KNN interpolation method to add new points. In the point cloud dataset, set m, The point cloud density in each sparse area is increased by 20%, and the point cloud in the high-density area is reduced by 15%. Finally, a density-optimized point cloud dataset is obtained to achieve uniform distribution.

[0109] Step S123: constructing a three-dimensional voxel grid based on the density-optimized point cloud dataset, setting the voxel size to [0.01, 0.5] for spatial discretization, and performing voxel center mapping on the point cloud within each voxel to obtain a voxelized downsampled point cloud dataset;

[0110] In this embodiment, based on the generated density-optimized point cloud dataset, a 3D voxel grid is constructed and downsampled to reduce the amount of computation while maintaining the structural features of the point cloud. First, the point cloud is divided into equidistant cubic voxels and the voxel size is set. m, to control the downsampling accuracy. The voxel center mapping method is used, that is, for each voxel Point set within Calculate the center point: ;in, is the number of points in a voxel. In point cloud data processing, set , which reduces the number of point clouds by more than 80%, while ensuring that the error is less than 0.05m, and finally obtains the voxelized downsampled point cloud dataset.

[0111] Step S124: adjusting the voxel spacing in the voxelized downsampled point cloud dataset to [0.5, 2.0], and performing edge detail optimization to obtain a uniformly distributed point cloud dataset;

[0112] In this embodiment, for the generated voxel downsampled point cloud dataset, the voxel spacing is adjusted and the edge details are optimized to reduce the information loss caused by downsampling. First, the voxel spacing is set , using Optimal KNN Interpolation (OKI) to complete the point cloud in low-density areas, ensuring that key edge information is not lost. Edge optimization uses a curvature-based boundary refinement method. Calculate the local curvature of each point: ;in, , , The eigenvalues ​​calculated for PCA, The points are marked as boundary points and curvature guided interpolation is performed. In the optimization of building edge point cloud, , which reduces the boundary error by 15% and finally obtains a uniformly distributed point cloud dataset.

[0113] Step S125: performing coordinate normalization processing on the uniformly distributed point cloud data set, and correcting the direction of the point cloud normal vector to generate denoised point cloud data.

[0114] In this embodiment, the generated uniformly distributed point cloud dataset is subjected to coordinate normalization processing, and the direction of the point cloud normal vector is corrected to ensure the consistency of the data. First, the point cloud coordinates are normalized using the mean normalization method, and the normalization range is set to [-1 1], the conversion formula is as follows: ;in, and are the mean and standard deviation of the point cloud coordinates, respectively, to ensure that the point cloud coordinate range is within a controllable range. For normal vector correction, the PCA normal estimation algorithm is used to calculate the local normal vector of each point. , and adjust the normal consistency according to the main direction. If the angle between the normal vectors of two adjacent points Exceeding the threshold , then flip the direction: ; In bridge structure point cloud correction, set , ensuring that the normal consistency is improved by 90%, and finally generating denoised point cloud data, which can be used for high-precision point cloud reconstruction and modeling.

[0115] Optionally, step S2 is specifically:

[0116] Step S21: performing elevation histogram statistical analysis on the point cloud fusion data, calculating the elevation frequency histogram to identify elevation anomaly areas, and performing outlier detection on the elevation anomaly areas to generate elevation normalized point cloud data;

[0117] In this embodiment, the elevation histogram statistical analysis is performed on the point cloud fusion data. First, the elevation information z of the point cloud is extracted and the elevation frequency histogram is constructed. The elevation interval width is set to Δz = 0.1m, and the number of points f(z) in each elevation interval is calculated. By analyzing the elevation distribution, abnormal elevation areas are identified. These areas are usually characterized by extreme values ​​or mutations where the elevation values ​​deviate from the normal distribution. The outlier detection algorithm (such as the detection based on z-score) is used to screen the elevation abnormal areas, and the z-score threshold is set. As an abnormal standard, when the elevation of a point exceeds the mean value When the standard deviation is greater than 1, the point is considered abnormal. Finally, by removing these abnormal values, the elevation normalized point cloud data is obtained to ensure that the point cloud elevation data conforms to the normal geographical distribution. For example, in mountainous terrain data, Filter outlier points with a removal rate of approximately 5% to 10%.

[0118] Step S22: extracting ground points from the elevation-normalized point cloud data based on a cloth simulation filtering algorithm, performing cloth simulation iteration to approximate the terrain surface, separating ground points from non-ground points, and generating a preliminary ground point cloud dataset;

[0119] In this embodiment, ground point extraction is performed from elevation-normalized point cloud data using a cloth simulation filtering algorithm. This cloth simulation filtering utilizes the principle of simulating cloth forces to approximate the terrain surface through an iterative algorithm, gradually separating ground points from non-ground points. In each iteration, the cloth simulation's elasticity parameter E and gravity parameter g are set, and the ground is gradually approximated based on the local elevation characteristics of the point cloud data. Initially, a high initial cloth tension is set to accommodate areas with large elevation variations. The tension is then gradually reduced to ensure a smoother approximation. The goal of cloth simulation is to gradually approximate the ground point cloud to the actual ground through simulated deformation. Cloth simulation can effectively handle complex terrain in mountainous areas and urban environments. Ultimately, the cloth simulation algorithm outputs a preliminary ground point cloud dataset. This data is repeatedly simulated to approximate the ground, eliminating non-ground points that do not belong to the ground area.

[0120] Step S23: performing plane regression analysis on the preliminary ground point cloud dataset, setting the plane inlier threshold to [10, 500] and the inlier residual threshold to [0.001, 0.05], removing floating points and abnormal points with local elevation mutations, and generating an optimized ground point cloud dataset;

[0121] In this embodiment, a plane regression analysis is performed on the preliminary ground point cloud data set to remove floating points and abnormal points with local elevation mutations. The ground point cloud data is fitted with a plane using the Least Squares Method (LSM). The threshold of points in the plane is set. , that is, the minimum number of points that each plane model needs to include in the plane regression analysis; set the internal point residual threshold , which is the maximum allowable distance between the point cloud and the fitted plane. When the residual between the point cloud data point and the fitted plane model exceeds the threshold, the point is considered an outlier and needs to be removed. In particular, for the processing of local elevation mutation areas and suspended points, in areas with large terrain undulations, a smaller residual threshold (such as ), while a larger residual threshold is used in flat areas (such as Finally, optimizing the ground point cloud dataset removes local outliers and preserves the true ground structure. This step is particularly useful in urban and plateau areas, ensuring the removal of artificial or noisy data.

[0122] Step S24: Calculate the local slope and curvature of the optimized ground point cloud dataset, and mark the partitions according to the local slope and curvature to generate a terrain feature enhanced point cloud dataset;

[0123] In this embodiment, based on the optimized ground point cloud dataset, the local slope and curvature of each point are calculated to further enhance the terrain features. Every three adjacent points in the point cloud The slope is calculated by using the three-point method to calculate the normal vector of each point and the angle with the horizontal plane. The slope calculation formula is: ; Local curvature The least squares fitting plane of each point neighborhood is obtained to reflect the curvature of the terrain. Set the thresholds of slope and curvature, such as , marking areas with strong terrain features, such as slopes and hills. This information enhances the topographic representation of the point cloud data by marking and reinforcing local terrain features, making the recognition of complex terrain more accurate. After calculating slope and curvature, the resulting terrain-enhanced point cloud dataset effectively improves the point cloud's terrain understanding capabilities.

[0124] Step S25: Segment the shoreline structure point cloud based on the terrain feature enhanced point cloud dataset, set the neighborhood search radius to [0.5, 5.0] and the terrain similarity threshold to [0.5, 1.0] for structure segmentation, and generate a ground point cloud segmentation dataset.

[0125] In this embodiment, the point cloud dataset is enhanced based on the terrain features to perform shoreline structure point cloud segmentation. First, the neighborhood search radius is set. , perform neighborhood search for each point and calculate the similarity of terrain features of its neighboring points. By calculating the terrain similarity measure , which considers the similarity of terrain features such as slope and curvature: ; When the similarity measure Exceeding the set threshold When the points are in the same terrain area, the two points are considered to belong to the same terrain area. Hierarchical clustering is used to segment the point cloud. The segmented structure area is used to identify terrain features such as coastline and ridgeline. When segmenting the point cloud of coastline and urban area, , an accurate ground structure segmentation dataset is obtained, and the segmented ground point cloud area is of great significance for further 3D modeling and geographic analysis.

[0126] Optionally, step S22 is specifically as follows:

[0127] Step S221: Grid-segment the elevation normalized point cloud data, set the grid size to [0.1, 5.0], construct a grid point cloud distribution matrix, and calculate the lowest elevation point of each grid cell to generate an initial terrain skeleton point cloud dataset;

[0128] In this embodiment, the elevation-normalized point cloud data is gridded and segmented. First, the grid size is set to [0.1, 5.0] to meet the resolution requirements of different regions. During the gridding process, a smaller grid size (such as 0.5m) is selected for fine-grained areas (such as urban and mountainous areas), while a larger grid size (such as 3.0m) is suitable for flat or large areas. By mapping the spatial coordinates (x, y, z) of the point cloud data, a grid point cloud distribution matrix is ​​obtained. The point cloud within each grid cell is statistically analyzed based on its elevation value, and the lowest elevation point within each grid cell is selected as the representative of the grid, called the grid lowest point. In this way, a preliminary terrain skeleton for the region can be extracted. This skeleton point cloud dataset can be used for subsequent terrain feature extraction and ground point extraction. For example, in urban data, the lowest point of each grid cell typically represents the foundation of a building, facilitating subsequent processing. The resulting initial terrain skeleton point cloud dataset accurately captures the main trends of ground changes.

[0129] Step S222: Initialize the cloth simulation on the initial terrain skeleton point cloud dataset, set the cloth elastic coefficient to [0.1, 1.0] and the gravity factor to [9.0, 9.8], calculate the cloth contact points using an iterative approximation method, and generate terrain surface fitting point cloud data;

[0130] In this embodiment, a cloth simulation method is used to fit the terrain surface based on the initial terrain skeleton point cloud dataset. First, the cloth elastic coefficient E∈[0.1,1.0] is set to control the tension of the cloth simulation. A smaller elastic coefficient is suitable for complex or variable terrain, while a larger elastic coefficient is suitable for flatter areas. The gravity factor g∈[9.0,9.8] is set to the Earth's standard gravitational acceleration. Through the cloth simulation algorithm, the changes in the cloth under the action of gravity are simulated in each iteration, and the matching of each cloth contact point with the original point cloud is calculated. During the simulation process, the cloth nodes adjust their positions to approximate the terrain surface, thereby generating a surface fitting model in three-dimensional space that is closer to the actual terrain. Through repeated approximation of the cloth simulation, an accurate terrain surface fitting point cloud dataset is ultimately formed. This method has high adaptability and stability for complex terrain such as mountains, rivers, and urban environments. The final terrain surface fitting point cloud can be further used for accurate terrain analysis and ground point segmentation.

[0131] Step S223: Calculate the elevation deviation between the terrain surface fitting point cloud data and the elevation normalized point cloud data, set the ground fitting error threshold to [0.01, 0.5], remove non-ground points that exceed the error threshold, and obtain a candidate set of ground points;

[0132] In this embodiment, the elevation deviation of each point is calculated between the terrain surface fitting point cloud data set and the elevation normalized point cloud data to identify ground points and non-ground points. Specifically, for each point Calculate its fitted elevation and actual elevation The deviation between , and set the elevation fitting error threshold . when deviation When the threshold is exceeded, the point is considered a non-ground point and is removed from the dataset. The threshold setting for elevation deviation depends on the accuracy requirements of the data and the complexity of the terrain. In mountainous or urban areas, a larger error threshold can be set (e.g. ) to adapt to the undulation of the ground; in relatively flat areas, the error threshold is appropriately reduced (such as ) can improve the accuracy of ground points. The candidate ground points screened by this method will effectively exclude elevation anomalies and non-ground area points.

[0133] Step S224: Density filtering is performed on the ground point candidate set, the density threshold is set to [1, 10], isolated points and sparse points are removed, and an optimized ground point cloud dataset is generated;

[0134] In this embodiment, after the ground point candidate set is generated, the point cloud data is further subjected to density filtering. Density filtering is used to remove isolated points and sparse points to ensure the continuity and integrity of the data. To this end, the density threshold of the point cloud data is set. , that is, there must be at least other points in the neighborhood to be considered a valid point. When the number of points in the neighborhood is less than the threshold, the point will be considered a sparse point or a noise point and removed from the dataset. The selection of the density threshold depends on the terrain characteristics of the area and the distribution density of the point cloud data. For example, in cities and dense areas, a larger density threshold (such as ), while in open areas, the density threshold should be appropriately lowered (such as 2). Through this step, the optimized ground point cloud dataset will remove isolated noise points, retain more representative ground data, and further improve data quality.

[0135] Step S225: Perform connected area detection on the optimized ground point cloud dataset, set the connectivity threshold to [0.5, 1.0], remove floating points and unconnected ground segments, and generate a preliminary ground point cloud dataset.

[0136] In this embodiment, the connected area detection is performed on the optimized ground point cloud dataset, and the connectivity threshold is used Determine the connectivity between ground points. For each pair of points and , calculate the Euclidean distance between them ,if If the connectivity threshold is less than the set value, the two points are considered to be connected. The purpose of connected area detection is to remove floating points and disconnected ground segments to ensure that the generated ground point cloud dataset has high continuity and integrity. The choice of should be adjusted according to the distribution characteristics of the point cloud data and the complexity of the ground structure. For example, in an urban environment, a larger threshold It is suitable for relatively regular ground distribution, while in complex natural terrain, it is more appropriate to use a smaller threshold (such as 5) Connected ground segments can be better detected. Ultimately, the generated preliminary ground point cloud dataset will contain real ground points and remove all disconnected ground areas and non-ground points.

[0137] Optionally, the step S3 of identifying geometrically salient points specifically includes:

[0138] Calculate the normal vector based on the edge detection results, set the field search radius to [0.1, 2.0], calculate the normal vector of each edge point and the angle deviation between the edge point and the neighboring point, and obtain the normal vector deviation point cloud data;

[0139] In this embodiment, after edge detection is performed on the point cloud data, the area search radius is set. , to determine the density and influence range of points in the neighborhood. By calculating the normal vector of each edge point, selecting points within the neighborhood, and fitting the neighborhood point cloud using the principal component analysis (PCA) method, the normal vector of each edge point is obtained. This normal vector represents the direction of the local plane of the point. Then, by calculating the angle deviation between the normal vector of the edge point and the normal vector of its neighborhood point A large deviation value indicates that there is a large difference in geometry between the point and its neighborhood. Set the threshold of angle deviation The threshold is [0.01, 0.2], and points with a value greater than this threshold are marked to generate normal vector deviation point cloud data. This can effectively distinguish points with large local geometric changes, providing a basis for subsequent curvature calculation and feature point extraction.

[0140] Calculate the principal curvature and curvature gradient of the normal vector deviation point cloud data, set the curvature threshold to [0.001, 0.1] to filter out low curvature points, and generate curvature enhanced point cloud data;

[0141] In this embodiment, the principal curvature and curvature gradient of each point are calculated based on the normal vector deviation point cloud data. and secondary curvature It is determined by the change of the point cloud in the local area and can be obtained by solving the eigenvalue of the curvature tensor matrix. Set the curvature threshold , filtering out low curvature points, that is, points whose main curvature is less than the set threshold. The height of the curvature directly affects the complexity of the terrain, so the range of the curvature threshold is used to remove irrelevant points in relatively flat areas (low curvature areas) and retain those points with strong geometric features in the local space. In the process of generating curvature enhanced point cloud data, smaller curvature values ​​will be excluded to ensure that significant terrain features with high curvature are retained in the dataset. For example, in mountainous areas, the curvature is large, and the points filtered are mainly in low slope areas. The final curvature enhanced point cloud dataset will contain more points that are closely related to terrain changes, enhancing the expressiveness of the point cloud.

[0142] Perform feature response analysis on the curvature-enhanced point cloud data. Based on the Harris 3D corner detection algorithm, set the response threshold to [0.01, 0.5], extract the local maximum feature point, and generate a preliminary salient point set.

[0143] In this embodiment, after generating the curvature enhanced point cloud dataset, the Harris 3D corner detection algorithm is used to perform feature response analysis. Through this algorithm, the corner response value of each point is first calculated based on the feature response function of the local neighborhood, reflecting the geometric changes of the point in the local area. Covariance matrix in the neighborhood. Set the neighborhood radius , take a point All points in the neighborhood of (i.e. distance and The distance between For each point , calculate its intersection with point The position difference and normal vector difference of are used to construct the local covariance matrix: ;in, Indicates a point Neighborhood, and Yes and The coordinates of the covariance matrix Perform eigenvalue decomposition to obtain eigenvalues (corresponding to the three eigenvectors of the covariance matrix). These eigenvalues ​​represent the local geometric information of the point in each direction. According to the size of the eigenvalue, the curvature and geometric significance of the point can be described. Using the eigenvalue Calculate Harris response value , indicating a point The standard Harris corner response function is: ; is a matrix The determinant of , reflects the local curvature of the point. is a matrix The trace of , which is equal to the sum of the eigenvalues, reflects the overall local change of the point. Is a constant, usually ranging from 0.04 to 0.06. Indicates the significance of the point, the higher Indicates that there is a strong geometric change in the local area of ​​the point, which may be a corner point or an edge point. Set the response threshold (like , used to filter out significant points from the calculated response function. , then the point is considered a significant point. Set the response threshold , when the response value of a point exceeds the threshold, the point is considered a potential salient point. This step extracts points with large geometric changes by detecting the local maximum response value. For example, in areas such as building corners and road intersections, the Harris algorithm can effectively identify points with obvious local shape changes. These salient points usually represent key features in the point cloud, such as the edges and corners of buildings and the intersections of roads. Finally, the local maximum feature points extracted by the Harris algorithm form a preliminary salient point set.

[0144] Perform spatial distribution analysis on the preliminary salient point set, set the cluster radius to [0.05, 1.0] and the minimum number of neighborhood points to [3, 20], remove isolated points, and generate an optimized salient point set;

[0145] In this embodiment, the spatial distribution analysis of the preliminary salient point set is performed and the clustering radius is set. and the minimum number of neighborhood points , in order to determine whether each salient point belongs to a reasonable local area. In this process, the DBSCAN (density clustering) method is used to set a suitable radius. , analyze whether the points in the salient point concentration form a reasonable dense area in space. If the number of points in the neighborhood of a point is less than the set minimum number of neighborhood points , the point is considered an isolated point and will be removed. This step can remove points that are not spatially significant or clustered, such as noise points or occasional non-salient points. In the process of optimizing the salient point set, reasonably setting the cluster radius and number of neighborhood points can more accurately retain representative salient points.

[0146] Based on the optimized salient point set, geometric consistency analysis is performed, the geometric structure consistency threshold is set to [0.5, 1.0], low consistency points are removed, and finally a candidate set of feature points is generated.

[0147] In this embodiment, geometric consistency analysis is performed on the optimized salient point set to further improve the accuracy and reliability of the data. , which is used to determine whether the geometric features of each point are consistent with the structure of the surrounding points. By calculating the geometric structural similarity between the significant point and its neighborhood points, such as the local plane fitting error, its geometric consistency is evaluated. If the geometric consistency of the point is lower than the set threshold, the point will be eliminated. The purpose of this step is to remove points that are inconsistent in geometric structure, such as some error points that do not conform to the expected terrain changes. Through geometric consistency analysis, the final generated feature point candidate set will contain points with significant geometric changes and high structural consistency, which are suitable for further modeling and analysis. For example, if the data set contains a large range of roads and buildings, geometric consistency analysis can remove noise points that do not belong to the main structure and retain key feature points. The final generated feature point candidate set will provide accurate and reliable geometric feature data for subsequent applications.

[0148] Optionally, the ground feature type classification described in step S4 is specifically:

[0149] Perform local geometric feature extraction on the feature point candidate set, extract the surface normal vector, curvature, point density and morphological characteristics of each feature point, and generate a ground feature geometric feature dataset;

[0150] In this embodiment, for each feature point selected from the feature point candidate set, , first we need to extract the local geometric features of the point. This includes surface normal vector, curvature, point density and morphological features. The selected neighborhood radius To define the local neighborhood of each feature point. By calculating all points in the neighborhood The local surface normal vector is obtained and the normal vector is estimated using the weighted least squares method, where the weight is determined by the distance between each point and the center point, and closer points are given a larger weight. The principal curvature of each feature point is calculated using the quadratic surface model that fits the local surface. The principal curvature can be calculated by taking the first and second order derivatives of the quadratic surface fitted by the local surface. Within a given neighborhood radius, the number of points in the neighborhood is counted as an indicator of point density. By calculating the geometric shape of the point cloud in the neighborhood (such as edge angle, convexity, etc.), the morphological features of the point are extracted. These features will form a dataset of geometric features of the ground object. ,in is the feature point position, is the normal vector, is the curvature, is the point density, For morphological characteristics.

[0151] Perform scale space analysis on the ground feature geometric feature dataset, calculate the feature change rate at different scales, and generate a multi-scale ground feature feature dataset;

[0152] In this embodiment, after obtaining the ground feature geometric feature dataset, scale space analysis is performed to evaluate the feature changes of the ground features at different scales. To this end, by changing the neighborhood radius To create feature point sets at different scales, and thus calculate the feature change rate at different scales. By continuously adjusting the neighborhood radius , extract the local geometric features (such as normal vector, curvature, etc.) of each feature point and calculate the rate of change of each feature at different scales. The rate of change can be calculated by the difference in features between adjacent scales, for example: ;in and Scale and By summarizing the calculation results of these change rates, a multi-scale ground feature dataset is generated. , used for subsequent classification training.

[0153] Obtain a target feature type database, perform classification training based on the target feature type database, and build a feature classification model;

[0154] In this embodiment, in order to classify the land features, it is necessary to first obtain a feature database of the target land feature type. This database contains a typical feature set of different land feature types (dikes, scarps, slopes, roads, houses, etc.), including normal vectors, curvature, density, morphological features, etc. The database is composed of remote sensing data (satellite images, aerial photography, etc.), LiDAR point cloud data, geographic information system (GIS) database, and field survey data. Based on these data, the geometric features, texture features, spatial features, etc. of each target land feature are extracted, and the extracted feature data are compiled into a database. Target land feature type feature database ,in For the feature type, is the feature set of this type. Use machine learning classification algorithms (such as random forest, support vector machine SVM, deep neural network DNN, etc.) to train the feature type database. The model input is the feature data , the output is the feature type The best classification model is selected through cross-validation, and the model parameters are optimized using a training set from the target feature database. The training process uses multi-scale features to ensure that the model can adapt to features at different scales.

[0155] The multi-scale feature dataset is classified into target feature categories using the feature classification model, and high-confidence classification results are screened to obtain a preliminary feature classification dataset.

[0156] In this embodiment, the trained object classification model is used to classify the multi-scale object feature dataset. In this step, the classification model assigns a feature category label to each feature point. Data set to classification model to get each feature point Corresponding category prediction , and calculate the classification confidence for each point For example, when using random forest, the prediction value is determined by voting of multiple decision trees, and the final result is the category with the most votes. Filter out high confidence classification results, such as These high-confidence classification results are regarded as the final classification results, and the preliminary land feature classification dataset is obtained. .

[0157] The classification error point categories of the preliminary object classification dataset are adjusted to generate the target classification dataset.

[0158] In this embodiment, for the preliminary land feature classification dataset Since there may be misclassification in the classification process, it is necessary to adjust the error point category. Error points mainly refer to those points with low confidence or unclear categories. For points with low confidence in the classification results (such as ), the category adjustment can be performed through neighborhood analysis. The categories of these error points are adjusted according to the categories of the high confidence points in the neighborhood. For example, if most of the points in the neighborhood of a misclassified point belong to the category , then adjust the misclassified point to Category. After adjusting the error points, the target classification data set is generated , which is the final result of ground feature classification, is the adjusted category label.

[0159] Optionally, the allocation symbol style and binding point cloud feature attributes described in step S4 are specifically:

[0160] Based on the target classification data set, calculate the global and local statistical characteristics of each category of ground objects and generate a statistical table of ground object characteristics;

[0161] In this embodiment, after obtaining the target classification data set After that, the global and local statistical characteristics of the feature categories are calculated. , calculate its related statistical features, including: 1) global statistical features: for example, feature categories The maximum, minimum, and average values ​​of the spatial distribution of all points under the given condition are calculated as follows: ;in, For category The average position of For category The total number of midpoints, is the location of the point. 2) Local statistical features: For example, for each feature point , calculate the density of points in its neighborhood, and other local geometric features (such as curvature, change in normal vector, etc.). The spherical neighborhood method can be used to calculate local density: ; These global and local features will be summarized into a feature statistics table , which contains statistics for each feature category.

[0162] Based on the predefined symbol style library and the statistical table of land feature characteristics, the symbol style of the target classification dataset is matched and mapped to obtain a preliminary symbol mapping dataset;

[0163] In this embodiment, the symbol style can be defined in advance based on geographic information standards (such as S-57 / S-100 maritime standards) (such as a rectangle + anchor icon for a dock, and a dotted line for a tide line) to obtain a symbol style library. The predefined symbol style library contains different symbol shapes (such as circles, squares, triangles, etc.), colors (such as red, green, blue, etc.) and symbol scales (such as size ranges). The point cloud feature attributes (type, elevation, confidence) are mapped to symbol color, size, and transparency (such as a red warning symbol for a high-risk area). According to the statistical characteristics of the landform (such as the height and density of the category), it is matched to a specific symbol in the symbol style library. For example, if the average height of the landform category is greater than the set threshold, a larger symbol size is selected for the landform of that category; if the density is higher, a darker color is selected. Assume that according to the statistical table, the category Average height =10 m, density =50, then select a larger symbol and dark green. The mapping of the symbol style library can be done as follows: ; This generates a preliminary symbol map dataset: ,in for point symbol.

[0164] Adaptively adjust the preliminary symbol mapping dataset based on the point cloud feature attributes in the feature statistics table, set the symbol scale adjustment factor to [0.5, 2.0] and the color mapping parameter to [0, 255], adjust the symbol shape according to the height, density and classification confidence of the feature, and generate an optimized symbol mapping dataset;

[0165] In this embodiment, after obtaining the preliminary symbol mapping data set, the symbol style is adaptively adjusted. In order to make the symbol mapping more consistent with the actual data distribution and requirements, the scale, color and shape of the symbol are fine-tuned. Set the symbol scale adjustment factor and colormap parameters Adjust. Adaptively adjust the size of the symbol based on the characteristics of the feature category (such as height, density, classification confidence, etc.). For example, if the height of a certain category of feature is large, the size of the symbol It can be adjusted by the following formula: ;in, is the height adjustment factor, which is usually dynamically calculated based on the average height of the objects. The symbol color can be adjusted based on the density of the objects and the confidence level of the category. For example, if the density of the object category is high, the brightness of the symbol color can be increased appropriately. Color Mapping Parameters Can be based on density To set, for example Through these adaptive adjustments, an optimized symbol mapping dataset is generated. ,in This is the optimized symbol style.

[0166] Bind the optimized symbol mapping dataset to the target classification dataset, set the point cloud and symbol matching constraint to [0.1, 1.0], and generate a symbolic point cloud dataset;

[0167] In this embodiment, after establishing a matching relationship between the optimized symbol mapping dataset and the target classification dataset, the two are bound to generate a symbolized point cloud dataset. , ensuring that each point cloud data and its corresponding symbol style have good spatial consistency. Matching constraint: By calculating the matching degree between point cloud data and symbols, ensure that each point With its symbol The spatial position of the symbol is consistent with the scale, color, etc. of the symbol style. It can be calculated by the following formula: ;in Yes ampersand If the constraints are met, the point cloud and the symbol are considered to be matched successfully. The final symbolized point cloud dataset is generated. , which contains the symbolized point cloud data.

[0168] Perform consistency checks on the symbolized point cloud dataset, detect and correct abnormal mapping points, and finally output the symbolized dataset.

[0169] In this embodiment, the consistency check of the symbolized point cloud dataset is performed to detect and correct abnormal mapping points. Abnormal mapping points refer to points where the symbol style does not match the actual point cloud data. These points may be caused by misclassification or incorrect symbol style settings. Symbol style The symbol style of the points in the neighborhood is consistent to identify abnormal mapping points. The symbol style and If inconsistent, is considered an outlier. The method to correct an outlier is to adjust the symbol style of its neighboring points. For example, if the symbol style of most neighboring points is , then The symbol style is adjusted to Through consistency checking and correction of outliers, a symbolic dataset is finally generated. , this is the final symbolized point cloud dataset.

[0170] Optionally, step S5 is specifically as follows:

[0171] Step S51: Organize the symbolized data set into multiple levels of symbols, construct a symbol hierarchy tree, and obtain a preliminary symbol hierarchy structure;

[0172] In this embodiment, in the symbolized point cloud dataset Based on the data set, a multi-level symbol organization is performed on the data set to construct a symbol hierarchy tree. The purpose of the symbol hierarchy tree is to classify and group symbols by their relevance and priority, so that symbols at different levels can be better displayed and managed during the visualization process. First, the symbols are divided into different levels according to their category, size, color and other characteristics. For example, some land feature categories may be at the top symbol level, while other categories may be at the bottom symbol level. The structure of the symbol hierarchy tree is: ;in For the Layer symbol hierarchy, including different symbol nodes. Each symbol node can represent a symbol of a specific feature category and has related symbol characteristics (such as size, shape, color, etc.).

[0173] Step S52: performing a hierarchical integrity check on the preliminary symbol hierarchy structure to obtain an optimized symbol hierarchy data set;

[0174] In this embodiment, when constructing the preliminary symbol hierarchy tree After that, it is subjected to a hierarchical integrity check. This process verifies whether each symbol node follows the rules of the symbol hierarchy to ensure that the structure of the symbol hierarchy is reasonable and that the hierarchical division of symbols meets the requirements of visual effects and operability. The verification content includes: 1) Symbol priority check: Ensure that the priority of the symbol is set correctly. For example, high-level symbols (such as "buildings" or "transportation facilities") should be at a higher level, while low-level symbols (such as "trees" or "grass") should be at a lower level. 2) Symbol hierarchy depth check: The depth of the hierarchy tree should be set reasonably to avoid visual overload caused by an overly deep hierarchy. 3) Symbol association check: Whether the parent-child relationship of the symbol conforms to the actual distribution and expression of the land features. For example, symbols of certain land feature categories should probably be bound to the same level, and should not span different levels. After verification, an optimized symbol hierarchy dataset is obtained. , which meets the completeness requirements and can effectively improve the symbol rendering and interaction performance in subsequent steps.

[0175] Step S53: obtaining a user interaction parameter set, and performing a user interaction perspective perception simulation on the user interaction parameter set and the optimized symbol hierarchy data set to obtain observation perspective character perception data;

[0176] In this embodiment, the user interaction parameter set is obtained through the 3D geographic interactive platform, which includes the user input parameters such as viewing angle, zoom, rotation, etc. For example, the user can set the viewing angle, zoom ratio, and the type of ground feature to be observed. By simulating these parameters and combining them with the optimization of the symbol level data set, the user can obtain the user interaction parameter set. , to simulate user interaction perspective perception. This simulation generates observation perspective character perception data by calculating the visualization effect of symbols under different interaction perspectives. , which contains information about the visibility, clarity, size change, etc. of each symbol at different viewing angles. The simulation process can be performed using the projection matrix and camera parameters: ;in, Indicates that under given user interaction parameters, the symbol Perspective transformation and perception data.

[0177] Step S54: performing symbol visibility evaluation based on the observation perspective character perception data, and optimizing the symbol layout in the symbol hierarchy dataset according to the symbol visibility to obtain a symbol dynamic adjustment dataset;

[0178] In this embodiment, after obtaining the character perception data of the observation angle After that, the visibility of the symbol is evaluated. During the evaluation process, the visibility, size change and occlusion of the symbol under different viewing angles are mainly considered. Specifically, for each symbol , calculate its visibility at a specific viewing angle , such as whether the symbol is obscured, whether the symbol needs to be dynamically resized to improve visibility, etc. The calculation formula for visibility evaluation is: ;in, Indicates whether the symbol is in view, is the visibility adjustment factor of the symbol, which is adjusted according to the symbol's position and viewing angle. Based on the results of the visibility evaluation, the symbol layout in the symbol hierarchy dataset is optimized to obtain the symbol dynamic adjustment dataset. , which contains dynamic visibility and layout adjustment information for all symbols.

[0179] Step S55: setting an interactive annotation mode between the symbol dynamic adjustment data set and the user interaction parameter set, constructing an interactive annotation model, and performing the symbol annotation task based on the interactive annotation model, thereby obtaining an interactive annotation data set;

[0180] In this embodiment, after obtaining the symbol dynamic adjustment data set After that, set the symbol to dynamically adjust the data set and the user interaction parameter set Interactive annotation mode between them, building an interactive annotation model This model is used to dynamically adjust the display of symbols during user interaction. For example, it updates the position, size, and color of symbols in real time based on user rotation, scaling, and other operations. The core of the interactive annotation model is to dynamically adjust the dataset and update the properties of symbols based on the interactive data and symbols entered by the user. The interactive annotation model works as follows: ;in, This dataset represents a symbol dataset that has been annotated through user interaction. This dataset can accurately reflect the optimal display effect of symbols from the perspective of user interaction.

[0181] Step S56: Perform dynamic visualization rendering on the interactive annotation dataset and set rendering parameters to obtain a dynamic view of the symbol.

[0182] In this embodiment, the interactive annotation dataset Perform dynamic visualization rendering. Set rendering parameters These parameters control the rendering effects of the symbol, including lighting, shadows, and reflections. The rendering process is performed by the graphics rendering engine to generate a dynamic view of the symbol. The rendered symbol dynamic view can reflect the effects of user interaction in real time and adjust the displayed content according to the user's operation. For example, the size of the symbol will scale with the change of viewing angle, and the color of the symbol will adjust according to the changes of light and shadow. The rendering formula is: ; Through dynamic rendering, the final output symbol dynamic view can reflect user interaction operations and symbol display effects in real time.

[0183] Optionally, this specification further provides a feature point automatic labeling system based on point cloud data, which is used to execute the feature point automatic labeling method based on point cloud data as described above. The feature point automatic labeling system based on point cloud data includes:

[0184] Point cloud fusion module, used to obtain regional multi-source point cloud data and perform multimodal data fusion on regional multimodal point cloud data to obtain point cloud fusion data;

[0185] The ground point cloud segmentation module is used to segment the shoreline structure ground point cloud based on the point cloud fusion data to obtain a ground point cloud segmentation dataset;

[0186] The salient point recognition module is used to perform edge detection on the ground point cloud segmentation dataset and identify geometric salient points based on the edge detection results to obtain a candidate set of feature points;

[0187] The feature point symbolization module is used to classify the ground object type based on the feature point candidate set to obtain the target classification data set; it assigns symbol styles to the target classification data set and binds point cloud feature attributes in combination with the predefined symbol style library to obtain the symbolized data set;

[0188] The dynamic interactive annotation module is used to dynamically adjust and interactively annotate multi-level symbols based on symbolic datasets to obtain a dynamic view of the symbols.

[0189] The present invention is therefore intended to be illustrative and non-restrictive in all respects, with the scope of the invention being defined by the appended claims rather than the foregoing description, and all changes that come within the meaning and range of equivalents of the application documents are intended to be embraced therein.

[0190] The foregoing description is intended only to provide specific embodiments of the present invention, which will enable those skilled in the art to understand and implement the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not intended to be limited to the embodiments shown herein, but is to be construed in the widest possible manner consistent with the principles and novel features disclosed herein.

Claims

1. A method for automatic feature point annotation based on point cloud data, characterized in that: The following steps are involved: Step S1: Acquire regional multi-source point cloud data, and perform multi-modal data fusion on the regional multi-modal point cloud data to obtain point cloud fusion data; Step S2: Segment the shoreline structure ground point cloud based on the point cloud fusion data to obtain a ground point cloud segmentation dataset; Step S3: performing edge detection on the ground point cloud segmentation dataset, and identifying geometrically significant points based on the edge detection results, thereby obtaining a candidate set of feature points; Step S4: classify the feature types based on the feature point candidate set to obtain a target classification dataset; assign a symbol style to the target classification dataset and bind point cloud feature attributes in combination with a predefined symbol style library to obtain a symbolized dataset; Step S5: Perform multi-level symbol dynamic adjustment and interactive annotation based on the symbolized dataset to obtain a symbol dynamic view; Step S5 specifically includes: Step S51: Organize the symbolized data set into multiple levels of symbols, construct a symbol hierarchy tree, and obtain a preliminary symbol hierarchy structure; Step S52: performing a hierarchical integrity check on the preliminary symbol hierarchy structure to obtain an optimized symbol hierarchy data set; Step S53: obtaining a user interaction parameter set, and performing a user interaction perspective perception simulation on the user interaction parameter set and the optimized symbol hierarchy data set to obtain observation perspective character perception data; Step S54: performing symbol visibility evaluation based on the observation perspective character perception data, and optimizing the symbol layout in the symbol hierarchy dataset according to the symbol visibility to obtain a symbol dynamic adjustment dataset; Step S55: setting an interactive annotation mode between the symbol dynamic adjustment data set and the user interaction parameter set, constructing an interactive annotation model, and performing the symbol annotation task based on the interactive annotation model, thereby obtaining an interactive annotation data set; Step S56: Perform dynamic visualization rendering on the interactive annotation dataset and set rendering parameters to obtain a dynamic view of the symbol.

2. The method for automatic feature point annotation based on point cloud data according to claim 1, characterized in that: Step S1 is specifically as follows: Step S11: acquiring regional multi-source point cloud data, and performing coordinate transformation and time stamp alignment on the regional multi-source point cloud data to obtain multi-source point cloud data to be processed; Step S12: performing noise filtering and voxel downsampling on the multi-source point cloud data to be processed to obtain denoised point cloud data; Step S13: performing point cloud registration and coordinate alignment on the denoised point cloud data to obtain aligned point cloud data; Step S14: identifying geometric differences of multi-source point cloud data based on the aligned point cloud data, and performing local morphological reconstruction optimization on the geometric differences of the multi-source point cloud data to obtain standardized point cloud geometric data; Step S15: performing point cloud data fusion on the aligned point cloud data according to the standardized point cloud geometry data to obtain point cloud fusion data.

3. The method for automatic feature point annotation based on point cloud data according to claim 2, characterized in that: Step S12 is specifically as follows: Step S121: performing outlier detection on the multi-source point cloud data to be processed, setting the local neighborhood point count threshold to [10, 100] and the mean distance standard deviation to [1.0, 3.0] to filter the point cloud subset and obtain a preliminary denoised point cloud dataset; Step S122: performing density equalization processing on the preliminary denoised point cloud dataset, setting the spatial radius to [0.1, 2.0] to remove sparse noise points, and generating a density-optimized point cloud dataset; Step S123: constructing a three-dimensional voxel grid based on the density-optimized point cloud dataset, setting the voxel size to [0.01, 0.5] for spatial discretization, and performing voxel center mapping on the point cloud within each voxel to obtain a voxelized downsampled point cloud dataset; Step S124: adjusting the voxel spacing in the voxelized downsampled point cloud dataset to [0.5, 2.0], and performing edge detail optimization to obtain a uniformly distributed point cloud dataset; Step S125: performing coordinate normalization processing on the uniformly distributed point cloud data set, and correcting the direction of the point cloud normal vector to generate denoised point cloud data.

4. The method for automatic feature point annotation based on point cloud data according to claim 1, characterized in that: Step S2 is specifically as follows: Step S21: performing elevation histogram statistical analysis on the point cloud fusion data, calculating the elevation frequency histogram to identify elevation anomaly areas, and performing outlier detection on the elevation anomaly areas to generate elevation normalized point cloud data; Step S22: extracting ground points from the elevation-normalized point cloud data based on a cloth simulation filtering algorithm, performing cloth simulation iteration to approximate the terrain surface, separating ground points from non-ground points, and generating a preliminary ground point cloud dataset; Step S23: performing plane regression analysis on the preliminary ground point cloud dataset, setting the plane inlier threshold to [10, 500] and the inlier residual threshold to [0.001, 0.05], removing floating points and abnormal points with local elevation mutations, and generating an optimized ground point cloud dataset; Step S24: Calculate the local slope and curvature of the optimized ground point cloud dataset, and mark the partitions according to the local slope and curvature to generate a terrain feature enhanced point cloud dataset; Step S25: Segment the shoreline structure point cloud based on the terrain feature enhanced point cloud dataset, set the neighborhood search radius to [0.5, 5.0] and the terrain similarity threshold to [0.5, 1.0] for structure segmentation, and generate a ground point cloud segmentation dataset.

5. The method for automatic feature point annotation based on point cloud data according to claim 4, characterized in that: Step S22 is specifically as follows: Step S221: Grid-segment the elevation normalized point cloud data, set the grid size to [0.1, 5.0], construct a grid point cloud distribution matrix, and calculate the lowest elevation point of each grid cell to generate an initial terrain skeleton point cloud dataset; Step S222: Initialize the cloth simulation on the initial terrain skeleton point cloud dataset, set the cloth elastic coefficient to [0.1, 1.0] and the gravity factor to [9.0, 9.8], calculate the cloth contact points using an iterative approximation method, and generate terrain surface fitting point cloud data; Step S223: Calculate the elevation deviation between the terrain surface fitting point cloud data and the elevation normalized point cloud data, set the ground fitting error threshold to [0.01, 0.5], remove non-ground points that exceed the error threshold, and obtain a candidate set of ground points; Step S224: Density filtering is performed on the ground point candidate set, the density threshold is set to [1, 10], isolated points and sparse points are removed, and an optimized ground point cloud dataset is generated; Step S225: Perform connected area detection on the optimized ground point cloud dataset, set the connectivity threshold to [0.5, 1.0], remove floating points and unconnected ground segments, and generate a preliminary ground point cloud dataset.

6. The method for automatic feature point annotation based on point cloud data according to claim 1, characterized in that: The identification of geometrically salient points in step S3 is specifically as follows: Calculate the normal vector based on the edge detection results, set the neighborhood search radius to [0.1, 2.0], calculate the normal vector of each edge point and the angle deviation between the edge point and the neighborhood point, and obtain the normal vector deviation point cloud data; Calculate the principal curvature and curvature gradient of the normal vector deviation point cloud data, set the curvature threshold to [0.001, 0.1] to filter out low curvature points, and generate curvature enhanced point cloud data; Perform feature response analysis on the curvature-enhanced point cloud data. Based on the Harris 3D corner detection algorithm, set the response threshold to [0.01, 0.5], extract the local maximum feature point, and generate a preliminary salient point set. Perform spatial distribution analysis on the preliminary salient point set, set the cluster radius to [0.05, 1.0] and the minimum number of neighborhood points to [3, 20], remove isolated points, and generate an optimized salient point set; Based on the optimized salient point set, geometric consistency analysis is performed, the geometric structure consistency threshold is set to [0.5, 1.0], low consistency points are removed, and finally a candidate set of feature points is generated.

7. The method for automatic feature point annotation based on point cloud data according to claim 1, characterized in that: The ground feature type classification described in step S4 is specifically as follows: Perform local geometric feature extraction on the feature point candidate set, extract the surface normal vector, curvature, point density and morphological characteristics of each feature point, and generate a ground feature geometric feature dataset; Perform scale space analysis on the ground feature geometric feature dataset, calculate the feature change rate at different scales, and generate a multi-scale ground feature feature dataset; Obtain a target feature type database, perform classification training based on the target feature type database, and build a feature classification model; The multi-scale feature dataset is classified into target feature categories using the feature classification model, and high-confidence classification results are screened to obtain a preliminary feature classification dataset. The classification error point categories of the preliminary object classification dataset are adjusted to generate the target classification dataset.

8. The method for automatic feature point annotation based on point cloud data according to claim 1, characterized in that: The allocation symbol style and binding point cloud feature attributes described in step S4 are specifically: Based on the target classification data set, calculate the global and local statistical characteristics of each category of objects and generate a statistical table of object characteristics; Based on the predefined symbol style library and the statistical table of land feature characteristics, the symbol style of the target classification dataset is matched and mapped to obtain a preliminary symbol mapping dataset; Adaptively adjust the preliminary symbol mapping dataset based on the point cloud feature attributes in the feature statistics table, set the symbol scale adjustment factor to [0.5, 2.0] and the color mapping parameter to [0, 255], adjust the symbol shape according to the height, density and classification confidence of the feature, and generate an optimized symbol mapping dataset; Bind the optimized symbol mapping dataset to the target classification dataset, set the point cloud and symbol matching constraint to [0.1, 1.0], and generate a symbolic point cloud dataset; Perform consistency checks on the symbolized point cloud dataset, detect and correct abnormal mapping points, and finally output the symbolized dataset.

9. A feature point automatic annotation system based on point cloud data, characterized in that: The method for automatically labeling feature points based on point cloud data according to claim 1 is configured to include: Point cloud fusion module, used to obtain regional multi-source point cloud data and perform multimodal data fusion on regional multimodal point cloud data to obtain point cloud fusion data; The ground point cloud segmentation module is used to segment the shoreline structure ground point cloud based on the point cloud fusion data to obtain a ground point cloud segmentation dataset; The salient point recognition module is used to perform edge detection on the ground point cloud segmentation dataset and identify geometric salient points based on the edge detection results to obtain a candidate set of feature points; The feature point symbolization module is used to classify the ground object type based on the feature point candidate set to obtain the target classification data set; it assigns symbol styles to the target classification data set and binds point cloud feature attributes in combination with the predefined symbol style library to obtain the symbolized data set; The dynamic interactive annotation module is used to dynamically adjust and interactively annotate multi-level symbols based on symbolic datasets to obtain a dynamic view of the symbols.

Images