Method for intelligently extracting key geometric features of target object by using three-dimensional laser scanner

Abstract

The invention provides a method for intelligently extracting key geometric features of a target object through a three-dimensional laser scanner, and belongs to the technical field of actual measurement of an internal forming surface of a wind tunnel body. Constructing an association degree matrix by using a minimum spanning tree algorithm in a graph theory, calculating three clustering center values based on K-means clustering analysis, determining a boundary threshold value, starting a scanning precision adaptive strategy, dynamically adjusting scanning parameters through a feature extraction optimization model, extracting key geometric features by using a geometric feature recognition algorithm, and carrying out feature extraction on the key geometric features; spatial context information is automatically extracted by using a Transform attention mechanism, and finally a key geometric feature extraction result of the target object is output through sequential processing of six parallel computing units, so that the technical problem of insufficient geometric feature extraction precision of three-dimensional laser scanning point cloud data in a complex environment is solved.

Description

Technical Field

[0001] This invention belongs to the field of actual measurement technology of the internal forming surface of wind tunnel, specifically, it relates to a method for intelligent extraction of key geometric features of target objects by a three-dimensional laser scanner. Background Technology

[0002] 3D laser scanning technology is an important measurement method that uses a laser beam to scan a target object from all directions to obtain high-precision 3D point cloud data. It is widely used in industrial inspection, architectural surveying, cultural relic protection, reverse engineering, and other fields. Traditional geometric feature extraction methods mainly rely on point cloud segmentation algorithms based on distance thresholds, feature point detection methods based on normal vectors, and surface reconstruction techniques based on region growing to identify and extract key geometric features of the target object. Traditional techniques have significant shortcomings when processing point cloud data in complex environments. Due to equipment system errors, environmental interference, and the surface characteristics of the target object during laser scanning, a large number of different types of interference data, such as noise points, anomalies, reflection points, and scattering points, are generated. Existing feature extraction algorithms lack effective identification and processing mechanisms for these different types of interference data, resulting in reduced accuracy of extracted geometric features. At the same time, traditional methods lack adaptive adjustment capabilities when facing different reflection environments and cannot dynamically optimize processing parameters according to actual scanning conditions. Traditional technologies struggle to effectively handle concurrent interference data of various types, especially under complex scanning conditions such as high-reflection, low-reflection, and mixed-reflection environments. Existing technologies cannot accurately identify and separate different types of interference data and lack targeted processing strategies, resulting in insufficient accuracy in geometric feature extraction. This affects the reliability and practicality of 3D laser scanning technology in precision measurement and industrial applications. Summary of the Invention

[0003] In view of this, the present invention provides a method for intelligent extraction of key geometric features of target objects from a 3D laser scanner, which can solve the technical problem of insufficient accuracy in geometric feature extraction of 3D laser scanning point cloud data in complex environments in the prior art.

[0004] This invention is implemented as follows: It provides a method for intelligent extraction of key geometric features of a target object using a 3D laser scanner, comprising: performing a 3D laser scanner to scan the target object from all directions to obtain raw point cloud data, and establishing noise point distribution matrices, anomaly point distribution matrices, reflection point distribution matrices, and scattering point distribution matrices; preprocessing the raw point cloud data using a multi-threaded parallel processing architecture, starting the maximum number of CUDA streams on the GPU, with six parallel computing units in each CUDA stream processing the anomaly point distribution matrix, reflection point distribution matrix, scattering point distribution matrix, and their fused feature matrix respectively; calculating the continuity parameter of each point using a spatial geometric analysis algorithm, and combining step feature recognition and fused feature extraction, utilizing... The minimum spanning tree algorithm in graph theory is used to calculate the correlation matrix of point cloud data. Based on this correlation matrix, K-means clustering analysis is performed to calculate the first, second, and third cluster center values. The boundary thresholds for each cluster are determined by the correlation of the minimum distance points. An adaptive scanning accuracy strategy is initiated, and the scanning parameters are dynamically adjusted based on the three cluster center values ​​using a feature extraction optimization model. A point cloud filtering function is then called to perform parallel filtering of the original point cloud data. A geometric feature recognition algorithm is used to extract key geometric features, and the Transformer attention mechanism automatically extracts spatial context information. Six parallel computing units establish a sequential relationship based on the number of data processing blocks, outputting the key geometric feature extraction results of the target object.

[0005] The noise point distribution matrix is ​​a data structure used to record the distribution pattern of invalid data points in three-dimensional space caused by equipment system errors, environmental interference and surface characteristics of the target object during the scanning process. The noise point distribution matrix is ​​used to calculate the input parameters of the point cloud filtering function and the gating weight function.

[0006] The anomaly distribution matrix is ​​a matrix used to store the location information of data points that deviate from normal geometric features. Anomalies are generated by abnormal reflection or occlusion of the target object surface by the laser beam. The anomaly distribution matrix is ​​used for processing in the first parallel computing unit, generation of fused feature matrix, input parameters of point cloud filtering function, and calculation of gating weight function.

[0007] The reflection point distribution matrix is ​​a matrix structure that records the spatial distribution of data points generated when the laser beam undergoes specular reflection on the surface of the target object. The reflection point distribution matrix is ​​used for processing by the second parallel computing unit, generation of fusion feature matrix, classification of environmental conditions in the scanning area, and calculation of gating weight function.

[0008] The scattering point distribution matrix is ​​a matrix that stores the distribution information of data points formed when the laser beam is scattered due to the surface roughness or material properties of the target object. The scattering point distribution matrix is ​​used for processing by the third parallel computing unit, generation of the fusion feature matrix, and calculation of the gate weight function.

[0009] Each point continuity parameter refers to a quantitative index of the continuity of the geometric relationship between adjacent points of the same type of data in three-dimensional space. It is obtained by calculating the distance, angle and the angle between the normal vectors of adjacent points. Each point continuity parameter is used for the calculation of step feature recognition and fusion feature extraction.

[0010] The correlation matrix is ​​a matrix structure based on graph theory used to describe the strength of spatial relationships between points in point cloud data. It is calculated using the minimum spanning tree algorithm and is used as an input parameter for K-means clustering analysis and point cloud filtering functions.

[0011] The first, second, and third cluster center values ​​are feature values ​​of three cluster centers obtained by analyzing the correlation matrix using the K-means clustering algorithm. The first cluster center value corresponds to a high-density point cloud region, the second cluster center value corresponds to a medium-density point cloud region, and the third cluster center value corresponds to a low-density point cloud region.

[0012] The minimum distance point correlation degree is a correlation measure determined by calculating the minimum Euclidean distance between the first cluster center value, the second cluster center value, and the third cluster center value. The minimum distance point correlation degree is used to determine the boundary threshold of each cluster and the input parameter of the point cloud filtering function.

[0013] The adaptive scanning accuracy strategy is a control strategy that dynamically adjusts the working parameters of the laser scanner based on the geometric features of the target object and the clustering analysis results. It includes adaptive adjustment of scanning resolution, sampling frequency and laser power. The adaptive scanning accuracy strategy is used to adjust the scanning parameters of the feature extraction optimization model.

[0014] The point cloud filtering function is used to remove noise points and outliers from the original point cloud data, thereby improving data quality and geometric feature extraction accuracy. The input includes the original point cloud data, noise point distribution matrix, outlier distribution matrix, correlation matrix, and minimum distance point correlation. The output is the optimized point cloud dataset.

[0015] The structure of the feature extraction optimization model is a convolutional neural network architecture based on an improved attention mechanism, which includes a multi-scale feature extraction layer, a spatial attention module, and a geometric feature fusion layer. The hierarchical fusion weights of the spatial attention module are determined based on the first cluster center value, the second cluster center value, and the third cluster center value.

[0016] The key geometric features include edge lines, surfaces, corners, step features, and fusion features, which are extracted from the optimized point cloud dataset through a geometric feature recognition algorithm. The key geometric features are used for the extraction of spatial context information and the output of the key geometric feature extraction results of the target object.

[0017] The spatial context information is automatically extracted from the optimized point cloud dataset using the Transformer attention mechanism, and is associated with key geometric features. This spatial context information is used to provide input data for the six parallel computing units.

[0018] The six parallel computing units respectively process the anomaly point distribution matrix, reflection point distribution matrix, scattering point distribution matrix, fusion feature matrix of anomaly point distribution matrix and reflection point distribution matrix, fusion feature matrix of reflection point distribution matrix and scattering point distribution matrix, and fusion feature matrix of anomaly point distribution matrix and scattering point distribution matrix. They construct the sequential relationship through the number of data processing blocks and output the key geometric feature extraction results of the target object.

[0019] The gate weight function is used to adjust the gate parameters of the spatial attention module in the feature extraction optimization model. The balance value is calculated based on four data: the distribution density of the noise point distribution matrix, the discreteness of the anomaly point distribution matrix, the intensity value of the reflection point distribution matrix, and the uniformity of the scattering point distribution matrix. The corresponding weight adjustment function is used according to the different balance ranges to which the balance value belongs.

[0020] This invention establishes a multi-dimensional data analysis system comprising noise point distribution matrices, anomaly point distribution matrices, reflection point distribution matrices, and scattering point distribution matrices. This enables precise identification and classification of different types of interference data during scanning, effectively addressing the technical limitations of traditional techniques in distinguishing and processing various types of interference data. Utilizing a multi-threaded parallel processing architecture and a collaborative working mechanism of six parallel computing units, this invention achieves efficient processing and accurate feature extraction of complex point cloud data through parallel processing of different types of distribution matrices and their fused feature matrices. Combined with K-means clustering analysis and the minimum spanning tree algorithm to calculate the correlation matrix, it also solves the problem of traditional techniques lacking adaptability under different reflection environments through an adaptive scanning accuracy strategy and dynamic parameter adjustment of the feature extraction optimization model. In summary, this invention solves the technical problem mentioned in the background art of insufficient geometric feature extraction accuracy of 3D laser scanning point cloud data in complex environments. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention.

[0022] Figure 2This is a comparative analysis chart of scanning accuracy under various environmental conditions in the embodiments. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0024] like Figure 1 The diagram shown is a flowchart of a method for intelligent extraction of key geometric features of a target object using a 3D laser scanner, provided by the present invention. This method includes the following steps:

[0025] S01. The target object is scanned from all directions using a 3D laser scanner to obtain raw point cloud data, and a noise point distribution matrix is ​​established to record the spatial distribution information of noise points generated during the scanning process. At the same time, anomaly point distribution matrix, reflection point distribution matrix and scattering point distribution matrix are constructed.

[0026] S02. Preprocess the raw point cloud data using a multi-threaded parallel processing architecture, start the maximum number of CUDA streams of the GPU, and start six parallel computing units in each CUDA stream. The six parallel computing units are respectively used for processing the abnormal point distribution matrix, the reflection point distribution matrix, the scattering point distribution matrix, the fusion feature matrix of the abnormal point distribution matrix and the reflection point distribution matrix, the fusion feature matrix of the reflection point distribution matrix and the scattering point distribution matrix, and the fusion feature matrix of the abnormal point distribution matrix and the scattering point distribution matrix.

[0027] S03. Calculate the continuity parameter of each point in the preprocessed raw point cloud data using spatial geometric analysis algorithms, and combine step feature recognition and fusion feature extraction to calculate the correlation matrix of the point cloud data using the minimum spanning tree algorithm in graph theory.

[0028] S04. Perform K-means clustering analysis based on the correlation matrix, calculate the first cluster center value, the second cluster center value, and the third cluster center value, and determine the boundary threshold of each cluster by the correlation degree of the minimum distance point;

[0029] S05. Start the scanning accuracy adaptive strategy. The feature extraction optimization model dynamically adjusts the scanning parameters based on the first cluster center value, the second cluster center value, and the third cluster center value. At the same time, the point cloud filtering function is called to perform parallel filtering processing on the original point cloud data.

[0030] S06. Use a geometric feature recognition algorithm to extract key geometric features of the target object from the optimized point cloud dataset processed by the point cloud filtering function, including edge lines, surfaces, corners, step features and fusion features. Use the Transformer attention mechanism to automatically extract spatial context information associated with key geometric features from the optimized point cloud dataset to provide input data for six parallel computing units.

[0031] S07. Start six parallel computing units to process the anomaly point distribution matrix, reflection point distribution matrix, scattering point distribution matrix, fusion feature matrix of anomaly point distribution matrix and reflection point distribution matrix, fusion feature matrix of reflection point distribution matrix and scattering point distribution matrix, and fusion feature matrix of anomaly point distribution matrix and scattering point distribution matrix respectively. Construct the sequential relationship through the number of data processing blocks, and output the final key geometric feature extraction results of the target object.

[0032] The noise point distribution matrix is ​​a data structure used to record the distribution pattern of invalid data points in three-dimensional space caused by equipment system errors, environmental interference and surface characteristics of the target object during the scanning process. The noise point distribution matrix is ​​used as the input parameter of the point cloud filtering function in step S05 and the calculation of the gating weight function in step S07.

[0033] The anomaly distribution matrix is ​​a matrix used to store the location information of data points that deviate from normal geometric features. The anomalies are generated by abnormal reflection or occlusion of the target object surface by the laser beam. The anomaly distribution matrix is ​​used for processing in the first parallel computing unit in step S02, generating the fusion feature matrix of the anomaly distribution matrix and the reflection point distribution matrix in step S04, generating the fusion feature matrix of the anomaly distribution matrix and the scattering point distribution matrix in step S04, the input parameters of the point cloud filtering function in step S05, and the calculation of the gating weight function in step S07.

[0034] The reflection point distribution matrix is ​​a matrix structure that records the spatial distribution of data points generated when the laser beam undergoes specular reflection on the surface of the target object. The reflection point distribution matrix is ​​used for processing in the second parallel computing unit in step S02, generating the fusion feature matrix of the anomaly point distribution matrix and the reflection point distribution matrix in step S04, generating the fusion feature matrix of the reflection point distribution matrix and the scattering point distribution matrix in step S04, classifying the environmental conditions of the scanning area, and calculating the gating weight function in step S07.

[0035] The scattering point distribution matrix is ​​a matrix that stores the distribution information of data points formed when the laser beam is scattered due to the surface roughness or material properties of the target object. The scattering point distribution matrix is ​​used for processing in the third parallel computing unit in step S02, generating the fusion feature matrix of the reflection point distribution matrix and the scattering point distribution matrix in step S04, generating the fusion feature matrix of the anomaly point distribution matrix and the scattering point distribution matrix in step S04, and calculating the gate weight function in step S07.

[0036] Each point continuity parameter refers to a quantitative index of the continuity of the geometric relationship between adjacent points of the same type of data in three-dimensional space. It is obtained by calculating the distance, angle and the angle between the normal vectors of adjacent points. Each point continuity parameter is used for the calculation of step feature recognition and fusion feature extraction in step S03.

[0037] The correlation matrix is ​​a matrix structure based on graph theory used to describe the strength of spatial relationships between points in point cloud data. It is calculated using the minimum spanning tree algorithm and is used as the input parameter for K-means clustering analysis in step S04 and point cloud filtering function in step S05.

[0038] The first, second, and third cluster center values ​​are feature values ​​of three cluster centers obtained by analyzing the correlation matrix using the K-means clustering algorithm. The first cluster center value corresponds to the center feature value of the high-density point cloud region and is used to identify the main geometric structure of the target object. The second cluster center value corresponds to the center feature value of the medium-density point cloud region and is used to identify the secondary geometric features of the target object. The third cluster center value corresponds to the center feature value of the low-density point cloud region and is used to identify the edge and detail features of the target object. The first, second, and third cluster center values ​​are used for calculating the minimum distance point correlation degree in step S04, adjusting the scanning parameters of the feature extraction optimization model in step S05, and determining the hierarchical fusion weight of the spatial attention module in the feature extraction optimization model.

[0039] The minimum distance point correlation degree is a correlation measure determined by calculating the minimum Euclidean distance between the first cluster center value, the second cluster center value, and the third cluster center value. The minimum distance point correlation degree is used to determine the boundary threshold of each cluster in step S04 and the input parameter of the point cloud filtering function in step S05.

[0040] Among them, the step feature refers to the feature point in the point cloud data where the geometric attributes of adjacent points change abruptly. It is identified by calculating the continuity parameter of each point. The step feature is used to extract the key geometric features in step S06.

[0041] Among them, the fusion feature refers to the composite feature formed by combining multiple basic geometric features in space, which is extracted by calculating the continuity parameter of each point. The fusion feature is used for the extraction of key geometric features in step S06.

[0042] The adaptive scanning accuracy strategy is a control strategy that dynamically adjusts the working parameters of the laser scanner based on the geometric features of the target object and the clustering analysis results. It includes adaptive adjustment of scanning resolution, sampling frequency and laser power. The adaptive scanning accuracy strategy is used to adjust the scanning parameters of the feature extraction optimization model in step S05.

[0043] The point cloud filtering function is used to remove noise points and outliers from the original point cloud data, thereby improving data quality and the accuracy of geometric feature extraction. The input includes the original point cloud data, the noise point distribution matrix, the outlier distribution matrix, the correlation matrix, and the minimum distance point correlation. The output is an optimized point cloud dataset, which is used for feature extraction in the geometric feature recognition algorithm in step S06.

[0044] The feature extraction optimization model is structured as a convolutional neural network architecture based on an improved attention mechanism, comprising a multi-scale feature extraction layer, a spatial attention module, and a geometric feature fusion layer. The hierarchical fusion weights of the spatial attention module are determined based on the first cluster center value, the second cluster center value, and the third cluster center value. The feature extraction optimization model is used in step S05 to dynamically adjust the scanning parameters based on the first cluster center value, the second cluster center value, and the third cluster center value.

[0045] The key geometric features include edge lines, surfaces, corners, step features, and fusion features, which are extracted from the optimized point cloud dataset using a geometric feature recognition algorithm. These key geometric features are used for the extraction of spatial context information in step S06 and the output of the key geometric feature extraction results of the target object in step S07.

[0046] The spatial context information is automatically extracted from the optimized point cloud dataset using the Transformer attention mechanism, and is associated with key geometric features. This spatial context information is used to provide input data for the six parallel computing units in step S06.

[0047] The number of data processing blocks is used to measure the computational load of the six parallel computing units when processing various types of matrix data. It is determined by counting the number of basic operations required for each parallel computing unit to complete data processing. The number of data processing blocks is used to construct the sequential relationship of the six parallel computing units in step S07.

[0048] The fusion feature matrix of the anomaly distribution matrix and the reflection point distribution matrix is ​​a comprehensive feature matrix obtained by performing feature-level fusion of the anomaly distribution matrix and the reflection point distribution matrix through matrix operations. The fusion feature matrix of the anomaly distribution matrix and the reflection point distribution matrix is ​​used for processing in the fourth parallel computing unit in step S02 and in the fourth parallel computing unit in step S07.

[0049] The fusion feature matrix of the reflection point distribution matrix and the scattering point distribution matrix is ​​a comprehensive feature matrix obtained by performing feature-level fusion of the reflection point distribution matrix and the scattering point distribution matrix through matrix operations. The fusion feature matrix of the reflection point distribution matrix and the scattering point distribution matrix is ​​used for processing in the fifth parallel computing unit in step S02 and in the fifth parallel computing unit in step S07.

[0050] The fusion feature matrix of the anomaly distribution matrix and the scattering point distribution matrix is ​​a comprehensive feature matrix obtained by performing feature-level fusion of the anomaly distribution matrix and the scattering point distribution matrix through matrix operations. The fusion feature matrix of the anomaly distribution matrix and the scattering point distribution matrix is ​​used for processing in the sixth parallel computing unit in step S02 and in the sixth parallel computing unit in step S07.

[0051] The six parallel computing units respectively process the anomaly distribution matrix, reflection point distribution matrix, scattering point distribution matrix, fusion feature matrix of anomaly distribution matrix and reflection point distribution matrix, fusion feature matrix of reflection point distribution matrix and scattering point distribution matrix, and fusion feature matrix of anomaly distribution matrix and scattering point distribution matrix. The order relationship is constructed by the number of data processing blocks. After the first parallel computing unit processes the anomaly distribution matrix, the second parallel computing unit starts processing the reflection point distribution matrix. After the second parallel computing unit processes the reflection point distribution matrix, the third parallel computing unit starts processing the scattering point distribution matrix. After the third parallel computing unit processes the scattering point distribution matrix, the fourth parallel computing unit starts processing the fusion feature matrix of anomaly distribution matrix and reflection point distribution matrix. After the fourth parallel computing unit processes the reflection point distribution matrix and scattering point distribution matrix, the fifth parallel computing unit starts processing the fusion feature matrix of reflection point distribution matrix and scattering point distribution matrix. After the fifth parallel computing unit processes the scattering point distribution matrix, the sixth parallel computing unit starts processing the fusion feature matrix of anomaly distribution matrix and scattering point distribution matrix. Finally, the key geometric feature extraction results of the target object are output.

[0052] The steps for establishing the training dataset of the feature extraction optimization model include collecting 3D laser scanning point cloud data of different types of target objects as raw samples, manually labeling each sample with key geometric features including edge lines, surfaces, corners, step features and fusion features, establishing distribution labels for noise point distribution matrices and outlier point distribution matrices, constructing a comprehensive training dataset containing various geometric shapes and surface characteristics, and dividing the data according to the ratio of training set, validation set and test set of 7:2:1.

[0053] The steps of training the feature extraction optimization model include training the model using a supervised learning method, calculating the difference between the predicted result and the true label using the cross-entropy loss function, updating the network parameters through the backpropagation algorithm, setting the learning rate to 0.001, the batch size to 32, the training epochs to 200, using an early stopping mechanism to prevent overfitting during training, and using a validation set to evaluate the model performance and save the optimal model parameters.

[0054] The gate weight function is used to adjust the gate parameters of the spatial attention module in the feature extraction optimization model. The balance value is calculated based on four data: the distribution density of the noise point distribution matrix, the dispersion of the outlier distribution matrix, the intensity value of the reflection point distribution matrix, and the uniformity of the scattering point distribution matrix. When the balance value is within the first balance range, a linearly increasing weight adjustment function is used; when the balance value is within the second balance range, a logarithmically increasing weight adjustment function is used; when the balance value is within the third balance range, a constant weight adjustment function is used; when the balance value is within the fourth balance range, an exponentially decaying weight adjustment function is used; and when the balance value is within the fifth balance range, a linearly decreasing weight adjustment function is used.

[0055] The environmental conditions of the scanning area are classified into three categories: high-reflectivity environment, low-reflectivity environment, and mixed-reflectivity environment. High-reflectivity environments are obtained by analyzing the reflection point distribution matrix using the K-means clustering algorithm and calculating the first environment cluster value. Low-reflectivity environments are obtained by analyzing the reflection point distribution matrix using the K-means clustering algorithm and calculating the second environment cluster value. Mixed-reflectivity environments are obtained by analyzing the reflection point distribution matrix using the K-means clustering algorithm and calculating the third environment cluster value. The impact of different environment classifications on the overall scheme is mainly reflected in scanning accuracy and feature extraction accuracy. High-reflectivity environments have higher data acquisition efficiency but are prone to overexposure points. Low-reflectivity environments have relatively lower data quality but fewer noise points. Mixed-reflectivity environments require more adaptive adjustment. The contribution rate of each environment classification is determined by calculating the ratio of the first environment cluster value to the total cluster value, the ratio of the second environment cluster value to the total cluster value, and the ratio of the third environment cluster value to the total cluster value. Different ratios have a significant impact on feature extraction accuracy. An excessively high proportion of high-reflectivity environments leads to an increase in overexposure points, affecting the accuracy of geometric feature recognition. An excessively high proportion of low-reflectivity environments reduces overall scanning efficiency. To reduce negative impacts, a dynamic threshold adjustment strategy is adopted to adjust the laser power and scanning parameters based on the first, second, and third environmental cluster values. Multispectral fusion technology is also introduced to improve the data acquisition quality under different reflection environments.

[0056] The specific implementation methods of the above steps are described in detail below.

[0057] The specific implementation of step S01 involves first emitting a laser beam towards the surface of the target object using a laser emitter. After reflection from the target object's surface, the laser beam is received by a laser receiver. Distance information is calculated by measuring the round-trip time of the laser beam, and the emission and reception angles of the laser beam are recorded simultaneously. The distance and angle information are combined to calculate the three-dimensional coordinate data of each point on the target object's surface, forming an original point cloud dataset. During the scanning process, a noise point distribution matrix is ​​simultaneously established. This matrix uses a three-dimensional grid structure to store the spatial distribution information of noise points, with a grid resolution of 0.1 mm. A distribution pattern model is established by statistically analyzing the number and density of noise points within each grid. Simultaneously, an anomaly point distribution matrix is ​​constructed. This matrix records the location information of data points deviating from normal geometric features due to abnormal reflections of the laser beam from the target object's surface. The anomaly point discrimination threshold is set to three times the distance standard deviation. A reflection point distribution matrix is ​​also constructed to record the spatial distribution of data points generated when the laser beam undergoes specular reflection on the target object's surface. The reflection intensity threshold is set to 1.5 times the average reflection intensity. A scattering point distribution matrix is ​​constructed to store the data point distribution information formed when the laser beam is scattered due to the surface roughness of the target object. Scattering point identification is based on the change angle of the surface normal vector, and the threshold is set to 15 degrees.

[0058] The specific implementation of step S02 involves activating the maximum number of CUDA streams in the graphics processor for multi-threaded parallel processing. Within each CUDA stream, six parallel computing units are activated to preprocess the raw point cloud data. The first parallel computing unit processes the outlier distribution matrix, using the isolated forest algorithm to identify outliers. Multiple decision trees within a random forest are constructed to score the anomaly of the data points, with an anomaly score threshold set to 0.6. The second parallel computing unit processes the reflection point distribution matrix, using an intensity threshold segmentation algorithm to mark points with reflection intensity exceeding a set threshold as reflection points. Simultaneously, the relationship between the reflection angle and the incident angle is calculated to verify the validity of the reflection points. The third parallel computing unit processes the scattering point distribution matrix, using a surface roughness estimation algorithm to calculate the local roughness parameters of the target object's surface and determining the intensity of the scattering phenomenon based on the roughness values. The fourth parallel computing unit processes the fusion feature matrix of the outlier distribution matrix and the reflection point distribution matrix, using a weighted fusion algorithm to combine the feature information of the two matrices, with a weight ratio of 0.7 to 0.3. The fifth parallel computing unit processes the fusion feature matrix of the reflection point distribution matrix and the scattering point distribution matrix, using principal component analysis to extract and fuse the main feature components of the two matrices. The sixth parallel computing unit performs the fusion feature matrix processing of the anomaly distribution matrix and the scattering point distribution matrix, and uses the mutual information algorithm to calculate the correlation between the two matrices and perform feature fusion.

[0059] The specific implementation of step S03 involves using a spatial geometric analysis algorithm to calculate the continuity parameters of each point in the preprocessed raw point cloud data. The continuity parameter calculation is based on three geometric relationship indicators: distance, angle, and the angle between normal vectors of adjacent points. Distance continuity is obtained by calculating the Euclidean distance between adjacent points, with a distance threshold set to twice the average point spacing. Angle continuity is obtained by calculating the angle change between adjacent points and a reference point, with an angle change threshold set to 30 degrees. Normal vector continuity is obtained by calculating the angle between the normal vectors of adjacent points, with a normal vector angle threshold set to 45 degrees. Combined with a step feature recognition algorithm, feature points with abrupt changes in geometric attributes are identified by detecting abrupt changes in continuity parameters. A first-order difference algorithm is used for abrupt change detection, with a difference threshold set to 2.5 times the standard deviation of the continuity parameters. The fusion feature extraction algorithm identifies composite features by analyzing the combination patterns of multiple basic geometric features in space, and uses a clustering fusion method to combine spatially adjacent and similar basic features to form fused features. The minimum spanning tree algorithm is used to calculate the correlation matrix of point cloud data. Based on graph theory, this algorithm treats each point in the point cloud data as a node of the graph, and the distance between points as the weight of the edge. The minimum spanning tree is constructed by Prim's algorithm. The weight of the edge in the tree reflects the strength of the spatial relationship between points, thus forming the correlation matrix.

[0060] The specific implementation of step S04 is to perform K-means clustering analysis based on the correlation matrix, setting the number of clusters to 3, and using Euclidean distance as the similarity metric. The first cluster center corresponds to a high-density point cloud region, determined by calculating data points in the correlation matrix whose weights are greater than 1.8 times the average weight, and is used to identify the main geometric structures of the target object, such as basic geometric shapes like planes and cylinders. The second cluster center corresponds to a medium-density point cloud region, determined by calculating data points in the correlation matrix whose weights are between 0.8 and 1.8 times the average weight, and is used to identify secondary geometric features of the target object, such as grooves and protrusions. The third cluster center corresponds to a low-density point cloud region, determined by calculating data points in the correlation matrix whose weights are less than 0.8 times the average weight, and is used to identify the edges and detailed features of the target object, such as sharp edges and small holes. The minimum distance point correlation is determined by calculating the minimum Euclidean distance between the three cluster centers; this value reflects the degree of correlation between regions of different densities. The boundary thresholds for each cluster are determined based on the minimum distance point correlation. The boundary threshold between high-density and medium-density regions is set to 1.2 times the minimum distance point correlation, and the boundary threshold between medium-density and low-density regions is set to 0.8 times the minimum distance point correlation.

[0061] The specific implementation of step S05 involves activating a scanning accuracy adaptive strategy. This strategy dynamically adjusts the operating parameters of the laser scanner based on the complexity of the target object's geometric features and the clustering analysis results. The feature extraction optimization model dynamically adjusts three key parameters—scanning resolution, sampling frequency, and laser power—based on the first, second, and third cluster center values. When the first cluster center value dominates, the scanning resolution is set to 0.5 mm, the sampling frequency to 10000 Hz, and the laser power to 80% of the rated power to ensure accurate capture of the main geometric structures. When the second cluster center value dominates, the scanning resolution is set to 0.3 mm, the sampling frequency to 15000 Hz, and the laser power to 90% of the rated power to balance scanning efficiency and feature recognition accuracy. When the third cluster center value dominates, the scanning resolution is set to 0.1 mm, the sampling frequency to 20000 Hz, and the laser power to 100% of the rated power to ensure complete acquisition of edge and detail features. The point cloud filtering function employs a multi-level filtering strategy. The input parameters include the original point cloud data, noise point distribution matrix, outlier distribution matrix, correlation matrix, and minimum distance point correlation. First, obvious noise points are removed based on the noise point distribution matrix, with the noise point discrimination threshold set to 0.3 times the local density. Then, geometric outliers are removed based on the outlier distribution matrix. Finally, fine filtering is performed based on the correlation matrix and minimum distance point correlation, outputting an optimized point cloud dataset.

[0062] The specific implementation of step S06 involves using a geometric feature recognition algorithm to extract key geometric features of the target object from the optimized point cloud dataset processed by the point cloud filtering function. Edge line feature extraction employs a gradient detection algorithm, identifying geometric boundaries by calculating the gradient changes of the point cloud data in three-dimensional space, with the gradient threshold set to 2.0 times the average gradient. Surface feature extraction uses a principal curvature calculation algorithm, identifying different types of surfaces by analyzing the curvature of local surfaces, with the principal curvature threshold set to 0.01 / mm. Corner feature extraction uses a three-dimensional extended version of the Harris corner detection algorithm, identifying spatial corners by calculating the eigenvalues ​​of the local structure tensor, with the corner response threshold set to 0.1. Step features are obtained through abrupt change detection of continuity parameters, and fusion features are obtained through spatial combination analysis of multiple basic geometric features. The Transformer attention mechanism automatically extracts spatial context information associated with key geometric features from the optimized point cloud dataset. This mechanism employs a multi-head self-attention structure with 8 attention heads, identifying spatial association patterns by calculating the attention weights between points in the point cloud data, with the attention weight threshold set to 0.15. Spatial context information, including the neighborhood distribution of feature points, local geometric change trends, and spatial topological relationships between features, provides input data for the six parallel computing units.

[0063] The specific implementation of step S07 involves activating six parallel computing units to process different types of distribution matrices and fusion feature matrices. The order of the six parallel computing units is constructed using the number of data processing blocks, determined based on the number of basic operations required for each parallel computing unit to complete data processing. The first parallel computing unit processes the anomaly distribution matrix, with 100 basic operation blocks. After processing, the second parallel computing unit is activated. The second parallel computing unit processes the reflection point distribution matrix, with 120 basic operation blocks. After processing, the third parallel computing unit is activated. The third parallel computing unit processes the scattering point distribution matrix, with 110 basic operation blocks. After processing, the fourth parallel computing unit is activated. The fourth parallel computing unit processes the fusion feature matrix of the anomaly distribution matrix and the reflection point distribution matrix, with 150 basic operation blocks. After processing, the fifth parallel computing unit is activated. The fifth parallel computing unit processes the fusion feature matrix of the reflection point distribution matrix and the scattering point distribution matrix, with 160 basic operation blocks. After processing, the sixth parallel computing unit is activated. The sixth parallel computing unit processes the fusion feature matrix of the anomaly distribution matrix and the scattering point distribution matrix. The number of data processing blocks is set to 140 basic computing units, and the final output is the key geometric feature extraction result of the target object.

[0064] The feature extraction optimization model employs a convolutional neural network architecture based on an improved attention mechanism. This model comprises three main components: a multi-scale feature extraction layer, a spatial attention module, and a geometric feature fusion layer. The multi-scale feature extraction layer uses convolutional kernels of different sizes to process the input data in parallel. The kernel sizes are set to 3×3×3, 5×5×5, and 7×7×7. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function, capturing geometric feature information at different scales through multi-scale convolutional operations. The spatial attention module is designed based on a self-attention mechanism and includes three linear transformation layers: query, key-value, and value. The attention weights are calculated using a dot product attention mechanism, and a softmax function is used to normalize the attention score. This module adaptively focuses on important spatial locations and feature channels. The geometric feature fusion layer uses feature concatenation and dimensionality reduction operations to fuse the outputs of the multi-scale feature extraction layer and the spatial attention module. Dimensionality reduction is achieved through a 1×1×1 convolutional kernel, and the final output layer uses a fully connected layer to map the features to a classification space of key geometric features. The hierarchical fusion weights of the spatial attention module are determined based on the first cluster center value, the second cluster center value, and the third cluster center value. The weight for high-density regions is set to 0.5, the weight for medium-density regions is set to 0.3, and the weight for low-density regions is set to 0.2.

[0065] The detailed steps for establishing the training dataset for the feature extraction optimization model include four stages: data collection, annotation processing, data augmentation, and data partitioning. In the data collection stage, 3D laser scanning point cloud data of different types of target objects in the industrial production environment are collected as raw samples. Target object types include mechanical parts, building components, vehicle parts, and electronic equipment, with 1000 samples collected for each type. The sample data format is a standard point cloud file, containing the 3D coordinates, reflection intensity, and timestamp information of the points. In the annotation processing stage, each sample is manually annotated, identifying the accurate location and type information of key geometric features, including edge lines, surfaces, corners, step features, and fusion features. Distribution labels for noise point distribution matrices and outlier point distribution matrices are also established, with an annotation accuracy requirement of 0.1 mm. In the data augmentation stage, methods such as rotation, translation, scaling, and adding noise are used to augment the raw samples. The rotation angle range is set to 0 degrees to 360 degrees, the translation distance range is set to -10 mm to +10 mm, the scaling ratio range is set to 0.8 to 1.2, and the noise intensity is set to 5% of the original signal intensity. Five augmented samples are generated for each raw sample. During the data partitioning phase, the constructed comprehensive training dataset is divided into training, validation, and test sets in a ratio of 7:2:1 to ensure that different types of target objects are evenly distributed across the datasets. The training set is used for model parameter learning, the validation set is used for model performance evaluation and hyperparameter tuning, and the test set is used for verifying the model's generalization ability.

[0066] The feature extraction optimization model training employs a supervised learning method, using the cross-entropy loss function to calculate the difference between the predicted results and the true labels. This loss function effectively handles multi-class classification problems and provides stable gradient information. The optimizer uses the Adam algorithm, which combines the advantages of momentum and adaptive learning rates, enabling rapid convergence and avoiding local optima. The learning rate is set to 0.001, and a cosine annealing strategy is used to dynamically adjust the learning rate during training, with a minimum learning rate of 0.0001. The batch size is set to 32 to balance computational efficiency and memory consumption. The number of training epochs is set to 200, and an early stopping mechanism is used to prevent overfitting; training stops when the validation set loss does not decrease for 10 consecutive epochs. The model performance is evaluated using a validation set, with metrics including accuracy, precision, recall, and F1 score. Optimal model parameters are saved when the model performance reaches a preset threshold. During training, regularization techniques such as learning rate decay, batch normalization, and Dropout are used to improve the model's generalization ability, with the Dropout ratio set to 0.2.

[0067] The key technical ideas of this invention are mainly reflected in four aspects: multi-dimensional parallel processing architecture, adaptive scanning accuracy control, fusion feature matrix processing, and attention mechanism optimization. The multi-dimensional parallel processing architecture, by simultaneously processing three different types of point cloud data—anomalies, reflection points, and scattering points—significantly improves data processing efficiency compared to traditional serial processing methods. Simultaneously, the collaborative working mechanism of six parallel computing units ensures the full utilization of different types of feature information, avoiding the limitations of single feature extraction methods. The adaptive scanning accuracy control technology dynamically adjusts scanning parameters according to the complexity of the target object's geometric features. Compared to traditional scanning methods with fixed parameters, it optimizes scanning efficiency while maintaining scanning accuracy. Especially for complex objects with multi-layered geometric features, this technology can automatically adjust the scanning resolution and sampling frequency according to the feature density of different regions. The fusion feature matrix processing technology fuses different types of point distribution matrices at the feature level. Compared to simple data overlay, it better preserves the unique information of various features. Furthermore, it mines the intrinsic correlations between different features through matrix operations, improving the accuracy and completeness of feature extraction. The attention mechanism optimization technique employs a Transformer architecture to automatically extract spatial context information. Compared to traditional rule-based feature extraction methods, it can better understand the complex spatial relationships in point cloud data. It captures geometric feature association patterns at different scales and in different directions through a multi-head attention mechanism. The synergistic effect of these four key technical approaches forms a complete 3D laser scanning feature extraction solution. Compared to existing technologies, it significantly improves processing efficiency, extraction accuracy, and adaptability. Especially when processing target objects with complex geometric structures and diverse material properties, this synergistic mechanism can fully leverage the advantages of each technology to achieve high-quality key geometric feature extraction results.

[0068] Further explanation is needed regarding existing technologies. When processing 3D laser scanning point cloud data, current technologies typically employ a single-threaded serial processing approach, processing different types of data points one by one in a fixed order. This fails to effectively distinguish and specifically address different types of interference data, such as anomalies, reflection points, and scattering points, resulting in low processing efficiency and insufficient feature extraction accuracy. Traditional methods lack in-depth analysis of the relationships between multiple data types, only performing simple data filtering and feature recognition, making it difficult to fully mine the fused feature information between different types of data points. This invention, through the collaborative working mechanism of six parallel computing units, achieves parallel processing of different types of distribution matrices and in-depth mining of fused feature matrices, fundamentally solving the technical bottleneck of existing technologies' inability to simultaneously process multiple data types. The first to third parallel computing units specifically process the anomaly point distribution matrix, reflection point distribution matrix, and scattering point distribution matrix, respectively, ensuring accurate identification and analysis of each data type. The fourth to sixth parallel computing units specifically process the fused feature matrix of different types of distribution matrices, deeply mining the intrinsic relationships and composite feature information between different types of data points through feature-level fusion of anomalies and reflection points, reflection points and scattering points, and anomalies and scattering points. The sequential relationship mechanism built upon the number of data processing blocks ensures the orderly collaborative work among the six parallel computing units, avoiding potential data conflicts and processing chaos during parallel processing. This sequential relationship not only guarantees the logic and accuracy of data processing but also achieves optimized allocation and efficient utilization of computing resources. Compared to existing technologies, this solution significantly improves data processing efficiency through its parallel processing architecture and greatly enhances the accuracy and completeness of geometric feature extraction through in-depth analysis of the fused feature matrix. This provides more reliable and efficient technical support for the application of 3D laser scanning technology in complex environments, promoting the development of point cloud data processing technology towards intelligence and precision.

[0069] It should be noted that this invention also solves the following technical problems: Traditional 3D laser scanning technology lacks environmental adaptability when handling different reflection environments. Existing technologies typically use fixed scanning parameters and processing algorithms, which cannot be dynamically adjusted according to the actual environmental conditions of the scanning area. This leads to overexposure points affecting data quality in high-reflectivity environments, low data acquisition efficiency in low-reflectivity environments, and unstable processing results in mixed-reflectivity environments. This invention establishes a reflection point distribution matrix and uses the K-means clustering algorithm to classify the environmental conditions of the scanning area, identifying three types: high-reflectivity, low-reflectivity, and mixed-reflectivity environments. It calculates the corresponding environmental clustering values ​​and contribution rates, employs a dynamic threshold adjustment strategy to adjust laser power and scanning parameters according to different environmental classifications, and introduces multispectral fusion technology to improve data acquisition quality under different reflection environments, thereby achieving adaptive processing of the scanning system for different reflection environments. It also addresses the technical problem of low computational efficiency in traditional point cloud data processing methods when dealing with large-scale complex point cloud data. Existing technologies mostly employ single-threaded serial processing, which suffers from slow computation speed and low resource utilization when processing massive point cloud data. This is especially true when multiple types of data matrices need to be processed simultaneously, as traditional methods cannot fully utilize the parallel processing capabilities of modern computing devices. This invention constructs a multi-threaded parallel processing architecture, utilizing the maximum number of CUDA streams in a GPU. Each CUDA stream initiates six parallel computing units to process different types of distribution matrices and their fused feature matrices. By establishing a sequential relationship through the number of data processing blocks, efficient parallel collaborative processing is achieved, significantly improving the computational efficiency and system resource utilization of point cloud data processing. This provides an efficient technical solution for large-scale 3D laser scanning applications.

[0070] Specifically, the principle of this invention is as follows: This invention can solve the technical problem of insufficient accuracy in geometric feature extraction of 3D laser scanning point cloud data in complex environments. Its fundamental principle lies in constructing a complete multi-dimensional data analysis and processing system. By establishing noise point distribution matrices, anomaly point distribution matrices, reflection point distribution matrices, and scattering point distribution matrices, it achieves accurate identification and classification of different types of interference data generated during the scanning process, providing a high-quality data foundation for subsequent feature extraction. The key to the logical consistency of the technical solution of this invention lies in the adoption of a multi-layered data processing architecture. First, it uses the multi-threaded parallel processing capability of the GPU to start six parallel computing units to process different types of distribution matrices and their fused feature matrices respectively. It calculates point continuity parameters through spatial geometric analysis algorithms, combines step feature recognition and fused feature extraction techniques, and uses the minimum spanning tree algorithm in graph theory to construct a correlation matrix, providing scientific data support for K-means clustering analysis. The core logic of the technical solution lies in the fact that the three cluster center values ​​obtained through K-means clustering analysis can accurately reflect the feature distribution of point cloud regions with different densities. The first cluster center value corresponds to high-density regions and is used to identify the main geometric structures; the second cluster center value corresponds to medium-density regions and is used to identify secondary geometric features; and the third cluster center value corresponds to low-density regions and is used to identify edge and detail features. The boundary threshold of each cluster is determined by calculating the correlation degree of the minimum distance points, providing a scientific basis for the adaptive scanning accuracy strategy. This invention utilizes an improved attention mechanism convolutional neural network architecture for feature extraction optimization models, combining the collaborative work of multi-scale feature extraction layers, spatial attention modules, and geometric feature fusion layers. It automatically extracts spatial context information associated with key geometric features using the Transformer attention mechanism and adjusts the spatial attention through a gating weight function.

[0071] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0072] The specific implementation of step S01 is to obtain the three-dimensional coordinate data of each point on the surface of the target object through the principle of laser ranging. The formula for calculating the three-dimensional coordinates of the i-th data point in the point cloud data is:

[0073]

[0074] In the formula, P i Let d be the three-dimensional coordinate vector of the i-th data point; i The distance value obtained by laser ranging; θ i φ is the elevation angle of the laser beam. i Let be the azimuth angle of the laser beam. The formula for establishing the noise point distribution matrix N is:

[0075]

[0076] In the formula, N jkl This represents the noise point density at position (j, k, l) in the 3D mesh, in units of points / mm. 3 M is the total number of points; δ(P) i G jkl Let ) be the Kronecker function, when point P i Located in grid G jkl The value is 1 if the time interval is short, otherwise it is 0; G jkl A three-dimensional mesh unit with a resolution of 0.1 mm; w i The noise weight for the i-th point is dimensionless and is calculated using the measurement error and signal-to-noise ratio. The calculation formula is as follows: Where ε i The measurement error at the i-th point, in mm, is obtained through laser calibration experiments. The measurement method involves repeatedly scanning a standard geometric object and calculating the standard deviation. SNR i For signal-to-noise ratio, dimensionless, through Calculate, where S i The signal strength, measured in mW, is obtained through a laser receiver. (N) noise，i The noise intensity, expressed in mW, is obtained through background noise measurement. The formula for establishing the anomaly distribution matrix A is:

[0077]

[0078] In the formula, A i This is the anomaly identifier for the i-th point, and it is dimensionless. The average coordinates of the local neighborhood are obtained through... The calculation yields the result, where m is the number of neighboring points, with a default value of 6; σ local The standard deviation of the local neighborhood is expressed in mm and is calculated using the following formula: The formula for calculating the environmental cluster value of the scanned area environmental condition classification is as follows:

[0079]

[0080] In the formula, E1, E2, and E3 are the cluster values ​​of the first, second, and third environments, respectively; R1, R2, and R3 are the sets of reflection points in the high-reflection, low-reflection, and mixed-reflection environments, respectively; |R k | represents the number of reflection points for the k-th environment category; I R，i Let be the intensity value of the i-th reflection point.

[0081] The specific implementation of step S02 is to adopt a multi-threaded parallel processing architecture, and the outlier score calculation adopts the Isolation Forest algorithm, where the outlier score S of the i-th data point is... anomaly，i The calculation formula is:

[0082]

[0083] In the formula, E(h(P) i Point P) i The average path length in an isolated forest, dimensionless; c(n) forest ) is the normalization factor. H is the harmonic number. n forest This represents the total number of samples in the isolated forest. The formula for calculating the fused feature matrix is:

[0084] F AR =αA + βR + γA⊙R;

[0085] F RS =αR + βS + γR⊙S;

[0086] F AS =αA + βS + γA⊙S;

[0087] In the formula, F AR F RS F AS These are the fusion feature matrices for anomaly points and reflection points, reflection points and scattering points, and anomaly points and scattering points, respectively; A, R, and S are the distribution matrices for anomaly points, reflection points, and scattering points, respectively; α, β, and γ are the fusion weight coefficients, dimensionless, with default values ​​of 0.4, 0.4, and 0.2, respectively; ⊙ is the Hadamard product operator, representing the multiplication of corresponding matrix elements.

[0088] The specific implementation of step S03 is to calculate the continuity parameter of each point, and the distance continuity parameter C. d，i The calculation formula is:

[0089]

[0090] In the formula, C d，i k is the distance continuity parameter for the i-th point, which is dimensionless; neighbor P represents the number of neighboring points, with a default value of 8. j Let σ be the j-th neighbor of the i-th point; d This is the distance standard deviation parameter, in mm, with a default value of 0.5 times the average point spacing. Angular continuity parameter C. a，i The calculation formula is:

[0091]

[0092] In the formula, C a，i Let v be the angular continuity parameter at the i-th point, which is dimensionless; ij From point P iTo neighboring point P j The vector, calculated by the formula v ij =P j -P i ;v ref The reference direction vector is used, and the local principal direction is determined through principal component analysis; ∠(·,·) is the function for calculating the angle between the vectors. The continuity parameter C of the normal vector... n，i The calculation formula is:

[0093]

[0094] In the formula, C n，i Let n be the continuity parameter of the normal vector at the i-th point, which is dimensionless; i and n j Point P i and neighboring point P j The unit normal vector is obtained through local plane fitting. The correlation matrix G is calculated based on the minimum spanning tree algorithm, and its elements G ij The calculation formula is:

[0095]

[0096] Where the edge weight w ij The calculation formula is:

[0097]

[0098] In the formula, G ij The element in the i-th row and j-th column of the correlation matrix, in mm; w ij σ represents the edge weight between points i and j, in mm; MST stands for Minimum Spanning Tree; σ c This represents the standard deviation of continuous differences, is dimensionless, and has a default value of 0.1.

[0099] The specific implementation of step S04 is based on K-means clustering analysis using the correlation matrix. The formula for calculating the three cluster center values ​​is as follows:

[0100]

[0101] In the formula, K1, K2, and K3 are the center values ​​of the first, second, and third clusters, respectively, in mm; C1, C2, and C3 are the point sets of the three clusters; |C k | represents the number of points in the k-th cluster; The mean of the i-th row of the correlation matrix is ​​calculated using the following formula: Where N total The total number of point clouds. Minimum distance point correlation degree D. min The calculation formula is:

[0102] D min =min{||K1-K2||, ||K2-K3||, ||K1-K3||};

[0103] In the formula, D min The minimum Euclidean distance between the three cluster centers is given. The formula for calculating the cluster boundary threshold is:

[0104] T1 = 1.2·D min ;

[0105] T2 = 0.8·D min ;

[0106] In the formula, T1 is the boundary threshold between high-density and medium-density regions; T2 is the boundary threshold between medium-density and low-density regions.

[0107] The specific implementation of step S05 is to activate the scanning accuracy adaptive strategy, and the calculation formula for dynamically adjusting the scanning parameters is as follows:

[0108] R scan =R base ·(1+λ1K1+λ2K2+λ3K3) -1 ;

[0109] F sample =F base ·(1+μ1K1+μ2K2+μ3K3);

[0110] P laser =P base ·(1+v1K1+v2K2+v3K3);

[0111] In the formula, R scan This is the dynamically adjusted scan resolution; F sample The sampling frequency is dynamically adjusted; P laser The laser power is dynamically adjusted; R base F base P base These represent the base scanning resolution, base sampling frequency, and base laser power, respectively; λ1, λ2, and λ3 are resolution adjustment coefficients, with default values ​​of 0.3, 0.5, and 0.8, respectively; μ1, μ2, and μ3 are sampling frequency adjustment coefficients, with default values ​​of 0.2, 0.4, and 0.6, respectively; v1, v2, and v3 are laser power adjustment coefficients, with default values ​​of 0.1, 0.2, and 0.3, respectively. The point cloud filtering function F... filter The calculation formula is:

[0112] P filtered =F filter (P raw N, A, G, D min);

[0113] Where the filter weight W i The calculation formula is:

[0114]

[0115] In the formula, P filtered The filtered optimized point cloud dataset; P raw Raw point cloud data; W i N represents the filter weight at the i-th point, which is dimensionless; density，i This represents the noise density value at the i-th point, in units of particles / mm. 3 Through N density，i =N jkl We obtain, where (j, k, l) is point P. i The grid location; G i The correlation value of the i-th point is expressed in mm. Obtain; σ n σ a , respectively, are the standard deviation parameters for noise and outliers, dimensionless, with default values ​​of 0.2 and 0.3, respectively. The formula for calculating the gate weight function Gate(B) is:

[0116] B=ω1ρ N +ω2D A +ω3I R +ω4U S ;

[0117]

[0118] In the formula, B is the equilibrium value, which is dimensionless; ρ N This represents the distribution density of the noise point distribution matrix, in units of points / mm. 3 ,pass Calculate, where V total Total scan volume, in mm. 3 ;D A The discreteness of the outlier distribution matrix is ​​dimensionless, expressed through... Calculation, where The mean of the outlier identifiers; I R The intensity values ​​of the reflection point distribution matrix are dimensionless and are obtained through... Calculate, where M R I represents the total number of reflection points. R，i U represents the intensity value at the i-th reflection point. S To ensure the uniformity of the scattering point distribution matrix, which is dimensionless, through... Calculate, where Var(S) is the variance of the scattering point distribution, calculated using the following formula: M SS represents the total number of scattering points. i Let i be the distribution value of the i-th scattering point. ω1, ω2, ω3, ω4 are the weighting coefficients, dimensionless, with default values ​​of 0.3, 0.2, 0.3, 0.2 respectively; a1, a2, a3, a4, b1, b2, b3, b4, c are the adjustment function parameters, dimensionless; B1, B2, B3, B4, B5, B6 are the boundary values ​​of the equilibrium range, dimensionless, with default values ​​of 0.2, 0.4, 0.6, 0.8, 1.0, 1.2 respectively.

[0119] The specific implementation of step S06 involves using a geometric feature recognition algorithm to extract key geometric features. The gradient calculation formula for edge line feature extraction is as follows:

[0120]

[0121] Where the gradient magnitude M i The calculation formula is:

[0122]

[0123] In the formula, M is the gradient vector at the i-th point; i This represents the gradient magnitude, used to determine edge line features. The formula for calculating the principal curvature of the surface feature extraction is:

[0124]

[0125] In the formula, κ1 and κ2 are the principal curvatures, with units of 1 / mm; f xx f yy f xy and are the second-order partial derivatives of the surface function z = f(x, y), in units of 1 / mm, obtained through local quadratic surface fitting; f x f y , respectively, are the first-order partial derivatives of the surface function, dimensionless, and calculated using numerical difference methods. Attention weight Att ij The calculation formula is:

[0126]

[0127] In the formula, Att ij Q represents the attention weight of the i-th query point to the j-th key-value point, which is dimensionless; i For the query vector, through linear transformation Q i =W Q ·F i Obtain, of which W Q To query the weight matrix, Fi K is the eigenvector of the i-th point; j For key-value vectors, through K j =W K ·F j Obtain, of which W K d is the key-value weight matrix; k N represents the dimension of the key-value vector, with a default value of 64; att This represents the total number of points involved in the attention calculation.

[0128] The specific implementation method of step S07 is the same as described above, and will not be repeated in detail here.

[0129] The principles and effects of each formula and mathematical model are explained below. The 3D coordinate calculation formula is based on the principle of spherical coordinate system transformation, directly obtaining the spatial position of points through laser ranging and angle measurement. Compared with traditional multi-view reconstruction methods, it has higher accuracy and real-time performance, avoiding the influence of viewpoint occlusion and feature matching errors. The formula for establishing the noise point distribution matrix adopts the principle of weighted statistics, quantifying the noise distribution pattern through gridded space and weight allocation, where the noise weights... Taking into account both measurement error and signal-to-noise ratio, this method more accurately reflects the spatial characteristics of noise compared to simple density statistics, providing a precise reference for subsequent filtering. The outlier identification formula is based on the 3σ criterion in statistics, identifying outliers by the mean and standard deviation of the local neighborhood. The average coordinates of the local neighborhood... and standard deviation The calculation ensures an accurate description of local geometric features and has stronger adaptability compared to the global thresholding method, effectively handling complex objects with large differences in features across different regions.

[0130] Environmental clustering value calculation formula Based on the statistical analysis principle of reflectance intensity, this method calculates the average reflectance intensity of points in different environmental categories to achieve quantitative classification of scanning environmental conditions. Compared with subjective judgment methods, this approach offers greater objectivity and repeatability, providing a quantitative basis for adaptive adjustment of scanning parameters. The isolated forest anomaly scoring formula is based on the path length principle, where the harmonic number... and normalization factor This approach ensures mathematical rigor in scoring. By constructing a decision tree through random partitioning to calculate the isolated difficulty of points, it exhibits better robustness and computational efficiency compared to traditional distance-based anomaly detection methods, making it particularly suitable for anomaly detection in high-dimensional point cloud data. The fusion feature matrix calculation formula combines linear combination and Hadamard product, preserving the independent information of various features while uncovering the interaction relationships between features. Compared to simple feature concatenation methods, it can more comprehensively express composite feature information.

[0131] The continuity parameter calculation formula is based on the neighborhood consistency principle, quantifying the degree of geometric continuity through Gaussian weights and vector operations. Compared to rule-based continuity judgment methods, it exhibits better numerical stability and noise resistance. The correlation matrix calculation formula combines spatial distance and feature similarity, ensuring global connectivity through the minimum spanning tree algorithm. Compared to simple adjacency matrix methods, it more accurately reflects the topological relationships of point clouds. The cluster center value calculation formula is based on the centroid principle of the K-means algorithm, where the row mean... The calculation ensures comprehensive utilization of correlation information, finding the optimal cluster center through iterative optimization. Compared to fixed-threshold segmentation methods, it has stronger adaptability and can automatically adjust the classification boundary according to the data distribution. The minimum distance point correlation calculation formula quantifies the degree of separation between clusters using Euclidean distance, providing a quantitative basis for setting the boundary threshold. Compared to empirical threshold methods, it has a stronger theoretical foundation and repeatability.

[0132] The dynamic adjustment formula for scanning parameters is based on the feedback control principle. It guides scanning parameter optimization through clustering results, and compared to fixed-parameter scanning methods, it can adaptively adjust according to the complexity of target features, significantly improving scanning efficiency and feature capture completeness. The gate weight function Gate(B) is based on the piecewise function principle, where the balance value B = ω1ρ N +ω2D A +ω3I R +ω4U S This method comprehensively considers four key factors: noise density, anomaly dispersion, reflection intensity, and scattering uniformity. By employing different adjustment functions within different ranges, it achieves refined weight control, exhibiting stronger adaptability and control precision compared to single-weight adjustment methods. The filter weight calculation formula comprehensively considers noise, anomaly, and correlation information, achieving precise filtering through multi-factor weighting. Compared to single-filter criterion methods, it demonstrates better data preservation and noise suppression capabilities.

[0133] The gradient calculation formula is based on differential geometry principles, using partial derivatives to calculate the surface change rate and identify edge features. Compared to template-matching-based edge detection methods, it exhibits higher accuracy and stronger scale invariance. The principal curvature calculation formula is based on surface theory in differential geometry, quantifying the surface curvature through second-order partial derivatives. Compared to fitting-based surface analysis methods, it demonstrates better local adaptability and computational stability. The attention weight calculation formula is based on the self-attention mechanism of the Transformer architecture, where the query vector Q... i =W Q ·F i and key value vector K j =W K ·F jThe calculation achieves feature space mapping through linear transformation and calculates spatial correlation degree through query key value matching. Compared with traditional rule-based feature association methods, it has stronger learning ability and generalization performance and can automatically discover complex spatial relationship patterns.

[0134] To better understand and implement this invention, a specific application scenario is provided below as Example 2: A technical team undertook the task of three-dimensional geometric accuracy inspection of the internal surface of a large wind tunnel. The internal surface of the wind tunnel has complex and diverse shapes, including multiple functional areas such as contraction sections, test sections, and diffusion sections. After the initial construction of the internal forming surface of the wind tunnel is completed, a comprehensive and detailed inspection of the construction accuracy of the internal surface is required. This is to more accurately grasp the accuracy deviation and to promptly process and improve it, ensuring that the final internal surface accuracy meets the design requirements. The internal surface of the wind tunnel in this project is large in size and complex in shape. If the conventional coordinate measurement method is used for inspection, the workload is large and the data is not clear enough. In order to better obtain the three-dimensional spatial geometric information of the overall internal surface, three-dimensional laser scanning technology is used for the accuracy inspection of the internal surface of the wind tunnel.

[0135] The technical team used the key geometric feature extraction method for target objects using a 3D laser scanner, as described in this invention, to perform precision testing on the surface profile inside a wind tunnel. First, the basic scanning parameters for the 3D laser scanner were established, and the basic scanning resolution R was set... base Set to 0.8mm, base sampling frequency F base Set to 12000Hz, base laser power P base The power was set to 75% of the rated power. The wind tunnel was divided into fifteen scanning sections, each with sufficient visibility. A 3D laser scanner was used to quickly acquire 3D point data of the spatial positions of the wind tunnel's surfaces by identifying and extracting key geometric features of the target objects, forming a scanned 3D solid point cloud. The data was then processed using Cyclone 9.1 software to obtain the complete spatial geometric dimensions of the tunnel. Based on the segmentation and coordinate positioning requirements of each section, an instrument deployment scheme was designed. To reduce coordinate errors and scanning errors caused by station relocation of the 3D laser scanner, the installation distance of the 3D laser scanner in each visible section was controlled within 10 meters.

[0136] During the scanning process, the technical team first performed a 360-degree scan of each scanning section to acquire raw point cloud data, obtaining an average of 3.5 million 3D data points per section. Simultaneously, a noise point distribution matrix was established with a grid resolution of 0.1 mm, and the noise point density distribution pattern was statistically analyzed. The anomaly point distribution matrix used three times the local standard deviation as a threshold to identify anomalies caused by abnormal reflections of the laser beam from the wind tunnel's inner wall; these anomalies accounted for approximately 2.8%. The reflection point distribution matrix recorded data points where the laser beam underwent specular reflection from the wind tunnel's inner wall; the reflection intensity threshold was set to 1.5 times the average reflection intensity, resulting in a reflection point proportion of approximately 15.3%. The scattering point distribution matrix stored data points caused by scattering phenomena due to the roughness of the wind tunnel's inner wall surface; scattering point identification was based on a threshold of 15 degrees for the surface normal vector change angle, resulting in a scattering point proportion of approximately 12.1%.

[0137] The technical team implemented a multi-threaded parallel processing architecture, employing eight CUDA streams on the GPU for data preprocessing, with six parallel computing units running within each CUDA stream. The first parallel computing unit processed the outlier distribution matrix, using the Isolation Forest algorithm to calculate anomaly scores with an anomaly score threshold set to 0.6, taking 45 seconds. The second parallel computing unit processed the reflection point distribution matrix, using an intensity threshold segmentation algorithm to label reflection points, taking 52 seconds. The third parallel computing unit processed the scattering point distribution matrix, using a surface roughness estimation algorithm, taking 48 seconds. The fourth, fifth, and sixth parallel computing units respectively processed the fusion feature matrix using a weighted fusion algorithm with weight ratios set to 0.4, 0.4, and 0.2, with fusion processing times of 63 seconds, 68 seconds, and 65 seconds, respectively.

[0138] Spatial geometric analysis algorithms calculate the continuity parameters for each point. For distance continuity parameters, eight neighboring points are used, with a distance threshold set to twice the average point spacing of 2.3 mm. For angle continuity parameters, the angle change threshold is set to 30 degrees, and for normal vector continuity parameters, the normal vector angle threshold is set to 45 degrees. Step feature recognition detects abrupt changes in continuity parameters using a first-order difference algorithm, with a difference threshold set to 2.5 times the standard deviation of the continuity parameters, identifying a total of 1847 step feature points. Fusion feature extraction uses clustering fusion methods to identify composite features, extracting 523 fusion features. The minimum spanning tree algorithm is used to calculate the correlation matrix of the point cloud data, and the Prim algorithm is used to construct the minimum spanning tree, with a computation time of 156 seconds.

[0139] K-means clustering analysis was performed based on the correlation matrix, with 3 clusters and Euclidean distance as the similarity metric. The first cluster centers correspond to high-density point cloud regions, with data points having a weight greater than 1.8 times the average weight, accounting for 32.4% of the total points. These primarily identify the planar and cylindrical surfaces of the wind tunnel's main structure. The second cluster centers correspond to medium-density point cloud regions, with weights between 0.8 and 1.8 times the average weight, accounting for 45.7% of the total points. These primarily identify transition sections and connections on the wind tunnel's inner wall. The third cluster centers correspond to low-density point cloud regions, with weights less than 0.8 times the average weight, accounting for 21.9% of the total points. These primarily identify edges, corners, and detailed features on the wind tunnel's inner wall. The minimum distance correlation coefficient was calculated to be 1.35 mm. The boundary threshold between high-density and medium-density regions was set to 1.62 mm, and the boundary threshold between medium-density and low-density regions was set to 1.08 mm.

[0140] As shown in Table 1, the scanning accuracy adaptive strategy dynamically adjusts the scanning parameters based on the three cluster center values.

[0141] Table 1 Results of dynamic adjustment of scanning parameters in different regions

[0142] Region Type Cluster center ratio Scan resolution sampling frequency laser power High-density areas 32.4％ 0.5mm 10000Hz 80％ Medium density area 45.7％ 0.3mm 15000Hz 90％ low-density areas 21.9％ 0.1mm 20000Hz 100％

[0143] The point cloud filtering function performs multi-level filtering on the original point cloud data. First, noise points are removed based on the noise point distribution matrix, with the noise point discrimination threshold set to 0.3 times the local density, removing 97,800 noise points. Then, geometric outliers are removed based on the outlier distribution matrix, removing 145,600 outliers. Finally, fine filtering is performed based on the correlation matrix and the correlation between the minimum distance points, resulting in an optimized point cloud dataset with a data retention rate of 94.2%.

[0144] Geometric feature recognition algorithms were used to extract key geometric features of the wind tunnel's internal surfaces from an optimized point cloud dataset. Edge line feature extraction employed a gradient detection algorithm with a gradient threshold set to 2.0 times the average gradient, identifying 2156 edge lines. Surface feature extraction used a principal curvature calculation algorithm with a principal curvature threshold set to 0.01 / mm, identifying planar regions (67.3%), convex surfaces (18.9%), and concave surfaces (13.8%). Corner feature extraction used a 3D Harris corner detection algorithm with a corner response threshold set to 0.1, identifying 834 corner features. Step features and fusion features were obtained using the aforementioned algorithms, resulting in a total of 2370 key geometric feature points extracted.

[0145] As shown in Table 2, the Transformer attention mechanism automatically extracts spatial context information from the optimized point cloud dataset.

[0146] Table 2. Statistics of Spatial Context Information Extraction Results

[0147] Feature type Number of feature points Attention weight range Average weight value Edge line features 2156 0.15-0.92 0.68 Surface features 1847 0.18-0.87 0.52 Corner features 834 0.22-0.95 0.74 Step characteristics 1847 0.16-0.89 0.61 Fusion features 523 0.31-0.96 0.78

[0148] The attention mechanism employs a multi-head self-attention structure, with 8 attention heads, a key-value vector dimension of 64, and an attention weight threshold of 0.15. Spatial context information, including the neighborhood distribution pattern of feature points, local geometric change trends, and spatial topological relationships between features, provides input data for the six parallel computing units.

[0149] The six parallel computing units are started sequentially according to the order in which they are constructed based on the number of data processing blocks. The first parallel computing unit processes the anomaly distribution matrix, with 100 basic computational units and a processing time of 68 seconds. The second parallel computing unit processes the reflection point distribution matrix, with 120 basic computational units and a processing time of 82 seconds. The third parallel computing unit processes the scattering point distribution matrix, with 110 basic computational units and a processing time of 75 seconds. The fourth parallel computing unit processes the fusion feature matrix of anomalies and reflection points, with 150 basic computational units and a processing time of 103 seconds. The fifth parallel computing unit processes the fusion feature matrix of reflection points and scattering points, with 160 basic computational units and a processing time of 109 seconds. The sixth parallel computing unit processes the fusion feature matrix of anomalies and scattering points, with 140 basic computational units and a processing time of 95 seconds.

[0150] For surface smoothness analysis, since the focus is on the relative values ​​of deviations at various points, the error of the 3D laser scanner system has a very small impact; the main control is over instrument operation error. The test results show that the surface smoothness deviation of the wind tunnel's inner wall is within...

[0151] Within the ±0.8mm range, the accuracy meets the design requirement of ±1.0mm. For shape deviation analysis, the systematic and random errors of the 3D laser scanner are the main errors. Before scanning, the instrument must undergo systematic error detection to correct the scanning results. Shape deviation detection results show that the roundness deviation of the contraction section is 0.65mm, the straightness deviation of the test section is 0.43mm, and the curve deviation of the diffusion section is 0.72mm, all within the design tolerance range. For coaxiality detection, the target ball recognition error, the systematic error of the total station distance measurement, and operational gross errors are the main errors affecting the detection results. The coaxiality deviation measurement result is 0.38mm, meeting the design requirement of 0.5mm.

[0152] The gating weighting function calculates a balance value based on four factors: noise point distribution density, anomaly distribution dispersion, reflection point distribution intensity, and scattering point distribution uniformity. The noise point distribution density is 0.024 points / mm. 3The dispersion of the outlier distribution is 0.167, the intensity of the reflection point distribution is 0.683, and the uniformity of the scattering point distribution is 0.431, resulting in a calculated equilibrium value of 0.526. Since the equilibrium value falls within the second equilibrium range, a logarithmic growth weighting adjustment function is used, resulting in a final gating weighting adjustment coefficient of 0.734.

[0153] Table 3 shows the statistical results of the geometric feature extraction accuracy of different regions of the wind tunnel surface.

[0154] Table 3. Statistics on the accuracy of geometric feature extraction of wind tunnel interior surfaces.

[0155] Wind tunnel section Edge line precision Surface recognition accuracy Corner detection accuracy Overall accuracy Inlet contraction section 0.15mm 0.12mm 0.18mm 0.14mm front section of the test section 0.11mm 0.08mm 0.13mm 0.10mm Middle section of the test section 0.09mm 0.07mm 0.11mm 0.08mm Rear section of the test section 0.12mm 0.09mm 0.14mm 0.11mm Front of diffuser section 0.16mm 0.13mm 0.19mm 0.15mm Rear part of the diffuser section 0.18mm 0.15mm 0.21mm 0.17mm

[0156] The entire process of extracting key geometric features of the wind tunnel's internal surface took 432 minutes, with data acquisition accounting for 65% of the time, data preprocessing for 18%, feature extraction for 12%, and result output for 5%. The final output of the key geometric feature extraction results for the target object contains complete three-dimensional geometric information of the wind tunnel's internal surface, with a feature extraction accuracy of 0.08mm-0.17mm, meeting engineering inspection requirements. Figure 2 A comparative bar chart is provided to illustrate the analysis of various accuracy indicators for different sections of the wind tunnel.

[0157] Environmental condition classification analysis results show that the scanning environment inside the wind tunnel is a mixed reflection environment. The cluster value for high-reflectivity environments is 0.342, for low-reflectivity environments it is 0.285, and for mixed reflection environments it is 0.373. High-reflectivity environments are mainly distributed on the metal surface areas of the wind tunnel's inner wall, low-reflectivity environments are mainly distributed on the coating surface areas, and mixed reflection environments are mainly distributed in the interface areas between different materials. For different environmental classifications, a dynamic threshold adjustment strategy was adopted to adjust the laser power and scanning parameters, and multispectral fusion technology was introduced to improve data acquisition quality, effectively reducing the negative impacts of overexposure points and missing data.

[0158] Compared with traditional coordinate measurement methods, the key geometric feature extraction method for target objects using a 3D laser scanner, as described in this invention, demonstrates significant technical advantages in surface accuracy detection within wind tunnels. From the perspective of data acquisition efficiency, traditional methods require point-by-point measurement, necessitating a large number of measurement points to obtain complete geometric information for complex surfaces. In contrast, this invention, through laser scanning, can rapidly acquire millions of 3D point cloud data points, significantly improving data acquisition efficiency. From the perspective of geometric feature recognition accuracy, traditional methods rely on manual judgment and experience-based threshold setting, making them susceptible to subjective factors. This invention, through a multi-dimensional parallel processing architecture and adaptive scanning accuracy control technology, can automatically identify and extract key geometric features, avoiding interference from human error. From the perspective of adaptability to complex environments, traditional methods require frequent adjustments to measurement parameters when facing surfaces with different reflectivity. This invention, through the fusion of feature matrix processing and attention mechanism optimization, can adaptively handle mixed environments with various materials and geometries, significantly improving the robustness and consistency of measurements.

[0159] It should be noted that the variables involved in this invention are explained in detail in Tables 4 and 5 below.

[0160] Table 4. Variable Explanation Table (Part 1)

[0161]

[0162]

[0163] Table 5. Variable Explanation Table (Part Two)

[0164]

[0165]

[0166] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for extracting key geometric features of a target object using a 3D laser scanner, characterized in that, include: The target object is scanned from all directions using a 3D laser scanner to obtain raw point cloud data, and noise point distribution matrices, anomaly point distribution matrices, reflection point distribution matrices, and scattering point distribution matrices are established. The raw point cloud data is preprocessed using a multi-threaded parallel processing architecture, and the maximum number of CUDA streams of the GPU are started. Each CUDA stream starts six parallel computing units to process the anomaly point distribution matrix, reflection point distribution matrix, scattering point distribution matrix, and their fused feature matrix, respectively. The continuity parameters of each point are calculated using spatial geometric analysis algorithms. Combined with step feature recognition and fusion feature extraction, the correlation matrix of the point cloud data is calculated using the minimum spanning tree algorithm in graph theory. Based on the correlation matrix, K-means clustering analysis is performed to calculate the first, second, and third cluster center values. The boundary thresholds of each cluster are determined by the correlation degree of the minimum distance points. An adaptive scanning accuracy strategy is initiated, dynamically adjusting the scanning parameters based on the three cluster center values ​​using a feature extraction optimization model. The original point cloud data is then filtered in parallel using a point cloud filtering function. Key geometric features are extracted using a geometric feature recognition algorithm, and spatial context information is automatically extracted using the Transformer attention mechanism. The six parallel computing units establish a sequential relationship through the number of data processing blocks and output the key geometric feature extraction results of the target object.

2. The method for extracting key geometric features of a target object using a three-dimensional laser scanner according to claim 1, characterized in that, The noise point distribution matrix is ​​a data structure used to record the distribution pattern of invalid data points in three-dimensional space caused by equipment system errors, environmental interference, and surface characteristics of the target object during the scanning process. The noise point distribution matrix is ​​used for the input parameters of the point cloud filtering function and the calculation of the gating weight function.

3. The method for extracting key geometric features of a target object using a three-dimensional laser scanner according to claim 2, characterized in that, The anomaly distribution matrix is ​​a matrix used to store the location information of data points that deviate from normal geometric features. Anomalies are generated by abnormal reflection or occlusion of the target object surface by the laser beam. The anomaly distribution matrix is ​​used for processing in the first parallel computing unit, generation of fused feature matrix, input parameters of point cloud filtering function, and calculation of gating weight function.

4. The method for extracting key geometric features of a target object using a three-dimensional laser scanner according to claim 3, characterized in that, The reflection point distribution matrix is ​​a matrix structure that records the spatial distribution of data points generated when the laser beam undergoes specular reflection on the surface of the target object. The reflection point distribution matrix is ​​used for processing by the second parallel computing unit, generation of fusion feature matrix, classification of environmental conditions in the scanning area, and calculation of gating weight function.

5. The method for extracting key geometric features of a target object using a three-dimensional laser scanner according to claim 4, characterized in that, The scattering point distribution matrix is ​​a matrix that stores the distribution information of data points formed when the laser beam is scattered due to the surface roughness or material properties of the target object. The scattering point distribution matrix is ​​used for processing by the third parallel computing unit, generation of fusion feature matrix and calculation of gate weight function.

6. The method for extracting key geometric features of a target object using a three-dimensional laser scanner according to claim 5, characterized in that, Each point continuity parameter refers to a quantitative index of the continuity of the geometric relationship between adjacent points of the same type of data in three-dimensional space. It is obtained by calculating the distance, angle and the angle between the normal vectors of adjacent points. Each point continuity parameter is used for the calculation of step feature recognition and fusion feature extraction.

7. The method for extracting key geometric features of a target object using a three-dimensional laser scanner according to claim 6, characterized in that, The correlation matrix is ​​a matrix structure based on graph theory used to describe the strength of spatial relationships between points in point cloud data. It is calculated using the minimum spanning tree algorithm and is used as an input parameter for K-means clustering analysis and point cloud filtering functions.

8. The method for extracting key geometric features of a target object using a three-dimensional laser scanner according to claim 7, characterized in that, The first, second, and third cluster center values ​​are feature values ​​of three cluster centers obtained by analyzing the correlation matrix using the K-means clustering algorithm. The first cluster center value corresponds to a high-density point cloud region, the second cluster center value corresponds to a medium-density point cloud region, and the third cluster center value corresponds to a low-density point cloud region.

9. The method for extracting key geometric features of a target object using a three-dimensional laser scanner according to claim 8, characterized in that, The minimum distance point correlation is a measure of the degree of correlation determined by calculating the minimum Euclidean distance between the first cluster center value, the second cluster center value, and the third cluster center value. The minimum distance point correlation is used to determine the boundary threshold of each cluster and the input parameter of the point cloud filtering function.

10. The method for extracting key geometric features of a target object using a three-dimensional laser scanner according to claim 9, characterized in that, The adaptive scanning accuracy strategy is a control strategy that dynamically adjusts the working parameters of the laser scanner based on the geometric features of the target object and the clustering analysis results. It includes adaptive adjustment of scanning resolution, sampling frequency and laser power. The adaptive scanning accuracy strategy is used to adjust the scanning parameters of the feature extraction optimization model.

Images