A ship segmentation feature point high-precision identification method based on regional sampling and weighted evaluation

Abstract

The application discloses a kind of based on regional sampling and weighted evaluation's ship segmentation feature point high-precision identification method.The method includes: according to the theoretical coordinate of target feature point in three-dimensional design model, extracts theoretical local model and determines physical space initial guide position;Control 3D camera scanning obtains initial local measured point cloud, and carries out coarse registration with theoretical local model to obtain initial transformation matrix;Theoretical feature point is mapped to measured space using transformation matrix to obtain estimated position, and the center is defined as refinement sampling area and extracts refinement point cloud subset;The comprehensive feature matching degree score obtained by the distance score, curvature score, normal consistency score and spatial proximity score is calculated for each point in refinement subset, and the highest score point is selected as the best feature point;Output final coordinate and calculate confidence, when confidence is too low, trigger alarm or re-measurement.The application reduces the amount of calculation by several orders of magnitude through the gradual refinement sampling mechanism guided by theory, meets the timeliness requirement of industrial online detection;Through the four-dimensional complementary weighted scoring system, the influence of noise and data loss is effectively overcome, and sub-millimeter level recognition accuracy is realized;Through the configurable weight table and confidence self-diagnosis mechanism, the engineering applicability and measurement reliability are enhanced.

Description

Technical Field

[0001] This application relates to the fields of computer vision, 3D point cloud processing and precision measurement technology, specifically to a method for efficiently and accurately identifying the coordinates of key feature points from local point cloud data in the 3D measurement of large and complex structural components such as ships and aerospace equipment, and particularly to a high-precision identification method for feature points of ship segments based on regional sampling and weighted evaluation. Background Technology

[0002] In shipbuilding, aerospace, and other fields, the dimensional and positional accuracy of large and complex structural components (such as ship sections and aircraft wing panels) directly determines the assembly quality and performance of the final product. With the rapid development of 3D vision measurement technology, non-cooperative target measurement methods based on structured light or laser scanning have become an important means of precision on-site inspection due to their non-contact, high-efficiency, and high-density data acquisition capabilities. One of the core challenges of such methods lies in how to stably and accurately extract key points representing the geometric features of the structure from the massive, discrete 3D point cloud data collected, such as rib intersections (corner points), points on plate seams (edge ​​points), and detection points on planes or curved surfaces.

[0003] In existing technologies, the main methods for extracting feature points from point clouds can be categorized as follows: 1. Global registration-based methods: For example, the Iterative Closest Point (ICP) algorithm performs global optimal registration between the measured point cloud and the theoretical CAD model, and then directly reads the position corresponding to the theoretical point from the registered model as the measurement result. This method has theoretically high accuracy, but its computational cost is extremely high, requiring the processing of hundreds of thousands or even millions of point cloud data for iterative optimization, which is very time-consuming and cannot meet the requirements of online, fast inspection cycles on production lines.

[0004] 2. Local geometric fitting-based methods: As mentioned in some patents, this method analyzes the local neighborhood of the point cloud to fit geometric primitives such as planes and cylinders, and then calculates feature points through methods such as intersection. This method has high requirements for the local quality and integrity of the point cloud. In actual industrial settings, due to factors such as wear on parts edges, surface corrosion, environmental reflection, or occlusion, the measured point cloud often contains noise, outliers, and missing data, leading to failure of local geometric fitting or significant deviations.

[0005] 3. Simple sampling method: Randomly select a point near the theoretical point as the feature point. This method is the fastest, but its accuracy is completely uncontrollable, it is extremely susceptible to noise and local anomalies, and its stability is very poor, making it unsuitable for precision measurement.

[0006] In summary, the core contradiction of existing technologies lies in the fact that high-precision recognition relies on sufficient, high-quality point cloud data, but efficiency requirements limit the amount of data used. Complex calculations on the entire point cloud are inefficient; simple random sampling cannot guarantee accuracy and stability. Therefore, there is an urgent need for a novel feature point recognition method that can intelligently focus on key areas, achieve the highest accuracy using the most concise amount of data, and is robust to point cloud quality. Summary of the Invention

[0007] The purpose of this invention is to overcome the contradiction between processing efficiency and recognition accuracy in existing technologies, and to provide a novel, efficient, and robust feature point recognition method. Specifically, this invention aims to achieve the following objectives: Highly controllable computational load and significantly improved efficiency: By introducing the "region pre-limitation" and "secondary refined sampling" mechanisms, the complex feature evaluation calculations are strictly limited to a very small subset of the most valuable key data, which fundamentally avoids processing massive amounts of irrelevant point cloud data, greatly reduces the computational load, and meets the timeliness requirements of industrial online detection.

[0008] Dual guarantee of recognition accuracy and robustness: Within the refined sampling area, multiple complementary geometric evaluation criteria, such as distance, curvature, and normal, are integrated to construct a configurable weighted scoring system. Through comprehensive evaluation of multiple criteria, the instability of a single criterion (such as relying solely on distance or curvature) in the presence of noise, wear, or missing points in the point cloud is effectively overcome, ensuring stable output of optimal results under various complex working conditions.

[0009] Provide a clear and implementable engineering process: Design a complete deterministic algorithm process from data input to result output, and provide specific mathematical formulas and configurable parameter tables to make it easy to program and implement in industrial measurement software to meet the requirements of reliable field applications.

[0010] This invention provides a high-precision identification method for ship segment feature points based on regional sampling and weighted evaluation, comprising the following steps: Step S1, Local Region Extraction and Spatial Pre-positioning of Theoretical Model: Based on the theoretical coordinates P-theory of the target feature point in the 3D design model, extract the theoretical local model M-local of a specific range around the point, and determine the initial guide position P-guide of P-theory in physical space based on the real-time coordinate system established by the measurement system; wherein, M-local is a theoretical surface patch, curve segment or a set of intersecting theoretical planes, and its range is defined by a preset 3D bounding box or radius R; Step S2, Local Acquisition and Coarse Registration of Measured Point Cloud: Control the 3D camera to scan the actual physical region centered on P-guide to obtain the initial local measured point cloud Cloud-raw; use a fast registration algorithm to perform preliminary coarse registration between Cloud-raw and M-local to obtain the initial transformation matrix T-init that roughly aligns the model space with the measured space; Step S3: Determination of the sampling frame and data sampling of key area point cloud: Use T-init to map the theoretical feature points P-theory to the measured space to obtain the estimated feature point positions P-estimated; define a refined sampling area Ω that is much smaller than the initial scanning range with P-estimated as the center; extract all points falling within the area Ω from Cloud-raw to form a refined point cloud subset Cloud-subset; Step S4: Selection of the best feature point based on multi-criteria weighted evaluation: Calculate the comprehensive feature matching score S(pi) for each point pi in the refined point cloud subset Cloud-subset, and select the point with the highest score as the best feature point P-best; wherein, the comprehensive feature matching score S(pi) is obtained by weighted summation of the distance scoring function F-dist(pi), the curvature scoring function F-curv(pi), the normal consistency scoring function F-normal(pi), and the spatial proximity scoring function F-prox(pi); Step S5, Coordinate Output and Confidence Assessment: Output the three-dimensional coordinates of P-best as the final measurement coordinates P-final of the target feature point, and calculate the confidence of this identification. If the confidence is too low, trigger an alarm or remeasure the process.

[0011] In some embodiments, the extraction method of the theoretical local model M-local in step S1 is as follows: taking the target feature point P-theory as the center, and according to the preset three-dimensional bounding box size or radius R, a local model containing the feature point and its surrounding geometry is extracted from the complete three-dimensional design model.

[0012] In some embodiments, the fast registration algorithm in step S2 includes the Iterative Closest Point (ICP) algorithm based on normal vectors or the feature point matching algorithm.

[0013] In some embodiments, the refined sampling region Ω in step S3 is a cubic space with a side length of L or a spherical space with a radius of r. L or r is preset according to the feature size and accuracy requirements, and the number of points M of the refined point cloud subset Cloud-subset is much smaller than the total number of points N of the initial local measured point cloud Cloud-raw.

[0014] In some embodiments, the distance scoring function F-dist(pi) in step S4 is calculated as follows: For a point pi, calculate the nearest distance from it to all triangular patches or analytical surfaces on the theoretical local model M-local, and take the minimum value as d-min(pi); The distance is converted into a score by applying a thresholded decay function: F-dist(pi) = max(0, 1 d-min(pi) / D-threshold) Where D-threshold is the preset distance tolerance threshold, and F-dist(pi) = 0 when d-min(pi)>D-threshold.

[0015] In some embodiments, the curvature scoring function F-curv(pi) in step S4 is calculated as follows: Centered on point pi, select its K nearest neighbor points in Cloud-subset, perform principal component analysis (PCA) on this set of points, calculate the three eigenvalues ​​of the covariance matrix λ1≥λ2≥λ3, and calculate the surface variation σ(pi) = λ3 / (λ1+λ2+λ3) of point pi. Applying curvature matching functions: F-curv(pi) = σ(pi) / (1 + |σ(pi) σ-target|) Where σ-target is the ideal curvature reference value of the target feature point.

[0016] In some embodiments, the normal consistency scoring function F-normal(pi) in step S4 is calculated as follows: Using the results of principal component analysis (PCA), the eigenvector corresponding to the smallest eigenvalue λ3 is taken as the local normal vector np of point pi; the nearest projection point qi of point pi is found on the theoretical model M-local, and the theoretical normal vector nm at qi is obtained. calculate: F-normal(pi) = |np · nm| Here, "·" represents the vector dot product operation.

[0017] In some embodiments, the spatial proximity scoring function F-prox(pi) in step S4 is calculated as follows: Calculate the Euclidean distance d-prox(pi) from point pi to the estimated feature point location P-estimated, and apply a Gaussian decay function to convert it into a score: F-prox(pi) = exp( d-prox(pi)² / (2δ²)) Where δ is a preset scale parameter.

[0018] In some embodiments, the formula for calculating the comprehensive feature matching score S(pi) in step S4 is: S(pi) = w1× F-dist(pi) + w2× F-curv(pi) + w3× F-normal(pi) + w4×F-prox(pi) Among them, w1, w2, w3, and w4 are weight coefficients, satisfying w1+w2+w3+w4=1, and the weight coefficients are preset and adjusted according to the geometric type of the feature points to be identified.

[0019] In some embodiments, the preset configuration of the weighting coefficients is as follows: For corner features: w1=0.5, w2=0.3, w3=0.1, w4=0.1; For edge point features: w1=0.4, w2=0.2, w3=0.2, w4=0.2; For planar feature points: w1=0.4, w2=0.1, w3=0.3, w4=0.2; For the surface feature points: w1=0.5, w2=0.1, w3=0.2, w4=0.2.

[0020] In some embodiments, the formula for calculating the confidence level in step S5 is: Confidence = S(P-best) / avg(S(pi)) Where S(P-best) is the comprehensive score of the best feature point, and avg(S(pi)) is the average comprehensive score of all points in the refined point cloud subset Cloud-subset; when Confidence is lower than the preset threshold, an alarm or remeasurement process is triggered.

[0021] In some embodiments, the method is applied to three-dimensional visual precision measurement of ship sections and large complex aerospace structural components, and is used to identify rib intersections, seam edge points, and detection points on planes or curved surfaces from local point cloud data.

[0022] Compared with the prior art, the present invention has the following advantages: High efficiency: Through theoretically guided coarse registration and refined sampling, complex calculations are limited to a very small subset of data, reducing the computational load by several orders of magnitude and meeting the requirements of online detection.

[0023] High precision and robustness: It integrates four complementary geometric criteria—distance, curvature, normal, and spatial proximity—with the distance criterion as the core, ensuring that the selected points are both on the correct geometric structure and near their theoretical location. The multi-criteria weighting effectively overcomes the effects of noise and missing local data, achieving sub-millimeter level recognition accuracy.

[0024] Highly applicable to engineering projects: Configurable weight tables can parameterize engineering experience, facilitating on-site adjustments; confidence assessment provides self-diagnostic capabilities, ensuring measurement reliability. Attached Figure Description

[0025] The accompanying drawings illustrate, by way of example and not limitation, the various embodiments discussed herein.

[0026] Figure 1 This is a schematic diagram of the high-precision identification method for ship segment feature points based on regional sampling and weighted evaluation according to an embodiment of the present invention; Figure 2 This is a schematic diagram of refined sampling. Detailed Implementation

[0027] In order to gain a more detailed understanding of the features and technical content of the embodiments of this application, the implementation of the embodiments of this application will be described in detail below with reference to the accompanying drawings. The accompanying drawings are for reference and illustration only and are not intended to limit the embodiments of this application.

[0028] In the embodiments described in this application, it should be noted that, unless otherwise stated and limited, the term "connection" should be interpreted broadly. For example, it can be an electrical connection, or a connection between two internal components. It can be a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can understand the specific meaning of the above term according to the specific circumstances.

[0029] It should be noted that the terms "first," "second," and "third" used in the embodiments of this application are merely used to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first," "second," and "third" can be interchanged in a specific order or sequence where permitted. It should be understood that the objects distinguished by "first," "second," and "third" can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in an order other than those illustrated or described herein.

[0030] This invention provides a high-precision identification method for ship segment feature points based on regional sampling and weighted evaluation, such as... Figure 1 As shown, it includes the following steps: Step S1: Local Region Extraction and Spatial Pre-positioning of the Theoretical Model: Based on the theoretical coordinates P-theory of the target feature point in the 3D design model, extract the theoretical local model M-local of a specific range around the point. Then, based on the real-time coordinate system established by the measurement system, determine the initial guide position P-guide of P-theory in physical space. M-local refers to a theoretical surface patch, curve segment, or a set of intersecting theoretical planes, and its range is defined by a preset 3D bounding box or radius R.

[0031] Step S2, Local Acquisition and Coarse Registration of Measured Point Cloud: Control the 3D camera to scan the actual physical region centered on P-guide to obtain the initial local measured point cloud Cloud-raw. Use a fast registration algorithm to perform preliminary coarse registration between Cloud-raw and M-local to obtain the initial transformation matrix T-init, which roughly aligns the model space with the measured space.

[0032] Step S3: Determination of Key Region Point Cloud Sampling Boundaries and Data Sampling: Using T-init, the theoretical feature points P-theory are mapped to the measured space to obtain the estimated feature point positions P-estimated. A refined sampling region Ω, much smaller than the initial scanning range, is defined centered on P-estimated. All points falling within region Ω are extracted from Cloud-raw to form a refined point cloud subset Cloud-subset.

[0033] Step S4: Selection of the best feature point based on multi-criteria weighted evaluation: Calculate the comprehensive feature matching score S(pi) for each point pi in the refined point cloud subset Cloud-subset, and select the point with the highest score as the best feature point P-best. The comprehensive feature matching score S(pi) is obtained by weighted summation of the distance scoring function F-dist(pi), the curvature scoring function F-curv(pi), the normal consistency scoring function F-normal(pi), and the spatial proximity scoring function F-prox(pi).

[0034] Step S5, Coordinate Output and Confidence Assessment: Output the three-dimensional coordinates of P-best as the final measurement coordinates P-final of the target feature point, and calculate the confidence of this identification. If the confidence is too low, trigger an alarm or remeasure the process.

[0035] In some embodiments, the extraction method of the theoretical local model M-local in step S1 is as follows: taking the target feature point P-theory as the center, and according to the preset three-dimensional bounding box size or radius R, a local model containing the feature point and its surrounding geometry is extracted from the complete three-dimensional design model.

[0036] In some embodiments, the fast registration algorithm in step S2 includes the iterative nearest point (ICP) algorithm based on normal vectors or the feature point matching algorithm.

[0037] In some embodiments, the refined sampling region Ω in step S3 is a cubic space with a side length of L or a spherical space with a radius of r. L or r is preset according to the feature size and accuracy requirements, and the number of points M of the refined point cloud subset Cloud-subset is much smaller than the total number of points N of the initial local measured point cloud Cloud-raw.

[0038] In some embodiments, the distance scoring function F-dist(pi) in step S4 is calculated as follows: For a point pi, calculate the nearest distance from it to all triangular patches or analytical surfaces on the theoretical local model M-local, and take the minimum value as d-min(pi).

[0039] The distance is converted into a score by applying a thresholded decay function: F-dist(pi) = max(0, 1 d-min(pi) / D-threshold) Where D-threshold is the preset distance tolerance threshold, and F-dist(pi) = 0 when d-min(pi)>D-threshold.

[0040] In some embodiments, the curvature scoring function F-curv(pi) in step S4 is calculated as follows: Centered on point pi, select its K nearest neighbors in the Cloud-subset, perform principal component analysis (PCA) on this set of points, calculate the three eigenvalues ​​of the covariance matrix λ1≥λ2≥λ3, and calculate the surface variation of point pi σ(pi) = λ3 / (λ1+λ2+λ3).

[0041] Applying curvature matching functions: F-curv(pi) = σ(pi) / (1 + |σ(pi) σ-target|) Where σ-target is the ideal curvature reference value of the target feature point.

[0042] In some embodiments, the normal consistency scoring function F-normal(pi) in step S4 is calculated as follows: Using the results of Principal Component Analysis (PCA), the eigenvector corresponding to the smallest eigenvalue λ3 is taken as the local normal vector np of point pi. The nearest projection point qi of point pi is found on the theoretical model M-local, and the theoretical normal vector nm at qi is obtained.

[0043] calculate: F-normal(pi) = |np · nm| Here, "·" represents the vector dot product operation.

[0044] In some embodiments, the spatial proximity scoring function F-prox(pi) in step S4 is calculated as follows: Calculate the Euclidean distance d-prox(pi) from point pi to the estimated feature point location P-estimated, and convert it into a score using a Gaussian decay function: F-prox(pi) = exp( d-prox(pi)² / (2δ²)) Where δ is a preset scale parameter.

[0045] In some embodiments, the formula for calculating the comprehensive feature matching score S(pi) in step S4 is: S(pi) = w1× F-dist(pi) + w2× F-curv(pi) + w3× F-normal(pi) + w4×F-prox(pi) Among them, w1, w2, w3, and w4 are weight coefficients, which satisfy w1+w2+w3+w4=1, and the weight coefficients are preset and adjusted according to the geometric type of the feature points to be identified.

[0046] In some embodiments, the preset configuration of the weighting coefficients is as follows: For corner features: w1=0.5, w2=0.3, w3=0.1, w4=0.1.

[0047] For edge point features: w1=0.4, w2=0.2, w3=0.2, w4=0.2.

[0048] For planar feature points: w1=0.4, w2=0.1, w3=0.3, w4=0.2.

[0049] For the surface feature points: w1=0.5, w2=0.1, w3=0.2, w4=0.2.

[0050] In some embodiments, the formula for calculating the confidence level in step S5 is: Confidence = S(P-best) / avg(S(pi)) Where S(P-best) is the comprehensive score of the best feature point, and avg(S(pi)) is the average comprehensive score of all points in the refined point cloud subset Cloud-subset. When Confidence falls below a preset threshold, an alarm or remeasurement process is triggered.

[0051] In some embodiments, the method is applied to three-dimensional visual precision measurement of ship sections and large complex aerospace structural components, and is used to identify rib intersections, seam edge points, and detection points on planes or curved surfaces from local point cloud data.

[0052] Compared with the prior art, the present invention has the following advantages: High efficiency: Through theoretically guided coarse registration and refined sampling, complex calculations are limited to a very small subset of data, reducing the computational load by several orders of magnitude and meeting the requirements of online detection.

[0053] High precision and robustness: It integrates four complementary geometric criteria—distance, curvature, normal, and spatial proximity—with the distance criterion as the core, ensuring that the selected points are both on the correct geometric structure and near their theoretical location. The multi-criteria weighting effectively overcomes the effects of noise and missing local data, achieving sub-millimeter level recognition accuracy.

[0054] Highly adaptable to various engineering projects: Configurable weight tables parameterize engineering experience, facilitating on-site adjustments. Confidence assessment provides self-diagnostic capabilities, ensuring measurement reliability.

[0055] The technical route of this invention is as follows: extraction of local regions of theoretical model and spatial pre-positioning → local acquisition and coarse registration of measured point cloud → determination of sampling frame of key region point cloud and data sampling → selection of best feature evaluation based on multi-criteria weighted evaluation → coordinate output and confidence evaluation.

[0056] I. Local Region Extraction and Spatial Pre-positioning of Theoretical Model Based on the theoretical coordinates P-theory of the target feature point in the 3D design model (CAD), the theoretical local model M-local of a specific range around the point is extracted. Simultaneously, based on the real-time coordinate system established by the measurement system, the initial guiding position P-guide of P-theory in physical space is determined.

[0057] • M-local can be a theoretical surface patch, a curve segment, or a set of intersecting theoretical planes, the extent of which is defined by a preset 3D bounding box or radius R.

[0058] II. Local Acquisition and Coarse Registration of Measured Point Clouds Control the 3D camera to scan the actual physical area centered on P-guide, and obtain the initial local measured point cloud Cloud-raw. Use a fast registration algorithm (such as normal vector-based ICP or feature point matching) to perform preliminary coarse registration between Cloud-raw and M-local, so as to obtain an initial transformation matrix T-init that roughly aligns the model space and the measured space.

[0059] III. Determination of the Point Cloud Sampling Frame for Key Areas and Data Sampling ①. Determine the sampling area: Use T-init to map the theoretical feature point P-theory to the measured space, so as to obtain the estimated feature point position P-estimated. The refined sampling is as shown in Figure 2 , with P-estimated as the center, define a refined sampling area Ω that is much smaller than the initial scanning range (for example, a cubic space with side length L or a spherical space with radius r, where L or r is preset according to the feature size and accuracy requirements).

[0060] ②. Point cloud sampling: Extract all points falling within the area Ω from Cloud-raw to form a refined point cloud subset Cloud-subset = {pi | i=1,...,M}, M<<N (N is the total number of points of Cloud-raw). This step eliminates a large amount of irrelevant data through spatial constraints, so that subsequent calculations focus on the most critical area.

[0061] IV. Optimal Feature Point Selection Based on Multi-Criteria Weighted Evaluation This step is the core calculation, but since it is based on the small dataset Cloud-subset, the calculation amount is small. For each point pi in the refined subset Cloud-subset, calculate its comprehensive feature matching score S(pi). The scoring system consists of four functions with clear geometric meanings, which characterize the matching degree between the point pi and the target feature point from different dimensions.

[0062] 1. Distance Scoring Function F-dist(pi) — Core Foundation Meaning: Evaluate the geometric proximity between the point pi and the theoretical local model M-local. This is the most basic and most important item among all criteria, which ensures that the selected point is located on the correct geometric structure (plane, curved surface or intersection line) rather than an isolated outlier or noise point.

[0063] Calculation steps: ①. For the point pi, calculate the shortest distance from it to all triangular patches (or analytical surfaces) on M-local, and take the minimum value as d-min(pi).

[0064] ②. Apply a threshold decay function to convert distance into a rating. The closer the distance, the higher the rating, but when the distance exceeds a certain range, the rating drops sharply.

[0065] Specific formula: F-dist(pi) = max(0, 1 – d-min(pi) / D-threshold) (Where, D-threshold is a preset distance tolerance threshold (e.g., 2 mm). When d-min(pi) > D-threshold, F-dist(pi) = 0. This design can effectively filter out outliers that are far from the theoretical model due to registration residual errors or measurement noise.) 2. Curvature scoring function F-curv(pi) Significance: This function evaluates whether the local geometry of point pi matches the expected feature type. Different feature points have different curvature characteristics: corner points should have high curvature, planar points should have low curvature, and edge points are in between. This function helps the algorithm identify points with the correct local shape.

[0066] Calculation steps: ①. Using point pi as the center, select its K nearest neighbors in Cloud-subset (K is a preset parameter, such as K=20).

[0067] ②. Perform principal component analysis (PCA) on this set of points and calculate the three eigenvalues ​​of the covariance matrix: λ1 ≥ λ2 ≥ λ3 (usually λ3 is the smallest, representing the change in the normal direction).

[0068] ③. Calculate the surface variation (or approximate curvature) of point pi: σ(pi) = λ3 / (λ1 + λ2 + λ3). A value close to 0 for σ(pi) indicates that it is located in a flat region, and a value close to 1 / 3 indicates that it is isotropic (such as near a corner point).

[0069] Specific formula: F-curv(pi) = σ(pi) / (1 + |σ(pi) – σ-target|) (Where, σ-target is the ideal curvature reference value for the target feature point (e.g., 0.3 for corner points and 0.01 for planar points). This formula ensures that the closer the actual curvature is to the target value, the higher the score, thus achieving targeted filtering for specific feature types.) 3. Normality scoring function F-normal(pi) Significance: This criterion evaluates the consistency between the local surface orientation of point pi and the orientation of the theoretical model near that point. For points on a plane, the normal should be perpendicular to the theoretical plane; for points on a surface, the normal should coincide with the normal of the theoretical surface at that point. This criterion strengthens the spatial orientation matching of points.

[0070] Calculation steps: ①. Using the same PCA results as above, the local normal vector np of point pi is the eigenvector corresponding to the minimum eigenvalue λ3.

[0071] ②. On the theoretical model M-local, find the nearest projection point qi of point pi, and obtain the theoretical normal vector nm at qi.

[0072] Specific formula: F-normal(pi) = |np·nm| (That is, the absolute value of the cosine of the angle between the two normal vectors. When they are perfectly parallel, the score is 1; when they are perpendicular, the score is 0.) 4. Spatial proximity scoring function F-prox(pi) — location aid Significance: This function assesses the spatial proximity of point pi to the estimated feature point location P-estimated. Since the coarse registration in step two has already located P-estimated near the feature point, this function further ensures that the selected point does not deviate too far, acting as an "anchor." It complements the distance scoring function F-dist: F-dist ensures that the point falls on the correct geometric structure, while F-prox ensures that the point is close to the theoretical location. Together, they ensure that the final result has both geometric and positional accuracy.

[0073] Calculation steps: Calculate the Euclidean distance d-prox(pi) from point pi to the estimated feature point location P-estimated, and apply a Gaussian decay function to convert it into a score.

[0074] Specific formula: F-prox(pi) = exp(-d-prox(pi)² / (2δ²)) (Where, δ is a preset scale parameter, usually taken as half the radius of the sampling area or set empirically (e.g., 5mm). The closer the distance, the higher the score; the farther the distance, the faster the score decays.) 5. Overall Score and Weighting The weighted sum of the four scoring functions above yields the comprehensive feature matching score for point pi: S(pi) = w1 * F-dist(pi) + w2 * F-curv(pi) + w3 * F-normal(pi) + w4 *F-prox(pi) Where w1, w2, w3, and w4 are weight coefficients, satisfying w1 + w2 + w3 + w4 = 1. The weights can be preset and adjusted according to the geometric type of the feature points to be identified. Since distance is the most basic and important criterion, w1 is usually given a larger weight (e.g., 0.4~0.5) to highlight its core position.

[0075] The table below provides recommended weight configurations for typical characteristics of ship sections:

[0076] (Note: The values ​​in the table are empirical recommendations. In actual applications, they can be fine-tuned according to measurement specifications, but w1 is generally kept at its maximum value.) 6. Determination of optimal feature points Iterate through all points in the Cloud-subset, calculate the comprehensive score S(pi) for each point, and select the point with the highest score as the best feature point: P-best = pk, where k = argmax( S(pi) ), for all pi belonging to Cloud-subset.

[0077] V. Coordinate Output and Confidence Assessment The 3D coordinates of P-best are output as the final measurement coordinates P-final of the target feature point. At the same time, the confidence level of this identification is calculated as Confidence = S(P=best) / avg(S(pi)). If the confidence level is too low (below the preset threshold), an alarm or re-measurement process can be triggered to ensure the reliability of the measurement system.

[0078] The embodiments of the present invention provide a theory-guided progressive refinement sampling mechanism: through a progressive focusing strategy of "theoretical point extraction → coarse registration → spatial mapping → secondary refinement sampling," computational resources are precisely locked onto extremely small critical areas. The amount of point cloud data is drastically reduced from hundreds of thousands to hundreds of points, fundamentally solving the computational load problem caused by global registration or full-scale analysis, and laying the foundation for the timeliness of industrial online inspection.

[0079] The embodiments of the present invention provide a four-dimensional complementary weighted scoring system: a comprehensive scoring function S(pi) is constructed, comprising four dimensions: distance, curvature, normal, and spatial proximity, to perform a three-dimensional evaluation of each point within the refined region. The multi-dimensional information mutually verifies and complements each other, avoiding misjudgments caused by a single criterion (such as relying solely on curvature) in the presence of noise or missing data, significantly improving the robustness and accuracy of the recognition.

[0080] The embodiments of the present invention provide a configurable engineering weight table: abstract engineering experience is transformed into specific algorithm parameters, providing preset weight combinations (w1, w2, w3, w4) for different feature types (corner points, edge points, planar points, curved surface points). Corner points: emphasize curvature, supplemented by distance, to ensure the physical sharpness of the corner is found. Planar points: emphasize normal and distance, ensuring local flatness and correct orientation. Edge points: balance curvature and normal, taking into account both position and orientation. This makes the algorithm highly adaptable to the field; process engineers do not need to modify the core code, but can simply adjust the weights to match different measurement specifications and workpiece characteristics, lowering the application threshold.

[0081] The above embodiments of the present invention provide confidence level self-diagnosis and quality control: While outputting the final coordinates P-final, the system calculates the confidence score of the recognition result: Confidence = S(P-best) / avg(S(pi)). This ratio reflects the prominence of the best point's score relative to the overall level. When the confidence score falls below a preset threshold, it indicates that the current recognition result may be unreliable due to environmental interference or operational errors, and the system automatically triggers an alarm or retest process. This design embeds self-quality monitoring capabilities into the automated measurement system, effectively preventing the outflow of erroneous data.

[0082] The technical solutions described in the embodiments of this application can be combined arbitrarily without conflict.

[0083] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A high-precision identification method for feature points of ship sections based on regional sampling and weighted evaluation, characterized in that, Includes the following steps: Step S1, Local Region Extraction and Spatial Pre-positioning of Theoretical Model: Based on the theoretical coordinates P-theory of the target feature point in the 3D design model, extract the theoretical local model M-local of a specific range around the point, and determine the initial guide position P-guide of P-theory in physical space based on the real-time coordinate system established by the measurement system; wherein, M-local is a theoretical surface patch, curve segment or a set of intersecting theoretical planes, and its range is defined by a preset 3D bounding box or radius R; Step S2, Local Acquisition and Coarse Registration of Measured Point Cloud: Control the 3D camera to scan the actual physical region centered on P-guide to obtain the initial local measured point cloud Cloud-raw; use a fast registration algorithm to perform preliminary coarse registration between Cloud-raw and M-local to obtain the initial transformation matrix T-init that roughly aligns the model space with the measured space; Step S3: Determination of the sampling frame and data sampling of key area point cloud: Use T-init to map the theoretical feature points P-theory to the measured space to obtain the estimated feature point positions P-estimated; define a refined sampling area Ω that is much smaller than the initial scanning range with P-estimated as the center; extract all points falling within the area Ω from Cloud-raw to form a refined point cloud subset Cloud-subset; Step S4: Selection of the best feature point based on multi-criteria weighted evaluation: Calculate the comprehensive feature matching score S(pi) for each point pi in the refined point cloud subset Cloud-subset, and select the point with the highest score as the best feature point P-best; wherein, the comprehensive feature matching score S(pi) is obtained by weighted summation of the distance scoring function F-dist(pi), the curvature scoring function F-curv(pi), the normal consistency scoring function F-normal(pi), and the spatial proximity scoring function F-prox(pi); Step S5, Coordinate Output and Confidence Assessment: Output the three-dimensional coordinates of P-best as the final measurement coordinates P-final of the target feature point, and calculate the confidence of this identification. If the confidence is too low, trigger an alarm or remeasure the process.

2. The high-precision identification method for ship segment feature points based on regional sampling and weighted evaluation according to claim 1, characterized in that, The extraction method of the theoretical local model M-local in step S1 is as follows: taking the target feature point P-theory as the center, according to the preset three-dimensional bounding box size or radius R, the local model containing the feature point and its surrounding geometry is extracted from the complete three-dimensional design model.

3. The high-precision identification method for ship segment feature points based on regional sampling and weighted evaluation according to claim 1, characterized in that, The fast registration algorithm described in step S2 includes the Iterative Closest Point (ICP) algorithm based on normal vectors or the feature point matching algorithm.

4. The high-precision identification method for ship segment feature points based on regional sampling and weighted evaluation according to claim 1, characterized in that, In step S3, the refined sampling region Ω is a cubic space with a side length of L or a spherical space with a radius of r. L or r is preset according to the feature size and accuracy requirements, and the number of points M of the refined point cloud subset Cloud-subset is much smaller than the total number of points N of the initial local measured point cloud Cloud-raw.

5. The high-precision identification method for ship segment feature points based on regional sampling and weighted evaluation according to claim 1, characterized in that, The distance scoring function F-dist(pi) mentioned in step S4 is calculated as follows: For a point pi, calculate the nearest distance from it to all triangular patches or analytical surfaces on the theoretical local model M-local, and take the minimum value as d-min(pi); The distance is converted into a score by applying a thresholded decay function: F-dist(pi) = max(0, 1 d-min(pi) / D-threshold) Where D-threshold is the preset distance tolerance threshold, and F-dist(pi) = 0 when d-min(pi) > D-threshold.

6. The high-precision identification method for ship segment feature points based on regional sampling and weighted evaluation according to claim 1, characterized in that, The curvature scoring function F-curv(pi) mentioned in step S4 is calculated as follows: Centered on point pi, select its K nearest neighbor points in Cloud-subset, perform principal component analysis (PCA) on this set of points, calculate the three eigenvalues ​​of the covariance matrix λ1≥λ2≥λ3, and calculate the surface variation σ(pi) = λ3 / (λ1+λ2+λ3) of point pi. Applying curvature matching functions: F-curv(pi) = σ(pi) / (1 + |σ(pi) σ-target|) Where σ-target is the ideal curvature reference value of the target feature point.

7. The high-precision identification method for ship segment feature points based on regional sampling and weighted evaluation according to claim 1, characterized in that, The calculation method for the normal consistency scoring function F-normal(pi) mentioned in step S4 is as follows: Using the results of principal component analysis (PCA), the eigenvector corresponding to the smallest eigenvalue λ3 is taken as the local normal vector np of point pi; the nearest projection point qi of point pi is found on the theoretical model M-local, and the theoretical normal vector nm at qi is obtained. calculate: F-normal(pi) = |np · nm| Here, "·" represents the vector dot product operation.

8. The high-precision identification method for ship segment feature points based on regional sampling and weighted evaluation according to claim 1, characterized in that, The spatial proximity scoring function F-prox(pi) mentioned in step S4 is calculated as follows: Calculate the Euclidean distance d-prox(pi) from point pi to the estimated feature point location P-estimated, and apply a Gaussian decay function to convert it into a score: F-prox(pi) = exp( d-prox(pi)² / (2δ²)) Where δ is a preset scale parameter.

9. The high-precision identification method for ship segment feature points based on regional sampling and weighted evaluation according to claim 1, characterized in that, The formula for calculating the comprehensive feature matching score S(pi) in step S4 is as follows: S(pi) = w1× F-dist(pi) + w2× F-curv(pi) + w3× F-normal(pi) + w4× F-prox(pi) Among them, w1, w2, w3, and w4 are weight coefficients, satisfying w1+w2+w3+w4=1, and the weight coefficients are preset and adjusted according to the geometric type of the feature points to be identified.

10. The high-precision identification method for ship segment feature points based on regional sampling and weighted evaluation according to claim 9, characterized in that, The preset configuration of the weighting coefficients is as follows: For corner features: w1=0.5, w2=0.3, w3=0.1, w4=0.1; For edge point features: w1=0.4, w2=0.2, w3=0.2, w4=0.2; For planar feature points: w1=0.4, w2=0.1, w3=0.3, w4=0.2; For the surface feature points: w1=0.5, w2=0.1, w3=0.2, w4=0.2.

Images