A three-dimensional point AI registration and tolerance analysis method and system for non-contact measurement

Abstract

The application provides a three-dimensional point AI registration and tolerance analysis method and system for non-contact measurement, relates to the field of three-dimensional point AI registration and tolerance analysis, and comprises the following steps: synchronously collecting point cloud data through a plurality of sensors, and performing adaptive preprocessing to obtain a preprocessed point cloud pair; using local curvature and normal vectors to construct a multi-scale geometric descriptor, and performing multi-scale hierarchical feature extraction to obtain an enhanced point feature set; calculating a bidirectional matching probability matrix by using optimal transport theory, screening high-confidence point pairs through spatial compatibility constraints, obtaining a high-confidence matching point pair set and a matching score, performing transformation matrix estimation and quantitative uncertainty, and outputting an optimal rigid transformation and a covariance matrix; and performing adaptive optimization in layers, dynamically adjusting a search range, and verifying transformation parameters across scales to output a converged accurate registration result. The application is used to solve the defect of low registration accuracy at a low overlap rate in the prior art.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of three-dimensional point AI registration and tolerance analysis, in particular to a three-dimensional point AI registration and tolerance analysis method and system for non-contact measurement. BACKGROUND

[0002] Three-dimensional point cloud registration and tolerance analysis is a core technology in the fields of computer vision, intelligent manufacturing and industrial detection. With the development of non-contact measurement technologies such as LiDAR, structured light scanning and photogrammetry, the acquisition of three-dimensional point cloud data has become more convenient and efficient, which has promoted the evolution from traditional algorithms to artificial intelligence-driven solutions to meet the higher requirements of modern industry for precision, efficiency and automation.

[0003] Three-dimensional point cloud registration refers to the process of aligning point cloud data acquired from different perspectives or times to a unified coordinate system through spatial transformation, which is a key step in three-dimensional reconstruction, quality detection and other applications. Tolerance analysis is a key link connecting product design and manufacturing, which is used to evaluate and control the impact of part size, shape and position deviation on product functional quality.

[0004] In the prior art, the existing technology faces major challenges in low overlap rate scenarios (overlap rate 15%-30%), and the precision and robustness of low overlap rate point cloud processing are low. SUMMARY

[0005] The present application provides a three-dimensional point AI registration and tolerance analysis method and system for non-contact measurement to solve the low registration accuracy defect in the prior art in low overlap rate scenarios.

[0006] In one aspect, the present application provides a three-dimensional point AI registration and tolerance analysis method for non-contact measurement, comprising:

[0007] Synchronously collecting point cloud data by multiple sensors and performing adaptive preprocessing to obtain a pair of preprocessed point clouds;

[0008] Based on the pair of preprocessed point clouds, constructing multi-scale geometric descriptors using local curvature and normal vectors, and performing multi-scale hierarchical feature extraction to obtain an enhanced point feature set;

[0009] Based on the enhanced point feature set, calculating a bidirectional matching probability matrix using optimal transport theory, and filtering high-confidence point pairs through spatial consistency constraints to obtain a high-confidence matching point pair set and a matching score;

[0010] Based on the high-confidence matching point pair set and the matching score, estimating a transformation matrix and quantifying uncertainty to output an optimal rigid transformation and its covariance matrix;

[0011] Based on the optimal rigid transformation and its covariance matrix, adaptive optimization is performed in layers, the search range is dynamically adjusted, and the transformation parameters are verified across scales, and the converged accurate registration result is output;

[0012] Based on the converged accurate registration result, a three-dimensional error field is constructed and the partition error characteristics are counted, the abnormal areas are interactively displayed through the heat map and streamline map, and the error analysis report and interactive visualization interface are output.

[0013] Further, through multi-sensor synchronous acquisition of point cloud data, adaptive preprocessing is performed to obtain a pair of preprocessed point clouds, including:

[0014] The multi-sensor synchronously acquired point cloud data is subjected to space-time synchronization and unified coordinate system to obtain aligned multi-modal data;

[0015] Based on the aligned multi-modal data, the sampling density is dynamically adjusted using voxelization grid, and the sampling granularity is determined according to the local curvature change of the point cloud, to obtain filtered data;

[0016] Based on the filtered data, combined with statistical outlier removal and normal vector consistency detection, abnormal point filtering is performed to output clean and density-balanced multi-modal point cloud pairs, i.e. preprocessed point cloud pairs.

[0017] Further, based on the preprocessed point cloud pairs, local curvature and normal vector are used to construct multi-scale geometric descriptors, and multi-scale hierarchical feature extraction is performed to obtain an enhanced point feature set, including:

[0018] Based on the preprocessed point cloud pairs, the multi-scale covariance matrix of each point cloud is calculated, and the multi-scale feature values and normal vectors are extracted through singular value decomposition to obtain local geometric coding;

[0019] Based on the local geometric coding, spatial attention is used to focus on key structures, and channel attention is used to strengthen discriminative feature dimensions to obtain double-attention-enhanced features;

[0020] Based on the double-attention-enhanced features, different scale features are fused from bottom to top through a feature propagation module to obtain an enhanced point feature set.

[0021] Further, based on the enhanced point feature set, the optimal transport theory is used to calculate a bidirectional matching probability matrix, and high-confidence point pairs are selected through spatial compatibility constraints to obtain a high-confidence matching point pair set and a matching score, including:

[0022] Based on the enhanced point feature set, the inner product similarity of the source point cloud and the target point cloud features is calculated and normalized to obtain a constructed bidirectional matching probability matrix:

[0023] Based on the two-way matching probability matrix, the similarity matrix is expanded and the empty set is added, and the optimal transmission plan matrix with empty set constraint is solved to obtain the optimal transmission matching plan matrix;

[0024] Based on the optimal transmission matching plan matrix, the geometric consistent point pairs are screened through the distance invariance under rigid transformation and the triangle proportion consistency to obtain a high-confidence matching point pair set and a matching score.

[0025] Further, based on the high-confidence matching point pair set and the matching score, the transformation matrix is estimated, and the uncertainty is quantified, and the optimal rigid transformation and the covariance matrix thereof are output, including:

[0026] Based on the high-confidence matching point pair set and the matching score, the centroid of the matching point pair is calculated, and the initial transformation is obtained through covariance matrix decomposition;

[0027] Based on the initial transformation, the matching error distribution is simulated using Monte Carlo simulation to generate multiple groups of perturbed matching point pairs;

[0028] Based on the multiple groups of perturbed matching point pairs, the distribution characteristics are counted and the covariance matrix is established to represent the uncertainty of the transformation, and the constructed uncertainty propagation matrix is output;

[0029] Based on the uncertainty propagation matrix, a candidate transformation is generated for each high-confidence matching point pair subset, and the optimal transformation is selected through a voting mechanism, and the optimal rigid transformation and the covariance matrix thereof are output.

[0030] Further, based on the optimal rigid transformation and the covariance matrix thereof, hierarchical adaptive optimization is performed, the search range is dynamically adjusted, and the transformation parameters are verified across scales, and the converged accurate registration result is output, including:

[0031] Based on the optimal rigid transformation and the covariance matrix thereof, hierarchical accurate registration is performed on the rigid transformation parameters from coarse to fine, and the optimized rigid transformation parameters are obtained;

[0032] Based on the optimized rigid transformation parameters, the search range is dynamically adjusted, the point-to-local plane distance is used to accelerate convergence, and the transformation consistency under different resolutions is checked, and abnormal estimates are removed, and the converged accurate registration result is output.

[0033] Further, based on the converged accurate registration result, a three-dimensional error field is constructed and the partition error characteristics are counted, and the abnormal areas are interactively displayed through the heat map and streamline map, and an error analysis report and an interactive visualization interface are output, including:

[0034] Based on the converged accurate registration result, the dense corresponding error vectors of the registered point cloud and the target point cloud are calculated, and a three-dimensional error vector field is established;

[0035] Based on the three-dimensional error vector field, the mean, standard deviation and extreme value are calculated in the partition, the systematic error and random noise are distinguished, and the statistical characteristic analysis result is obtained.

[0036] Based on the statistical characteristic analysis result, the error amplitude is displayed by the heat map, and the error direction mode is shown by the streamline chart.

[0037] On the other hand, the application also provides a three-dimensional point AI registration and tolerance analysis system for non-contact measurement, comprising:

[0038] The acquisition module is used for synchronously collecting point cloud data by multiple sensors and performing adaptive preprocessing to obtain a preprocessed point cloud pair.

[0039] The processing module is used for constructing a multi-scale geometric descriptor using local curvature and normal vector based on the preprocessed point cloud pair, and performing multi-scale hierarchical feature extraction to obtain an enhanced point feature set; calculating a bidirectional matching probability matrix using optimal transport theory based on the enhanced point feature set, screening high-confidence point pairs through spatial compatibility constraint to obtain a high-confidence matching point pair set and a matching score; performing transformation matrix estimation and quantifying uncertainty based on the high-confidence matching point pair set and the matching score, and outputting an optimal rigid transformation and a covariance matrix thereof; performing adaptive optimization in layers based on the optimal rigid transformation and the covariance matrix thereof, dynamically adjusting the search range and verifying the transformation parameters across scales, and outputting a converged accurate registration result; constructing a three-dimensional error field and statistically analyzing partition error characteristics based on the converged accurate registration result, interactively displaying abnormal areas through a heat map and a streamline chart, and outputting an error analysis report and an interactive visualization interface.

[0040] On the other hand, the application also provides an electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor implements the above-mentioned any one of the three-dimensional point AI registration and tolerance analysis method for non-contact measurement when executing the program.

[0041] On the other hand, the application also provides a non-transitory computer readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the above-mentioned any one of the three-dimensional point AI registration and tolerance analysis method for non-contact measurement.

[0042] On the other hand, the application also provides a computer program product comprising a computer program, wherein the computer program is executed by a processor to implement the above-mentioned any one of the three-dimensional point AI registration and tolerance analysis method for non-contact measurement.

[0043] The application provides a three-dimensional point AI registration and tolerance analysis method and system for non-contact measurement, solves the low overlap registration problem through a hierarchical processing strategy, realizes collaborative optimization of registration and tolerance analysis, designs for a non-contact measurement scene, forms a complete closed loop from data acquisition to final analysis, and significantly improves the accuracy and robustness of low overlap point cloud processing; multi-modal feature fusion coding combines geometry, color and local context information, and enhances the representation ability of low overlap areas through hierarchical feature extraction; dynamic trusted area perception uses optimal transport theory combined with spatial compatibility constraints to adaptively identify and enhance overlapping areas, significantly reducing the interference of non-overlapping points on registration; the registration uncertainty is transmitted to the tolerance analysis link to establish a probabilistic tolerance evaluation model, and the influence of measurement error on tolerance analysis is quantified through Monte Carlo simulation to realize end-to-end uncertainty management from registration to tolerance analysis. BRIEF DESCRIPTION OF DRAWINGS

[0044] In order to more clearly illustrate the technical solutions in the application or prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0045] Figure 1 is a flowchart of the three-dimensional point AI registration and tolerance analysis method for non-contact measurement provided by the embodiment of the application;

[0046] Figure 2 is a schematic diagram of the three-dimensional point AI registration and tolerance analysis system for non-contact measurement provided by the embodiment of the application;

[0047] Figure 3 is a structural schematic diagram of an electronic device provided by the embodiment of the application. DETAILED DESCRIPTION

[0048] In order to make the objects, technical solutions and advantages of the application clearer, the technical solutions in the application will be described clearly and completely below with reference to the drawings in the application. Obviously, the described embodiments are some embodiments of the application, not all embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor belong to the protection scope of the application.

[0049] Figure 1 is one of the flowcharts of the three-dimensional point AI registration and tolerance analysis method for non-contact measurement provided by the embodiment of the application.

[0050] As Figure 1As shown, the three-dimensional point AI registration and tolerance analysis method for non-contact measurement provided by the embodiment of the application mainly comprises the following steps:

[0051] 11. Synchronously collect point cloud data by multiple sensors, and perform adaptive preprocessing to obtain a pair of preprocessed point clouds;

[0052] 12. Based on the pair of preprocessed point clouds, construct a multi-scale geometric descriptor using local curvature and normal vector, and perform multi-scale hierarchical feature extraction to obtain an enhanced point feature set;

[0053] 13. Based on the enhanced point feature set, calculate a bidirectional matching probability matrix using optimal transport theory, filter high-confidence point pairs through spatial compatibility constraints, and obtain a high-confidence matching point pair set and a matching score;

[0054] 14. Based on the high-confidence matching point pair set and the matching score, estimate a transformation matrix and quantify uncertainty, and output an optimal rigid body transformation and its covariance matrix;

[0055] 15. Based on the optimal rigid body transformation and its covariance matrix, perform adaptive optimization in layers, dynamically adjust the search range and verify the transformation parameters across scales, and output a converged accurate registration result;

[0056] 16. Based on the converged accurate registration result, construct a three-dimensional error field and statistically analyze the error characteristics in each partition, interactively display abnormal areas through a heat map and a streamline map, and output an error analysis report and an interactive visualization interface.

[0057] In the embodiments of the present application, multi-sensor synchronous acquisition and adaptive preprocessing are used to solve the problem of incomplete data of a single sensor by synchronous acquisition of multi-sensor (such as LiDAR and RGB camera), to enhance data richness by fusing multi-modal information (geometry, color, and reflection intensity), and to reduce data volume while retaining key features by adaptive preprocessing (voxelization down-sampling and statistical denoising), thereby significantly reducing subsequent calculation complexity; multi-scale hierarchical feature extraction is used to effectively capture subtle geometric structures in low overlap areas by using multi-scale descriptors based on local curvature and normal vectors and combining a double attention mechanism (space + channel), to maintain feature discrimination when the overlap rate is less than 20%, and to solve the problem of failure of traditional manual features (such as FPFH) in sparse areas; optimal transport matching and spatial compatibility verification are used to improve matching accuracy by using an entropy regularization process to handle noise interference and to further eliminate mis-matches with geometric inconsistencies by using spatial compatibility constraints (first-order distance invariance + second-order triangle proportion), thereby improving the registration success rate in low overlap rate (overlap rate 15%-30%) scenes; transformation matrix estimation and uncertainty quantification are used to not only output the optimal rigid transformation but also quantify the covariance matrix by using weighted SVD solution combined with Monte Carlo simulation, thereby realizing probabilistic expression of registration error and providing a reliable error propagation model for tolerance analysis; hierarchical optimization and cross-scale verification are used to improve calculation efficiency by using a hierarchical optimization strategy (adaptive ICP) from coarse to fine, to avoid local optimum by dynamically adjusting the search range based on overlap rate estimation, and to eliminate abnormal transformation parameters by cross-scale verification, thereby shortening the registration time and improving the accuracy of an aviation structural part (1.2M point cloud); and a three-dimensional error field is used to quantify systematic errors (such as offset) and random noise, to locate manufacturing defects by partitioning statistics (mean / standard deviation), and to help engineers quickly identify out-of-tolerance areas by interactive visualization of heat maps and streamline maps, thereby improving the accuracy of key profile tolerance analysis in industrial gear box detection.

[0058] As shown in Figure 1 , 11, point cloud data is acquired by multi-sensor synchronous acquisition, and adaptive preprocessing is performed to obtain a pair of preprocessed point clouds, including:

[0059] 111, the point cloud data acquired by the multi-sensor synchronous acquisition is subjected to time-space synchronization and unified coordinate system to obtain aligned multi-modal data;

[0060] 112, based on the aligned multi-modal data, a voxelization grid is used to dynamically adjust the sampling density, and the sampling granularity is determined according to the local curvature change of the point cloud to obtain screened data;

[0061] 113. Based on the filtered data, outlier filtering is performed by combining statistical outlier removal and normal vector consistency detection, and clean and density-balanced multimodal point cloud pairs are output, i.e., preprocessed point cloud pairs.

[0062] In this embodiment of the invention, spatiotemporal synchronization and coordinate system unification eliminate system errors in multi-sensor data, ensuring geometric consistency in subsequent processing and laying the foundation for low overlap registration. The point cloud space is divided into voxel grids (e.g., initial voxel side length 5mm), with each voxel retaining a representative point. Local curvature calculation: the covariance matrix is ​​calculated for the neighborhood of each point p (radius r = 3 times the voxel side length), and eigenvalues ​​and curvature are obtained through singular value decomposition. Dynamic granularity adjustment: in high curvature regions (κ > threshold), the voxel side length is reduced (e.g., 2mm) to preserve details, while in flat regions (κ ≤ threshold), the voxel side length is increased (e.g., 8mm) for sparsification. The k nearest neighbors of each point p (e.g., k = ...) are calculated. 50) Average distance μ and standard deviation σ are used to remove points that satisfy d(p)>μ+ασ. Normal vector consistency detection: Adaptive downsampling reduces the amount of data by more than 70% (e.g., from 1.2M points to 350K points) while retaining 98.5% of the key geometric features, significantly reducing the computational complexity of subsequent feature extraction and registration; the angle θ between the normal vector of point p and the average direction of the neighborhood normal vectors is calculated, and outliers with θ>30° are removed, outputting clean and density-balanced point cloud pairs; outlier filtering (e.g., statistical outlier removal and normal vector consistency detection) can reduce the noise level by 65% ​​(from 2.3mm to 0.8mm), effectively avoiding interference from flying points and measurement errors on registration.

[0063] like Figure 1 As shown in Figure 12, based on the preprocessed point cloud pairs, a multi-scale geometric descriptor is constructed using local curvature and normal vectors, and multi-scale hierarchical feature extraction is performed to obtain an enhanced set of point features, including:

[0064] 121. Based on the preprocessed point cloud pairs, calculate the multi-scale covariance matrix of each point cloud, and extract multi-scale eigenvalues ​​and normal vectors through singular value decomposition to obtain local geometric coding;

[0065] 122. Based on local geometric coding, spatial attention is used to focus on key structures, and channel attention is used to enhance the discriminative feature dimension, resulting in features enhanced by dual attention.

[0066] 123. Based on the features enhanced by dual attention, the features at different scales are fused from bottom to top through the feature propagation module to obtain the enhanced point feature set.

[0067] In the embodiment of the present application, by calculating the multi-scale covariance matrix of each point (such as neighborhood radius r1 < r2 < r3) and performing singular value decomposition (SVD), the multi-scale eigenvalues (λ1, λ2, λ3) and normal vectors (n1, n2, n3) are extracted, and the local geometric descriptors (such as curvature, normal vector angle, etc.) are constructed, which can capture both microscopic curvature changes and macroscopic surface trends, and significantly improve the feature discrimination in low overlap rate scenarios (overlap rate 15%-30%);

[0068] The spatial attention focuses on the high curvature area and the structure boundary by dynamically allocating the weight of the local geometric mode through the 3D convolution kernel, and suppresses the interference of the flat area; the channel attention evaluates the importance of the feature dimension through the normalized mutual information (DNMI), and strengthens the discriminative features robust to rotation and occlusion; the double attention mechanism combines local details and global context, so that the features still maintain high discriminability in low overlap areas (such as 15% overlap rate), and the false match rate is reduced;

[0069] The hierarchical feature fusion fuses the multi-scale features (such as local descriptors from 0.1m to 1.0m radius) from bottom to top through the feature propagation (FP) module, and constructs a 128-dimensional feature vector with local fine structure and global consistency; the hierarchical fusion strategy effectively solves the limitations of single-scale features in complex scenes, and reduces the registration error.

[0070] As shown in Figure 1 , based on the enhanced point feature set, the optimal transport theory is used to calculate the bidirectional matching probability matrix, high-confidence point pairs are selected through spatial compatibility constraints, and a high-confidence matching point pair set and a matching score are obtained, including:

[0071] 131、Based on the enhanced point feature set, the inner product similarity of the features of the source point cloud and the target point cloud is calculated, and normalized to obtain the constructed bidirectional matching probability matrix:

[0072] 132、Based on the bidirectional matching probability matrix, the similarity matrix is expanded and the empty set is increased, and the optimal transport plan matrix with empty set constraint is solved to obtain the optimal transport matching plan matrix;

[0073] 133、Based on the optimal transport matching plan matrix, geometric consistent point pairs are selected through the distance invariance under rigid transformation and the triangle proportion consistency to obtain a high-confidence matching point pair set and a matching score.

[0074] In the embodiment of the present application, the enhanced point feature set {F s , F t} (each point is a 128-dimensional vector) is input; the inner product similarity calculation: the feature similarity of the i-th point in the source point cloud P s and the j-th point in the target point cloud P t ;

[0075] where

[0076] where, S ij is the feature similarity, λ is the feature covariance weight (default 0.2) to strengthen the consistency of local feature distribution, ∑ ij is the neighborhood feature covariance matrix to improve the sensitivity to local geometric structure; bidirectional normalization: row normalization: column normalization: final similarity: where, α is the row and column weight balancing factor (default 0.6) to suppress outliers; output bidirectional matching probability matrix satisfies

[0077] Optimal transport matching plan matrix solving, first empty set expansion, expand the similarity matrix Add new row and column fill value δ = -log(N s ·N t ), indicating the "no matching" state; allow points to be unmatched with a cost of δ, avoid error propagation caused by forced pairing; entropy regularized optimal transport:

[0078]

[0079] Constraints:

[0080] where, T is the optimal transport matrix, β is the Mahalanobis distance weight, which is strengthened by ( the learnable projection matrix) to strengthen the alignment of feature space; γ is the prior distribution weight, T0 is initialized by point density distribution, to improve the uniformity of matching space; solve Sinkhorn iteration (complexity O(N s N t )) and converge in 10 iterations; output the optimal transport plan matrix T, element T ij represents the confidence score s i of the matching pair (p j , q ij );

[0081] Geometric consistency constraint verification, first-order distance constraint (rigid invariance):

[0082] |d(p i , p j )-d(q i , q j )|<τ1·(1+η·κ ij )

[0083] where, local curvature variation rate, adaptive adjustment threshold, η is the curvature sensitivity coefficient, high curvature area relaxes the constraint, τ1 is the basic distance tolerance;

[0084] Second-order proportional constraint (triangle similarity) :

[0085]

[0086] where σ local = std ( ∠ (p i p j , p j p k ) is a local surface flatness measure, ζ is a noise suppression factor (default 0.05) that tightens the constraint in flat areas; τ2 is the proportional tolerance threshold;

[0087] Compatibility matrix construction:

[0088]

[0089] The maximum connected subgraph is filtered by spectral clustering (Spectral Clustering) to retain geometric consistency matching pairs, and the high-confidence matching point pair set C = {(p i , q j} and the score s ij are output. Non-overlapping points are automatically excluded (such as s ij < 0.2);

[0090] Empty set constraint and covariance regularization solve the forced matching problem of traditional optimal transport at low overlap points, and the mismatch rate is reduced; adaptive geometric constraint dynamically adjusts the threshold through curvature and flatness, and maintains 95% matching accuracy on complex structures; end-to-end probability framework models the whole process from feature matching to geometric verification, providing a quantitative basis for uncertainty analysis for tolerance analysis; through the fusion of multi-order geometric constraints and probabilistic transport model, high-robust registration of low-overlap point clouds is realized, and reliable corresponding relationship is laid for subsequent tolerance analysis.

[0091] As shown in Figure 1 , based on the high-confidence matching point pair set and the matching score, the transformation matrix is estimated, and the uncertainty is quantified, and the optimal rigid transformation and its covariance matrix are output, including:

[0092] 141、Based on the high-confidence matching point pair set and the matching score, the centroid of the matching point pair is calculated, and the initial transformation is obtained through covariance matrix decomposition;

[0093] 142、Based on the initial transformation, the matching error distribution is simulated using Monte Carlo simulation, and multiple groups of perturbed matching point pairs are generated;

[0094] 143、Based on multiple sets of perturbed matching point pairs, the distribution characteristics are counted and the covariance matrix is established to represent the uncertainty of the transformation, and the constructed uncertainty propagation matrix is output;

[0095] 144、Based on the uncertainty propagation matrix, a candidate transformation is generated for each high-confidence matching point pair subset, the optimal transformation is selected through a voting mechanism, and the optimal rigid body transformation and its covariance matrix are output.

[0096] In the embodiments of the present application, by using weighted SVD and probabilistic disturbance modeling, the robustness problem of transformation matrix estimation in low overlap rate point cloud registration is solved, and the covariance propagation of registration error is realized, providing a reliable quantitative basis for tolerance analysis; the weighted SVD is used to solve the initial transformation: the input is the set of trusted point pairs C={(p i , q j )} and the matching score {s ij}, the curvature adaptive weight correction is performed: a local curvature weight is introduced for each matching pair, κ is the point curvature, σ κ is the curvature bandwidth (default 0.1), and the matching score is corrected: where λ f is the feature similarity gain (default 0.3);

[0097] Covariance matrix construction:

[0098]

[0099] where γ is a regularization factor to suppress singular matrix problems.

[0100] SVD decomposition and reflection processing are:

[0101]

[0102] Monte Carlo simulation of matching error distribution and error disturbance model: first define the disturbance parameters: position noise feature noise where σ p =0.1×point cloud spacing, σ f =0.05×feature norm;

[0103] Generate perturbed point pairs: feature vector

[0104] Spatial correlation constraint: add a spatial correlation matrix The perturbation term is modified as (τ is the correlation radius, default 1.0m); output N sets of perturbed matching point pairs (typically N=500);

[0105] Transform distribution statistics and covariance modeling, Lie algebra parameterization: convert rotation matrix R n to Lie algebra (6D vector: 3D rotation + 3D translation);

[0106] Covariance matrix calculation:

[0107]

[0108] where ξ n is the Lie algebra parameter of the nth Monte Carlo simulation, is the parameter mean, λ m is the conservative factor, σ max is the historical maximum standard deviation;

[0109] The uncertainty propagation matrix is:

[0110] Introduce Mahalanobis distance weight: where σ max is the historical maximum variance, λ m = 0.2 is the conservative factor;

[0111] Assuming fusion and optimal transform selection, generate subsets: divide the point pairs into K subsets based on Delaunay triangulation Each subset contains at least 3 non-coplanar points; candidate transform generation: for each subset C k Perform weighted SVD to get candidate transform {R k , t k};

[0112] Voting mechanism: define the voting score:

[0113] where β is the time decay factor (default 0.01), t is the iteration number; select the transform with the highest voting score as the output; subset division avoids local optimum, and the registration success rate is improved in low overlap rate (overlap rate 15%-30%) scenarios; the time decay factor preferentially selects new hypotheses, suitable for dynamic scanning scenarios.

[0114] As Figure 1 shown, based on the optimal rigid body transform and its covariance matrix, hierarchical adaptive optimization is performed, the search range is dynamically adjusted, and the transform parameters are verified across scales, and the converged accurate registration result is output, including:

[0115] 151、Based on the optimal rigid body transform and its covariance matrix, hierarchical fine registration is performed on the rigid body transform parameters from coarse to fine, and the optimized rigid body transform parameters are obtained;

[0116] 152、Based on the optimized rigid transformation parameters, the search range is dynamically adjusted, the point-to-local plane distance is used to accelerate convergence, and the consistency of transformation under different resolutions is checked, abnormal estimates are removed, and the accurate registration result is output.

[0117] In the embodiments of the application, through the hierarchical optimization strategy and the dynamic search mechanism, the convergence speed and local optimization problem in low overlap rate point cloud registration are solved, and the uncertainty of the transformation parameters is quantified by using the covariance matrix to realize robust registration from coarse to fine; the hierarchical fine registration is input with the initial transformation (R0, t0) and the covariance matrix Σ R,t , those multi-scale down-sampling: construct pyramid level k∈{1,…,K}, each layer down-sampling rate is γ k (default γ=0.5), the coarsest layer (k=1) point cloud resolution is reduced to 1 / 2 of the original data K ;

[0118] Dynamic adjustment of down-sampling granularity:

[0119]

[0120] Where λ κ is the curvature gradient sensitivity coefficient (default 0.2), more details are retained in high curvature areas, κ avg is the average curvature of the current layer; hierarchical optimization:

[0121] Hierarchical optimization: each layer solves the transformation parameters (R k ,t k ) by weighted SVD, the weight combines the matching score s ij and the covariance confidence:

[0122] Where ξ k is the Lie algebra parameter of the current layer, is the inverse covariance matrix;

[0123] Adaptive ICP improvement: change the point-to-point error of traditional ICP to the distance from the point to the local plane of the target point:

[0124] Where η is the normal vector alignment weight, which enforces the normal vector consistency;

[0125] Dynamic search range adjustment: adjust the search radius according to the overlap rate estimate overlap_est and the local point density ρ:

[0126] r search =α·overlap est +β·ρ+ζ·σ noise

[0127] where, a is the overlap rate weight, β is the point density coefficient, σ noise is the current layer noise standard deviation, calculated by the covariance matrix eigenvalue, and p is the local point density.

[0128] Step adaptation: adjust the iteration step size based on uncertainty:

[0129]

[0130] where, step_max is the maximum allowed step size (default 0.1).

[0131] As shown in Figure 1 , 16, based on the converged accurate registration results, construct a three-dimensional error field and calculate the partition error characteristics, through the interactive display of heat map and streamline map, output error analysis report and interactive visualization interface, including:

[0132] 161、Based on the converged accurate registration results, calculate the dense correspondence error vector of the registered point cloud and the target point cloud, and establish a three-dimensional error vector field.

[0133] 162、Based on the three-dimensional error vector field, calculate the mean, standard deviation and extreme value by partition, distinguish systematic error and random noise, and get the statistical characteristic analysis result.

[0134] 163、Based on the statistical characteristic analysis result, use heat map to display error amplitude and streamline map to show error direction mode.

[0135] In the embodiment of the application, a residual error vector field is constructed, a three-dimensional error field is generated by calculating the dense corresponding error vector E(x, y, z) = Pt(x, y, z) - Ps'(x, y, z) of the registered point cloud Ps' and the target point cloud Pt, the error field quantifies the registration deviation of each spatial point, and can intuitively reflect the local registration accuracy. For example, in the detection of industrial parts, the error vector field can accurately identify the micron-level deviation (such as 0.15 mm) of the tooth profile area, which is significantly better than the traditional method (0.5 mm); the error field is divided into grid areas, the mean value (reflecting the system deviation), the standard deviation (characterizing the random noise) and the extreme value (identifying the maximum deviation) of each area are calculated respectively, the system error and the random noise are distinguished, the registration algorithm can be optimized or the sensor parameters can be adjusted, for example, in the detection of aviation structural parts, the zoned statistics found that the system error of a certain area was 0.23 mm, which was traced to the sensor calibration deviation, and after correction, the error was reduced by 60%; the interactive visualization display, the heat map maps the error amplitude distribution with a color gradient, the red color identifies the high error area (such as the out-of-tolerance 0.3 mm or more), the blue color represents the low error area, the user can interactively locate the defect position through zooming and rotating, such as the abnormally high temperature area of the gear box tooth root crack; the streamline diagram shows the direction mode of the error vector through arrows or curves, and reveals the systematic deformation trend (such as overall translation or rotation); the heat map and the streamline diagram are combined, so that the detection rate of abnormal areas is improved; the zoned statistics quickly distinguish the sensor error and the environmental noise, shorten the debugging time; the interactive report supports engineers to adjust the detection strategy in real time, such as dynamically adjusting the laser scanning path in the detection of aviation skin.

[0136] As shown in Figure 2 A three-dimensional point AI registration and tolerance analysis system 20 for non-contact measurement, characterized by comprising:

[0137] An acquisition module 21 is configured to synchronously acquire point cloud data by multiple sensors and perform adaptive preprocessing to obtain a preprocessed point cloud pair.

[0138] The processing module 22 is configured to construct a multi-scale geometric descriptor using local curvature and normal vector based on the preprocessed point cloud pair, and perform multi-scale hierarchical feature extraction to obtain an enhanced point feature set; calculate a bidirectional matching probability matrix using optimal transport theory based on the enhanced point feature set, filter high-confidence point pairs through spatial compatibility constraint to obtain a high-confidence matching point pair set and a matching score; perform transformation matrix estimation and quantify uncertainty based on the high-confidence matching point pair set and the matching score, and output an optimal rigid transformation and a covariance matrix thereof; perform adaptive optimization in layers based on the optimal rigid transformation and the covariance matrix thereof, dynamically adjust a search range and verify transformation parameters across scales, and output a converged accurate registration result; and construct a three-dimensional error field and statistic partition error characteristics based on the converged accurate registration result, interactively display an abnormal area through a heat map and a streamline map, and output an error analysis report and an interactive visualization interface.

[0139] Figure 3 FIG. 1 is a structural schematic diagram of an electronic device provided by an embodiment of the present application.

[0140] As shown in Figure 3 The electronic device can include a processor 610, a communications interface 620, a memory 630, and a communications bus 640, wherein the processor 610, the communications interface 620, and the memory 630 can communicate with each other through the communications bus 640. The processor 610 can invoke a logical instruction in the memory 630 to execute a three-dimensional point AI registration and tolerance analysis method for non-contact measurement.

[0141] In addition, the logical instruction in the memory 630 described above can be implemented in the form of a software function unit and sold or used as an independent product, which can be stored in a computer-readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or part of the technical solutions can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a plurality of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the methods described in the embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0142] In another aspect, the present application also provides a computer program product comprising a computer program, the computer program being stored in a non-transitory computer-readable storage medium, and the computer program being executable by a processor to enable a computer to perform the non-contact measurement-oriented three-dimensional point AI registration and tolerance analysis method provided by the above methods.

[0143] In yet another aspect, the present application also provides a non-transitory computer-readable storage medium having stored thereon a computer program, the computer program being executable by a processor to implement the non-contact measurement-oriented three-dimensional point AI registration and tolerance analysis method provided by the above methods.

[0144] The device embodiments described above are merely illustrative, wherein the units described as separate components can or can not be physically separate, and the components displayed as units can or can not be physical units, i.e., can be located in one place, or can be distributed on multiple network units. Part or all of the modules can be selected to achieve the purpose of the present embodiment scheme according to actual needs. Those skilled in the art can understand and implement without creative labor.

[0145] From the above description of the embodiments, those skilled in the art can clearly understand that the embodiments can be realized by means of software plus necessary general hardware platforms, and of course can also be realized by hardware. Based on such understanding, the above technical solutions, essentially or in other words, the part that contributes to the prior art, can be embodied in the form of a software product, which can be stored in a computer-readable storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the methods described in each embodiment or some parts of the embodiments.

[0146] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to some technical features thereof; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A three-dimensional point AI registration and tolerance analysis method for non-contact measurement, characterized in that, The application comprises the following steps: Synchronously collecting point cloud data by multiple sensors, and performing adaptive preprocessing to obtain a pair of preprocessed point clouds; Based on the pair of preprocessed point clouds, a multi-scale geometric descriptor is constructed using local curvature and normal vector, and multi-scale hierarchical feature extraction is performed to obtain an enhanced point feature set, including: based on the pair of preprocessed point clouds, a multi-scale covariance matrix of each point cloud is calculated, and multi-scale feature values and normal vectors are extracted by singular value decomposition to obtain local geometric coding; based on the local geometric coding, the key structure is focused using spatial attention, and the discriminative feature dimension is strengthened using channel attention to obtain double-attention enhanced features; based on the double-attention enhanced features, the features of different scales are fused from bottom to top through a feature propagation module to obtain an enhanced point feature set; Based on the enhanced point feature set, a bidirectional matching probability matrix is calculated using optimal transport theory, high-confidence point pairs are screened through spatial compatibility constraints to obtain a high-confidence matching point pair set and a matching score, including: based on the enhanced point feature set, the inner product similarity of the features of the source point cloud and the target point cloud is calculated, and normalization is performed to obtain a constructed bidirectional matching probability matrix: based on the bidirectional matching probability matrix, the similarity matrix is expanded and the empty set is added, and the optimal transport plan matrix with empty set constraint is solved to obtain the optimal transport matching plan matrix; based on the optimal transport matching plan matrix, geometric consistent point pairs are screened through the distance invariance under rigid transformation and the triangle proportion consistency to obtain a high-confidence matching point pair set and a matching score; Based on the high-confidence matching point pair set and the matching score, a transformation matrix is estimated, and the uncertainty is quantified to output an optimal rigid transformation and a covariance matrix thereof; Based on the optimal rigid transformation and the covariance matrix thereof, adaptive optimization is performed in layers, the search range is dynamically adjusted, and the transformation parameters are verified across scales to output a converged accurate registration result; Based on the converged accurate registration result, a three-dimensional error field is constructed, and the partition error characteristics are counted, the abnormal areas are interactively displayed through a heat map and a streamline map, and an error analysis report and an interactive visualization interface are output.

2. The non-contact measurement oriented three-dimensional point AI registration and tolerance analysis method according to claim 1, characterized in that, Synchronously collecting point cloud data by multiple sensors, and performing adaptive preprocessing to obtain a pair of preprocessed point clouds, including: The point cloud data synchronously collected by multiple sensors is subjected to time-space synchronization and unified coordinate system to obtain aligned multi-modal data; Based on the aligned multi-modal data, the sampling density is dynamically adjusted using voxelization grid, and the sampling granularity is determined according to the local curvature change of the point cloud to obtain filtered data; Based on the filtered data, combined with statistical outlier removal and normal vector consistency detection, abnormal point filtering is performed to output clean and balanced density multi-modal point cloud pairs, i.e., preprocessed point clouds.

3. The non-contact measurement-oriented three-dimensional point AI registration and tolerance analysis method according to claim 2, characterized in that, Based on the high-confidence matching point pair set and the matching score, a transformation matrix is estimated, and the uncertainty is quantified to output an optimal rigid transformation and a covariance matrix thereof, including: Based on the high-confidence matching point pair set and the matching score, the centroids of the matching point pairs are calculated, and the initial transformation is obtained through covariance matrix decomposition; Based on the initial transformation, the matching error distribution is simulated using Monte Carlo simulation to generate multiple groups of perturbed matching point pairs; Based on multiple sets of disturbance matching point pairs, the statistical distribution characteristics are calculated and the covariance matrix is established to represent the uncertainty of the transformation, and the constructed uncertainty propagation matrix is output; Based on the uncertainty propagation matrix, a candidate transformation is generated for each high-confidence matching point pair subset, the optimal transformation is selected through a voting mechanism, and the optimal rigid body transformation and its covariance matrix are output.

4. The non-contact measurement-oriented three-dimensional point AI registration and tolerance analysis method according to claim 3, characterized in that, Based on the optimal rigid body transformation and its covariance matrix, hierarchical adaptive optimization is performed, the search range is dynamically adjusted, and the transformation parameters are verified across scales, and the converged accurate registration result is output, including: Based on the optimal rigid body transformation and its covariance matrix, hierarchical accurate registration is performed on the rigid body transformation parameters from coarse to fine, and the optimized rigid body transformation parameters are obtained. Based on the optimized rigid body transformation parameters, the search range is dynamically adjusted, the point-to-local plane distance is used to accelerate convergence, and the consistency of the transformation under different resolutions is checked, and abnormal estimates are removed, and the converged accurate registration result is output.

5. The non-contact measurement-oriented three-dimensional point AI registration and tolerance analysis method according to claim 4, characterized in that, Based on the converged accurate registration result, a three-dimensional error field is constructed and the partition error characteristics are calculated, and the abnormal areas are interactively displayed through the heat map and streamline map, and the error analysis report and interactive visualization interface are output, including: Based on the converged accurate registration result, the dense corresponding error vectors of the registered point cloud and the target point cloud are calculated, and a three-dimensional error vector field is established. Based on the three-dimensional error vector field, the mean, standard deviation and extreme value are calculated in each partition to distinguish systematic errors and random noise, and the statistical characteristic analysis result is obtained. Based on the statistical characteristic analysis result, the error amplitude is displayed by the heat map and the error direction pattern is displayed by the streamline map.

6. A three-dimensional point AI registration and tolerance analysis system for non-contact measurement, characterized by, The acquisition module is used to synchronously acquire point cloud data by multiple sensors and perform adaptive preprocessing to obtain a preprocessed point cloud pair. ​ The processing module is used for constructing a multi-scale geometric descriptor using local curvature and normal vector based on the pre-processed point cloud pair, and performing multi-scale hierarchical feature extraction to obtain an enhanced point feature set, including: based on the pre-processed point cloud pair, calculating a multi-scale covariance matrix of each point cloud, and extracting multi-scale feature values and normal vectors through singular value decomposition based on the local geometric code to obtain local geometric coding; based on the local geometric code, using spatial attention to focus on key structures and channel attention to strengthen discriminative feature dimensions to obtain double-attention enhanced features; based on the double-attention enhanced features, the features of different scales are fused from bottom to top through a feature propagation module to obtain an enhanced point feature set; based on the enhanced point feature set, a bidirectional matching probability matrix is calculated using optimal transport theory, high-confidence point pairs are selected through spatial compatibility constraints to obtain a high-confidence matching point pair set and a matching score, including: based on the enhanced point feature set, the inner product similarity of the source point cloud and the target point cloud features is calculated and normalized to obtain a constructed bidirectional matching probability matrix; based on the bidirectional matching probability matrix, the similarity matrix is expanded and the empty set is increased, and the optimal transport plan matrix with empty set constraint is solved to obtain the optimal transport matching plan matrix; based on the optimal transport matching plan matrix, geometric consistent point pairs are selected through distance invariance under rigid transformation and triangle proportion consistency to obtain a high-confidence matching point pair set and a matching score; based on the high-confidence matching point pair set and the matching score, a transformation matrix is estimated and uncertainty is quantified to output an optimal rigid transformation and a covariance matrix; based on the optimal rigid transformation and the covariance matrix, adaptive optimization is performed in layers, the search range is dynamically adjusted and the transformation parameters are verified across scales to output a converged accurate registration result; based on the converged accurate registration result, a three-dimensional error field is constructed and partition error characteristics are counted, and abnormal areas are interactively displayed through a heat map and a streamline map, and an error analysis report and an interactive visualization interface are output.

7. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to realize the three-dimensional point AI registration and tolerance analysis method for non-contact measurement according to any one of claims 1 to 5.

8. A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to realize the three-dimensional point AI registration and tolerance analysis method for non-contact measurement according to any one of claims 1 to 5.

Images