A method for matching measurement data based on three-dimensional local feature constraints

Abstract

The present application relates to the field of three-dimensional machine vision and industrial automation detection technology, and particularly relates to a kind of measurement data matching method based on three-dimensional local feature constraint, comprising: executing local geometric feature extraction step: obtain the source measurement data point set and target measurement data point set of target object surface, generate local feature saliency score;Perform constraint modal evaluation step: establish the point pair correspondence between source measurement data point set and target measurement data point set;The condition number of constraint covariance matrix is calculated;Perform consistency matching verification step: based on the joint consistency entropy of geometric attribute calculation normal vector and point pair distance of point pair;Robust pose solving step is executed: based on the joint consistency entropy filtering effective matching point pair, establish weighted least squares objective function and output final registration transformation parameter.The present application can accurately extract the local feature saliency score distinguishing plane and corner point.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional machine vision and industrial automation inspection technology, specifically to a measurement data matching method based on three-dimensional local feature constraints. Background Technology

[0002] In current industrial 3D optical scanning and comparison inspection scenarios, the system needs to collect source measurement point cloud data of the workpiece and register and compare it with the standard design model; however, actual inspection objects such as automotive body panels and aircraft blades often contain large areas of smooth curved surfaces with extremely low overall geometric feature significance.

[0003] To achieve data registration, the system relies on geometric features for constraint solving. However, scanning devices often generate highly dense redundant data in flat areas, while key edge feature data is relatively sparse, resulting in severe spatial sampling unevenness. At the same time, due to the lack of effective geometric barriers in the direction of large-area smooth tangential planes, multiple degrees of freedom of rigid body transformation cannot be subject to balanced and strong constraints. This geometric feature degradation phenomenon increases the condition number of the constraint covariance matrix, leading to ill-conditioned risks in the equation system. When iteratively calculating the pose, the algorithm is prone to getting trapped in unconstrained tangential sliding oscillations and degree-of-freedom drift, making it difficult to achieve stable convergence.

[0004] Therefore, overcoming the constraint mode degradation problem caused by flat structures and effectively suppressing the risk of slippage during registration, thereby improving the robustness and convergence accuracy of 3D measurement data matching under extremely degraded geometries, has become an urgent technical problem to be solved. Summary of the Invention

[0005] The purpose of this invention is to provide a measurement data matching method based on three-dimensional local feature constraints, thereby solving the following technical problems:

[0006] It avoids unconstrained sliding effects and degree-of-freedom drift in scenarios with extremely low geometric saliency or severely degraded geometric features, such as large flat areas. It is also more likely to suppress spurious feature matching caused by local overlap or noise in extremely degraded geometric structures and ensure the uniqueness of the solution, thereby continuously improving the robustness and automatic convergence probability of pose solving.

[0007] The objective of this invention can be achieved through the following technical solutions:

[0008] A measurement data matching method based on three-dimensional local feature constraints, comprising:

[0009] Perform the local geometric feature extraction step: obtain the source measurement data point set and the target measurement data point set of the target object surface, and perform neighborhood geometric analysis on the source measurement data point set and the target measurement data point set to generate a local feature saliency score;

[0010] Perform the following constraint modal evaluation steps: establish the point-to-point correspondence between the source measurement data point set and the target measurement data point set; construct a set of geometric constraint equations describing the point-to-point correspondence; construct the constraint covariance matrix of the geometric constraint equations based on the local feature saliency score, perform stability analysis on the constraint covariance matrix, and calculate the condition number of the constraint covariance matrix.

[0011] Perform the consistency matching verification steps: Calculate the joint consistency entropy of the normal vector and the distance between the point pairs based on the geometric attributes of the point pairs;

[0012] The robust pose solving steps are as follows: based on the joint consistency entropy, valid matching point pairs are screened, a weighted least squares objective function for the rigid body transformation matrix of the constraint covariance matrix of the target registration is established, and the rigid body transformation matrix is ​​iteratively solved using the local feature saliency score and condition number to output the final registration transformation parameters.

[0013] Optionally, a local geometric feature extraction step may be performed, specifically including:

[0014] For each measurement point in the source measurement data point set and the target measurement data point set, search for its K nearest neighbor set in its spatial neighborhood;

[0015] Principal component analysis is performed on the K nearest neighbor set to calculate the eigenvalues ​​of the covariance matrix;

[0016] Based on the distribution ratio of the feature values, the geometric curvature features of the region to which the current measurement point belongs are quantified, and a scalar form of local feature saliency score is generated, wherein the local feature saliency score is used to distinguish between planar regions, edge regions and corner regions.

[0017] Optionally, perform a constrained modal evaluation step, specifically including:

[0018] Using the local feature saliency score as a weighting factor, a set of linearized constraint equations describing the relative positional relationship between the source measurement data point set and the target measurement data point set is constructed;

[0019] Calculate the constraint covariance matrix corresponding to the linearized constraint equation set;

[0020] Perform singular value decomposition on the constraint covariance matrix to obtain the maximum and minimum singular values;

[0021] The ratio of the maximum singular value to the minimum singular value is calculated to obtain the condition number of the constraint covariance matrix, wherein the condition number is used to characterize whether the six degrees of freedom of the rigid body transformation matrix are subject to equilibrium constraints under the current correspondence.

[0022] Optionally, a consistency matching verification step is performed, which specifically includes:

[0023] For each pair of points that are established, calculate the Euclidean distance between the pairs and the angle between the normal vectors of the points.

[0024] The angle between the Euclidean distance and the normal vector is used as a joint random variable, and the probability distribution of the joint random variable in the local neighborhood is calculated.

[0025] Based on the aforementioned probability distribution, the information entropy is calculated to obtain the joint consistency entropy of the normal vector and the point-pair distance;

[0026] The joint consistency entropy is used to measure the statistical compatibility of matching point pairs in terms of spatial proximity and surface orientation consistency.

[0027] Optionally, perform robust pose solving steps, specifically including:

[0028] Set a consistency entropy threshold, mark point pairs whose joint consistency entropy is higher than the consistency entropy threshold as mismatches, and mark point pairs whose joint consistency entropy is lower than or equal to the consistency entropy threshold as valid matching point pairs;

[0029] Construct the weighted least squares objective function, wherein the point pair residual weights in the effective matching point pairs are determined by the local feature saliency score;

[0030] In the process of minimizing the weighted least squares objective function, a regularization term based on the condition number is introduced to suppress the degree-of-freedom drift of the registration transformation parameters in the direction of constraint mode degradation.

[0031] Solve the weighted least squares objective function to obtain the rigid body transformation matrix under the current iteration order.

[0032] Optionally, performing the constrained modal evaluation step may also include:

[0033] After calculating the condition number of the constraint covariance matrix, the condition number is compared with a preset ill-conditioned threshold.

[0034] If the condition number is less than or equal to the preset ill-conditioned threshold, the current constraint mode is determined to be a fully constrained state, and all degrees of freedom are updated with equal weights in the iterative solution.

[0035] If the condition number is greater than the preset ill-conditioned threshold, the current constraint mode is determined to be in a degenerate state, the corresponding unlocked degree of freedom direction is identified, and a damping constraint is applied to the unlocked degree of freedom direction to limit the sliding effect.

[0036] Optionally, the method also includes:

[0037] An iterative feedback closed loop is constructed, and the rigid body transformation matrix output by the robust pose solving step is applied to the source measurement data point set to obtain the transformed source measurement data point set.

[0038] The transformed source measurement data point set is used as the new source measurement data point set to re-trigger the consistency matching verification step and the robust pose solving step.

[0039] Calculate the increment of the transformation parameters between two consecutive iterations;

[0040] If the increment of the transformation parameter is less than the preset convergence threshold, the iteration is terminated and the final registration transformation parameter is output.

[0041] If the increment of the transformation parameter is greater than or equal to the preset convergence threshold, then the next iteration continues.

[0042] Optionally, the method also includes:

[0043] Before performing the local geometric feature extraction step, the source measurement data point set and the target measurement data point set are subjected to non-uniform sampling processing;

[0044] Based on the spatial density distribution of measurement points, a density threshold is set; regions with spatial density less than or equal to the density threshold are identified as low-density regions, and measurement points in the low-density regions are retained; regions with spatial density greater than the density threshold are identified as high-density regions, and measurement points in the high-density regions are voxelized and downsampled to balance the contribution of different geometric regions to the local feature saliency score.

[0045] The beneficial effects of this invention are:

[0046] 1. This invention uses non-uniform sampling and local geometric analysis based on spatial density distribution, which can effectively balance the amount of data in different geometric regions and prevent redundant point clouds in large flat areas from overwhelming key edge features, thereby accurately extracting the saliency score of local features that distinguish between planes and corners;

[0047] 2. This invention can accurately quantify the constrained state of rigid body transformation degrees of freedom by constructing a constraint covariance matrix and calculating its condition number; for degenerate states, damping constraints are applied in the directions of unlocked degrees of freedom, which effectively suppresses the large-area unconstrained sliding effect and degree of freedom drift caused by flat structures.

[0048] 3. This invention combines the spatial distance between point pairs with the angle between their normal vectors to calculate the joint consistency entropy, which can simultaneously measure the statistical compatibility of point pairs in terms of spatial proximity and surface orientation consistency, thereby accurately eliminating false feature mismatched point pairs caused by local overlap or physical topological conflicts.

[0049] 4. This invention introduces feature saliency as a weight into the objective function solution and constructs an iterative closed loop by combining a dynamic regularization term based on condition number, which significantly enhances the pose locking capability of key geometric regions and ensures that the algorithm can still converge stably and with high accuracy when facing extremely degenerate geometric structures. Attached Figure Description

[0050] Figure 1 This is a flowchart illustrating a measurement data matching method based on three-dimensional local feature constraints provided in an embodiment of this application. Detailed Implementation

[0051] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] Please see Figure 1 A measurement data matching method based on three-dimensional local feature constraints includes: performing a local geometric feature extraction step to obtain a source measurement data point set and a target measurement data point set, and performing neighborhood geometric analysis on the source measurement data point set and the target measurement data point set to generate a local feature saliency score;

[0053] Perform the constraint modal evaluation step, construct a set of geometric constraint equations based on the preliminary verification of the point pair correspondence and the local feature significance score, perform stability analysis on the set of geometric constraint equations, and calculate the condition number of the constraint covariance matrix.

[0054] Perform a consistency matching verification step to establish a point-to-point correspondence between the source measurement data point set and the target measurement data point set, and perform a preliminary verification by calculating the joint consistency entropy of the normal vector and the point-to-point distance based on the geometric attributes of the point pairs.

[0055] The robust pose solving steps are executed, effective matching point pairs are selected based on joint consistency entropy, and the rigid body transformation matrix is ​​iteratively solved using local feature saliency scores and condition numbers to output the final registration transformation parameters.

[0056] The local geometric feature extraction step includes: searching for the K nearest neighbor set in the spatial neighborhood of each measurement point in the source and target measurement data; performing principal component analysis on the K nearest neighbor set to calculate the eigenvalues ​​of the covariance matrix; quantifying the geometric curvature features of the region to which the current measurement point belongs based on the distribution ratio of the eigenvalues, and generating a scalar form of local feature saliency score, whereby the local feature saliency score is used to distinguish between planar regions, edge regions, and corner regions.

[0057] In the local geometric feature extraction step, the system acquires source measurement data and target measurement data. Taking the 3D optical scanning and standard CAD model comparison and detection scenario of a large flat body-in-white panel in an automobile manufacturing production line as an example, the source measurement data is the point cloud of the car door outer panel collected in real time by the on-site 3D scanner, and the target measurement data is the point cloud of the standard design model of the car door outer panel.

[0058] To address the physical characteristic of the large smooth curved surface on the outer panel of the car door resulting in extremely low geometric saliency, the system performs neighborhood geometric analysis on the source measurement data and the target measurement data. Specifically, for each measurement point in the point cloud, the algorithm extracts its K nearest neighbor set within its set spatial search radius, for example, searching for the 30 nearest points; principal component analysis is performed on the K nearest neighbor set to extract three feature values ​​describing the magnitude of the local spatial distribution variance.

[0059] In order to quantify the geometric curvature characteristics of the area to which the current measurement point belongs based on the distribution ratio of eigenvalues, the system calculates the ratio of the minimum eigenvalue to the sum of all eigenvalues, thereby reflecting the degree of normal fluctuation of the local surface.

[0060] As a concrete example, suppose a measurement point is extracted from a flat area in the center of a car door. The three eigenvalues ​​obtained from its principal component analysis, arranged in descending order, are 10.0, 9.5, and 0.1. Then, the curvature of the local surface where this point is located is calculated as follows: ;

[0061] The system presets multiple dynamic threshold ranges: when the curvature is in the low significance range, the curvature... When the local feature significance score of the mapped scalar form is 0.2, the current measurement point is determined to belong to a planar region; when the curvature is in the medium range... curvature When a score of 0.6 is generated, it is determined to belong to the edge region;

[0062] When the curvature exceeds the high significance threshold, the curvature... At that time, a score of 1.0 was generated to determine that it belonged to a corner region; in this example, the measurement point in the flat region received a low score of 0.2, thus generating a local feature saliency score to distinguish different geometric regions.

[0063] As an optional embodiment, the system performs a constraint modal evaluation step; the system uses the aforementioned generated local feature saliency score as a weighting factor to construct a linearized geometric constraint equation set describing the relative positional relationship between the source measurement data and the target measurement data; specifically, this linearized geometric constraint equation set is constructed based on the point-to-surface distance error metric;

[0064] Let the points in the source measurement data be... The corresponding point in the target measurement data is The local normal vector of the target point is The rigid body transformation parameters are determined by the rotation matrix. Translation vector Composition. Under the assumption of small motions, the rotation matrix can be linearized by a first-order Taylor expansion as follows: ,in A three-dimensional rotation vector The antisymmetric matrix. Therefore, the linearized geometric constraint equation for a single matching point pair is expressed as:

[0065] ;

[0066] The system uses the local feature saliency score as a weighting coefficient. Multiplying each equation by the corresponding point pairs, we can then summarize and construct a global system of weighted linearized constraint equations. , where state variables To characterize the six-dimensional pose increment of the current rigid body transformation, the matrix... This is the Jacobian design matrix used subsequently to calculate the constraint covariance matrix;

[0067] To perform stability analysis on the geometric constraint equations, the system calculates the constraint covariance matrix corresponding to the equations and performs singular value decomposition to extract the maximum and minimum singular values. The system then calculates the ratio of the maximum to the minimum singular value to obtain the condition number of the constraint covariance matrix.

[0068] In a physical sense, the condition number directly characterizes whether the six degrees of freedom of the rigid body transformation are subject to balanced and strong constraints under the current correspondence; for example, if the maximum singular value is calculated to be 150 and the minimum singular value is 3, then the condition number is... ;

[0069] When the condition number is in a relatively high range, such as the condition number At that time, it was objectively shown that the equation set had ill-conditioning risks, that is, there was a lack of effective geometric obstruction in the direction of the large flat tangential plane of the outer panel of the car door, and there were unlocked degrees of freedom, which could easily lead to unconstrained sliding effects during the registration process.

[0070] Optionally, the system performs a consistency matching verification step; for the point-to-point correspondence between the initially established source measurement data and the target measurement data, the system calculates the joint consistency entropy of the normal vector and the point-to-point distance based on the geometric attributes of the point pairs.

[0071] The specific logic is as follows: The system not only extracts the spatial Euclidean distance between point pairs, but also simultaneously extracts the angle between the normal vectors of the point pairs; the system uses these two as joint random variables, calculates their distribution probability in the local neighborhood, and calculates their negative logarithm based on the joint probability distribution to obtain the joint consistency entropy.

[0072] In this embodiment, the system employs a two-dimensional Gaussian kernel density estimation method, or calculates the joint distribution probability based on a preset binary normal distribution model. It is assumed that the system derives the Euclidean distance based on sample statistics within a local neighborhood. Angle with normal vector They approximately follow the mean respectively The standard deviations are respectively The probability density function of the joint distribution of any given pair of points is approximately calculated as follows: (The two points are normally distributed and relatively independent.)

[0073] ;

[0074] The system substitutes it into the negative logarithm function This allows us to obtain the precise scalarized joint consensus entropy;

[0075] The joint consistency entropy measures the statistical surprise of a single pair of matching points in terms of spatial proximity and surface orientation consistency, i.e. the degree of anomaly. For example, if a pair of matching points are only 0.1 mm apart in space, but their normal vectors are 45 degrees apart, this indicates that they belong to the front and chamfered sides of the workpiece respectively. In this case, the joint distribution probability is extremely low, and the calculated joint consistency entropy will increase significantly, such as reaching 2.8, thus exposing a serious conflict in the surface topology in a physical sense.

[0076] The system executes robust pose solving steps to output the final registration transformation parameters; the system sets a consistency entropy threshold, such as 1.5, and sets the joint consistency entropy... Point pairs marked as mismatches are removed, and the joint consensus entropy is... Point pairs are retained as valid matches;

[0077] The system uses local feature saliency scores and condition numbers to iteratively solve the rigid body transformation matrix. When constructing the weighted least squares objective function, the residual weights of effective matching point pairs directly adopt the aforementioned local feature saliency scores. For example, the residual weight of corner points is 1.0, and the residual weight of planar points is only 0.2.

[0078] During the iterative process of minimizing the objective function, the system introduces a regularization term based on the condition number; when the condition number indicates the existence of a degenerate state, such as the condition number... At this time, the regularization term applies a numerical penalty in the direction of constraint mode degradation to suppress the degree of freedom drift of the transformation parameters.

[0079] At the terminal human-computer interaction and business closed loop level, the final output registration transformation parameters and registration status will be synchronized to the visualization interface of the quality inspection industrial control computer; the upper half of the interface displays the registration error distribution in the form of a three-dimensional chromatogram, and the lower half provides operation suggestions according to priority.

[0080] If the system detects that the condition number during the registration process remains in the severely degraded range, the condition number... The red alert entry will be automatically activated on the right side of the interface, and a clear maintenance work order instruction will be generated: The geometric features of the current scanned area are severely degraded, and there is a risk of slippage. It is recommended that the operator attach at least 3 highly reflective reference targets to the flat area of ​​the workpiece.

[0081] Meanwhile, the system packages the original point cloud data under high condition number conditions and the registration matrix finally confirmed by manual verification into historical archives and feeds them back to the local server. This is used to adaptively fine-tune the threshold boundary of the feature value distribution ratio of the local feature saliency score in offline mode, so as to continuously optimize the algorithm's robustness in solving extreme degraded geometric structures during subsequent equipment operation and detection.

[0082] To achieve robust registration in the aforementioned industrial scenarios, in a preferred embodiment of the present invention, a constraint modal evaluation step is performed, specifically including: using local feature saliency scores as weighting factors to construct a linearized constraint equation set describing the relative positional relationship between the source measurement data and the target measurement data; calculating the constraint covariance matrix corresponding to the linearized constraint equation set; performing singular value decomposition on the constraint covariance matrix to obtain the maximum and minimum singular values; calculating the ratio of the maximum to the minimum singular values ​​to obtain the condition number of the constraint covariance matrix, wherein the condition number is used to characterize whether the six degrees of freedom of the rigid body transformation under the current correspondence are subject to equilibrium constraints.

[0083] The consistency matching verification step includes: for each established point pair correspondence, calculating the Euclidean distance between the point pairs and the angle between the normal vectors of the point pairs; using the Euclidean distance and the angle between the normal vectors as joint random variables, and statistically analyzing the distribution probability of the joint random variables in the local neighborhood; calculating the information entropy based on the distribution probability, and obtaining the joint consistency entropy of the normal vector and the point pair distance; wherein, the joint consistency entropy is used to measure the statistical compatibility of the matching point pairs in terms of spatial proximity and surface orientation consistency.

[0084] The robust pose solving steps include: setting a consensus entropy threshold, marking point pairs with joint consensus entropy higher than the threshold as mismatches, and marking point pairs with joint consensus entropy lower than or equal to the threshold as valid matches; constructing a weighted least squares objective function, where the residual weights of point pairs in valid matches are determined by local feature saliency scores; introducing a condition number-based regularization term during the minimization of the weighted least squares objective function to suppress the degree-of-freedom drift of the transformation parameters in the direction of constraint mode degradation; and solving the weighted least squares objective function to obtain the rigid body transformation matrix under the current iteration order.

[0085] As an optional embodiment, after the system completes the extraction of local geometric features, it triggers the execution of the constraint modal evaluation step; taking the automated three-dimensional optical inspection scenario of a large aero-engine wide-chord fan blade as an example, the source measurement data is the actual measured point cloud of the blade scanned on site, and the target measurement data is the digital twin standard model of the blade;

[0086] Since the fan blades have large areas of gently curved surfaces on the base and back, and only sharp features on the leading and trailing edges, the system uses the local feature saliency score generated in the previous steps as a weighting factor to construct a set of linearized constraint equations describing the relative positional relationship between the source measurement data and the target measurement data.

[0087] Specifically, when constructing the equation system, the system directly maps the saliency score to the equation coefficient weights of the corresponding point pairs. For example, assuming a matching point pair is located at the leading edge of the blade with a local feature saliency score of 0.9, while another matching point pair is located in the gentler blade basin region with a score of 0.15, the constraint contribution of the leading edge point pair to the rigid body transformation matrix is ​​amplified to that of the blade basin point pair. This effectively enhances the ability of significant geometric regions to lock onto the overall pose.

[0088] To quantify the locking state of local features to pose transformation in different dimensions, the system calculates the constraint covariance matrix corresponding to the linearized constraint equation set and performs singular value decomposition on the constraint covariance matrix to obtain the maximum and minimum singular values ​​that reflect the energy distribution in each principal axis direction.

[0089] The system calculates the ratio of the maximum singular value to the minimum singular value to obtain the condition number of the constraint covariance matrix. This condition number is used to physically characterize whether the six degrees of freedom of the rigid body transformation are subject to equilibrium constraints under the current correspondence. For example, if the calculated maximum singular value is 180 and the minimum singular value is 1.5, then the condition number is... ;

[0090] When the condition number enters the severely degenerate region, the condition number... This indicates that there are a large number of unlocked degrees of freedom along the tangential plane of the blade surface, which can easily cause unconstrained sliding drift in the algorithm during iterative optimization.

[0091] Optionally, the system performs a consistency matching verification step to eliminate false feature correspondences caused by local overlap or noise; for each pair of points initially established, the system not only calculates the Euclidean distance between the pairs of points, but also extracts the angle between the normal vectors of the points; the system uses the Euclidean distance and the angle between the normal vectors as joint random variables and calculates the distribution probability of the joint random variable in the local neighborhood.

[0092] In this scenario, the system calculates the information entropy based on the probability distribution, and obtains the joint consistency entropy of the normal vector and the distance between point pairs;

[0093] If the Euclidean distance between a pair of matching points is only 0.05 mm, but the angle between their normal vectors is as high as 85 degrees, it indicates that one point is on the front of the blade and the other point is mismatched on the extremely thin back of the blade. At this time, their joint distribution probability is extremely low, such as 0.015. According to the information entropy calculation logic, their joint consistency entropy will increase significantly to 3.8. This joint consistency entropy accurately measures the statistical compatibility of the matching point pair in terms of spatial proximity and surface orientation consistency.

[0094] The system executes robust pose solving steps; the system sets multi-level consistency entropy thresholds, such as setting the baseline consistency entropy threshold to 2.0, and then adjusting the joint consistency entropy if it exceeds the consistency entropy threshold. Point pairs that are topologically conflicting mismatches are marked and removed; conversely, joint consensus entropy below or equal to the consensus entropy threshold is set. The point pairs are marked as valid matches;

[0095] After selecting high-quality point pairs, the system constructs a weighted least squares objective function, where the residual weights of point pairs in the effective matching are directly determined by the aforementioned local feature saliency scores. To address the ill-conditioned solution problem caused by geometric degradation, the system dynamically introduces a regularization term based on the condition number during the process of minimizing the weighted least squares objective function.

[0096] Specifically, the regularization damping coefficient is not a fixed value, but is dynamically adjusted according to the condition number: when the condition number is in the normal range, the condition number... At this time, the damping coefficient approaches 0;

[0097] When the condition number reaches as high as 120 as in the previous example, the system will dynamically amplify the damping coefficient to the reference value. The weighted least squares objective function is multiplied to apply strong numerical damping in the direction of constraint mode degradation, effectively suppressing the drift of the degrees of freedom of the transformation parameters. Finally, the system solves the weighted least squares objective function to obtain the rigid body transformation matrix under the current iteration order.

[0098] At the business loop and terminal human-machine interaction level, the above-mentioned solution status and condition number indicators will be pushed to the visualization interface of the on-site quality inspection industrial control computer in real time; the upper half of the interface displays the current registration error with a dynamic three-dimensional chromatogram and displays the real-time condition number intuitively through the dashboard; the lower half of the interface dynamically refreshes operation suggestions according to priority.

[0099] If the system detects that the condition number is in the severely degraded range for three consecutive iterations, the condition number... The right side of the interface will automatically activate the red alert entry and generate a targeted maintenance work order: The geometric features of the current blade scanning area are severely degraded, and there is a risk of tangential slip. It is recommended that the operator adjust the scanner pose to supplement the acquisition of significant geometric point clouds of the leading and trailing edges of the blade.

[0100] Meanwhile, the system packages the joint distribution probability matrix under high condition number conditions and the registration log after final manual intervention into a historical archive and sends it back to the local server. When the equipment is idle and offline, the system uses this historical archive to adaptively fine-tune the decision boundary of the consistency entropy threshold, thereby continuously optimizing the algorithm's robustness in solving extreme degraded geometric structures during subsequent equipment operation and testing.

[0101] In a preferred embodiment of the present invention, the method further includes: after calculating the condition number of the constraint covariance matrix, determining the numerical range of the condition number; if the condition number is close to 1, determining that the current constraint mode is a fully constrained state, and updating all degrees of freedom with equal weights in the pose solution; if the condition number is greater than a preset ill-conditioned threshold, determining that the current constraint mode is a degenerate state, identifying the corresponding unlocked degree of freedom direction, and applying damping constraints in that direction to limit the sliding effect;

[0102] An iterative feedback loop is constructed, and the rigid body transformation matrix output by the robust pose solving step is applied to the source measurement data to obtain the transformed source measurement data. The transformed source measurement data is used as the new input to re-trigger the consistency matching verification step and the robust pose solving step. The transformation parameter increment between two adjacent iterations is calculated. If the transformation parameter increment is less than the preset convergence threshold, the iteration is terminated and the final registration transformation parameters are output. If the transformation parameter increment is greater than or equal to the preset convergence threshold, the next round of iteration is continued.

[0103] Before performing the local geometric feature extraction step, non-uniform sampling processing is performed on the source measurement data and the target measurement data; based on the spatial density distribution of the measurement points, measurement points in low-density areas are retained, and measurement points in high-density areas are voxelized downsampled to balance the contribution of different geometric regions to the local feature saliency score.

[0104] In a preferred embodiment of the present invention, the system triggers non-uniform sampling processing of the source measurement data and the target measurement data before performing the local geometric feature extraction step;

[0105] Taking the full-size automated assembly and inspection scenario of the three-dimensional morphology of large aluminum alloy extruded wall panels in high-speed rail carriages as an example, the source measurement data is the measured point cloud of the wall panels obtained by the laser tracking scanner;

[0106] Because the panel contains a continuous flat surface several meters long and locally scattered assembly positioning holes and reinforcing rib edges, the scanning equipment often generates excessively dense redundant data in the flat areas, while the data in the critical hole edge areas is relatively sparse. In order to eliminate the interference of this spatial sampling non-uniformity on subsequent geometric analysis, the system performs differential processing logic based on the spatial density distribution of the measurement points.

[0107] Specifically, the system counts the number of points in a unit spherical neighborhood to quantify local density and sets a density baseline threshold. For example, if a local density of up to 600 points per square centimeter is detected in a flat area of ​​the wall panel, far exceeding the preset density baseline threshold of 100 points per square centimeter, the system marks the area as a high-density area and forces voxel downsampling on it. The system sets the voxel grid size to 2.0 mm and retains only the centroid point within each voxel grid, thereby reducing the density of the area to approximately 25 points per square centimeter.

[0108] Conversely, for areas such as the edges of assembly positioning holes where the local density is only 40 points per square centimeter, the system marks them as low-density areas and retains all measurement points completely;

[0109] This non-uniform sampling process effectively balances the amount of data in different geometric regions, preventing redundant point clouds in large flat areas from overwhelming edge features in subsequent principal component analysis, thereby objectively ensuring the physical accuracy of local feature saliency scores.

[0110] After completing the aforementioned feature extraction and constraint modal evaluation, the system calculates the condition number of the constraint covariance matrix and then strictly determines the numerical range of the condition number in order to decide the update strategy of the rigid body transformation matrix.

[0111] The system sets up multi-level dynamic judgment intervals: when the condition number is in a low value range, for example... When the system determines that the current constraint mode is a fully constrained state, it indicates that the mounting holes and edge features of the panel are sufficient to provide geometric obstruction in all directions of three-dimensional space. At this time, the system updates all six degrees of freedom of translation and rotation with equal weights in the pose solution.

[0112] When the scanning field of view is limited and only the flat area without holes in the wall panel is obtained, the calculated condition number will rise sharply. If the system detects that the condition number is greater than the preset ill-conditioned threshold, for example, the ill-conditioned threshold is set to 40.0, and the current measured condition number is as high as 85.0, then the current constraint mode is determined to have entered a degenerate state.

[0113] After determining that the state is degenerate, the system immediately analyzes the eigenvectors of the 6×6 constraint covariance matrix and identifies the six-dimensional motion eigenvectors corresponding to the minimum singular value. This vector is the combination of unlocked degrees of freedom in a physical sense, which may manifest as tangential sliding along a certain direction, unconstrained rotation around a certain axis, or a coupling of the two. In order to suppress the algorithm from blindly searching for non-existent geometric correspondences in the six-dimensional degenerate subspace, the system extracts the translation and rotation components of the eigenvector and forcibly applies damping constraints in the corresponding combination directions.

[0114] As a specific numerical simulation, the system sets the dynamic damping coefficient to the ratio of the current condition number to the ill-conditioned threshold, i.e. The system directly multiplies the penalty weight of 2.125 into the regularization term corresponding to the sliding direction in the weighted least squares objective function, which is equivalent to applying effective damping to the tangential sliding effect at the mathematical solution level. This effectively limits the drift of the transformation parameters in the unconstrained direction and ensures the uniqueness of the solution.

[0115] Optionally, the system constructs an iterative feedback closed loop to approximate the optimal registration accuracy; the system directly applies the rigid body transformation matrix of the current round output by the aforementioned robust pose solving steps to the source measurement data, and obtains the transformed source measurement data with updated spatial position through coordinate mapping;

[0116] The system uses the transformed source measurement data as new input to re-trigger the consistency matching verification step and the robust pose solving step, so as to re-search for high-quality point pairs with higher joint consistency entropy at a position closer to the target model; at the end of each iteration, the system calculates the transformation parameter increment between two adjacent iterations, which includes the Euclidean distance change of the translation vector and the angular deflection change of the rotation matrix.

[0117] Among them, the angular deflection change of the rotation matrix It is achieved by extracting the relative rotation matrix between two adjacent iterations. It is calculated by converting the distance, and the specific physical calculation formula is as follows:

[0118] ;

[0119] This formula precisely maps complex matrix differences into a single equivalent rotation angle in three-dimensional space, providing a rigid criterion for algorithm convergence.

[0120] The system controls business flow by comparing parameter increments with preset convergence thresholds; for example, the translation convergence threshold is set to 0.01 mm and the rotation convergence threshold is set to 0.005 degrees.

[0121] If, after the 6th iteration, the system calculates that the translation increment is 0.007 mm and the rotation increment is 0.002 degrees, satisfying the condition that the transformation parameter increment is less than the preset convergence threshold, the system determines that the registration process has reached physical stability, terminates the iteration, and outputs the final registration transformation parameters; otherwise, if in an early iteration, such as the 2nd iteration, the translation increment is as high as 0.45 mm, which is greater than or equal to the preset convergence threshold, the system continues to execute the next round of iterations.

[0122] At the level of business closed loop and terminal human-machine interaction, the above-mentioned non-uniform sampling state, real-time condition number monitoring results and iterative convergence curve will be synchronously mapped to the industrial control computer terminal of the quality inspection engineer; the upper half of the interface displays the distribution of the retained point cloud of the wall panel in a three-dimensional high-brightness view and presents the convergence process of the registration error in a dynamic chromatogram; the lower half of the interface provides working condition diagnosis and operation suggestions according to priority.

[0123] If the system detects that the number of iterations has reached the hard upper limit and the increment of the transformation parameters continues to exceed the convergence threshold, and the condition number is in a severely degenerate range, such as the condition number... An orange warning entry will automatically pop up on the right side of the interface, and a clear business intervention instruction will be generated: The current wall panel scanning area is severely lacking in geometric constraint features, and the algorithm is stuck in tangential sliding oscillation. It is recommended that the quality inspector temporarily attach at least two standard gauge blocks in the flat area of ​​the wall panel to supplement the artificial constraint features.

[0124] Meanwhile, the system packages the original point cloud density distribution logs under such non-converged conditions with the transformation matrix of the final manual registration into an abnormal scene archive and feeds it back to the cloud server. During the system's dormancy period, the cloud algorithm uses this historical archive to adaptively fine-tune the grid size benchmark value of the voxelization downsampling and the judgment boundary of the ill-conditioned threshold, thereby continuously improving the algorithm's robustness and automatic convergence probability when dealing with extremely flat structures throughout the entire life cycle of the detection equipment.

[0125] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A measurement data matching method based on three-dimensional local feature constraints, characterized in that, include: Perform the local geometric feature extraction step: obtain the source measurement data point set and the target measurement data point set of the target object surface, and perform neighborhood geometric analysis on the source measurement data point set and the target measurement data point set to generate a local feature saliency score; Specifically, it includes: For each measurement point in the source measurement data point set and the target measurement data point set, search for its K nearest neighbor set in its spatial neighborhood; Principal component analysis is performed on the K nearest neighbor set to calculate the eigenvalues ​​of the covariance matrix; Based on the distribution ratio of the feature values, the geometric curvature features of the region to which the current measurement point belongs are quantified, and a scalar form of local feature saliency score is generated, wherein the local feature saliency score is used to distinguish between planar regions, edge regions and corner regions; The constraint modal evaluation steps are as follows: Establish point-to-point correspondences between the source and target measurement data point sets; construct a set of geometric constraint equations describing these point-to-point correspondences; construct a constraint covariance matrix for the geometric constraint equations based on the local feature saliency scores, perform stability analysis on the constraint covariance matrix, and calculate the condition number of the constraint covariance matrix; specifically including: Using the local feature saliency score as a weighting factor, a set of linearized constraint equations describing the relative positional relationship between the source measurement data point set and the target measurement data point set is constructed; Calculate the constraint covariance matrix corresponding to the linearized constraint equation set; Perform singular value decomposition on the constraint covariance matrix to obtain the maximum and minimum singular values; Calculate the ratio of the maximum singular value to the minimum singular value to obtain the condition number of the constraint covariance matrix, wherein the condition number is used to characterize whether the six degrees of freedom of the rigid body transformation matrix are subject to equilibrium constraints under the current correspondence. Perform the consistency matching verification steps: Calculate the joint consistency entropy of the normal vector and the distance between the point pairs based on the geometric attributes of the point pairs; The robust pose solving steps are as follows: based on the joint consistency entropy, valid matching point pairs are screened, a weighted least squares objective function for the rigid body transformation matrix of the constraint covariance matrix of the target registration is established, and the rigid body transformation matrix is ​​iteratively solved using the local feature saliency score and condition number to output the final registration transformation parameters.

2. The measurement data matching method based on three-dimensional local feature constraints according to claim 1, characterized in that, The consistency matching verification step specifically includes: For each pair of points that are established, calculate the Euclidean distance between the pairs and the angle between the normal vectors of the points. The angle between the Euclidean distance and the normal vector is used as a joint random variable, and the probability distribution of the joint random variable in the local neighborhood is calculated. Based on the aforementioned probability distribution, the information entropy is calculated to obtain the joint consistency entropy of the normal vector and the point-pair distance; The joint consistency entropy is used to measure the statistical compatibility of matching point pairs in terms of spatial proximity and surface orientation consistency.

3. The measurement data matching method based on three-dimensional local feature constraints according to claim 1, characterized in that, The robust pose solving steps specifically include: Set a consistency entropy threshold, mark point pairs whose joint consistency entropy is higher than the consistency entropy threshold as mismatches, and mark point pairs whose joint consistency entropy is lower than or equal to the consistency entropy threshold as valid matching point pairs; Construct the weighted least squares objective function, wherein the point pair residual weights in the effective matching point pairs are determined by the local feature saliency score; In the process of minimizing the weighted least squares objective function, a regularization term based on the condition number is introduced to suppress the degree-of-freedom drift of the registration transformation parameters in the direction of constraint mode degradation. Solve the weighted least squares objective function to obtain the rigid body transformation matrix under the current iteration order.

4. The measurement data matching method based on three-dimensional local feature constraints according to claim 1, characterized in that, The execution constraint modal evaluation step further includes: After calculating the condition number of the constraint covariance matrix, the condition number is compared with a preset ill-conditioned threshold. If the condition number is less than or equal to the preset ill-conditioned threshold, the current constraint mode is determined to be a fully constrained state, and all degrees of freedom are updated with equal weights in the iterative solution. If the condition number is greater than the preset ill-conditioned threshold, the current constraint mode is determined to be in a degenerate state, the corresponding unlocked degree of freedom direction is identified, and a damping constraint is applied to the unlocked degree of freedom direction to limit the sliding effect.

5. The measurement data matching method based on three-dimensional local feature constraints according to claim 1, characterized in that, The method further includes: An iterative feedback closed loop is constructed, and the rigid body transformation matrix output by the robust pose solving step is applied to the source measurement data point set to obtain the transformed source measurement data point set. The transformed source measurement data point set is used as the new source measurement data point set to re-trigger the consistency matching verification step and the robust pose solving step. Calculate the increment of the transformation parameters between two consecutive iterations; If the increment of the transformation parameter is less than the preset convergence threshold, the iteration is terminated and the final registration transformation parameter is output. If the increment of the transformation parameter is greater than or equal to the preset convergence threshold, then the next iteration continues.

6. The measurement data matching method based on three-dimensional local feature constraints according to claim 1, characterized in that, The method further includes: Before performing the local geometric feature extraction step, the source measurement data point set and the target measurement data point set are subjected to non-uniform sampling processing; Based on the spatial density distribution of measurement points, a density threshold is set; regions with spatial density less than or equal to the density threshold are identified as low-density regions, and measurement points in the low-density regions are retained; regions with spatial density greater than the density threshold are identified as high-density regions, and measurement points in the high-density regions are voxelized and downsampled to balance the contribution of different geometric regions to the local feature saliency score.

Images