A point set registration method and device based on kernel function and feature constraints

Abstract

The present disclosure relates to a point set registration method and device based on kernel functions and feature constraints, aiming to solve the problem of insufficient robustness of traditional ICP algorithms in scenarios with noise, occlusion, and large posture deviation. The method first generates a feature-enhanced point set through joint feature extraction and mixed sampling of curvature and density; then, in the coarse registration stage, the residual scale is dynamically estimated based on the median absolute deviation, combined with the Cauchy kernel function weighted suppression of abnormal matching; finally, in the fine registration stage, feature constraints are introduced to optimize the rigid body transformation parameters to improve the overall registration accuracy and convergence stability. The method provided in the embodiment of the present disclosure can effectively reduce errors, enhance robustness and convergence speed in point set registration tasks with occlusion, missing, noise interference, and large initial value deviation, and achieve high-precision point set registration, which has good engineering application value and promotion prospects.

Description

Technical Field

[0001] The present disclosure relates to the field of computer vision technology, and in particular to a point set registration method and device based on kernel functions and feature constraints. Background Art

[0002] Point cloud registration (Point Cloud Registration) is a core technology in computer vision and 3D reconstruction. Its goal is to align two point sets in space as accurately as possible by solving rigid body transformations. The Iterative Closest Point (ICP) algorithm and its numerous variants are fundamental technologies in this field. However, standard ICP algorithms are highly dependent on initial pose estimation and lack robustness in real-world scenarios with occlusion, noise, or outliers. They are prone to falling into local optimal solutions, significantly reducing registration accuracy and stability. Summary of the Invention

[0003] In order to solve the above technical problems, the present disclosure provides a point set registration method and device based on kernel function and feature constraints.

[0004] The present disclosure provides a point set registration method based on kernel function and feature constraint, comprising: executing step one on a source point set and a target point set, the step one comprising: extracting feature points to obtain an initial feature point set; screening an enhanced feature point set from the initial feature point set based on local curvature and local density; randomly sampling the initial feature point set to obtain a supplementary feature point set, mixing the supplementary feature point set with the enhanced feature point set in a preset ratio to obtain a mixed enhanced feature point set; step two comprising: performing spatial transformation on the source point set based on the mixed enhanced feature point set, comprising performing nearest neighbor matching on the source point set and the target point set to obtain a plurality of matching point pairs; dynamically estimating the residual scale of the matching point pairs based on the median absolute deviation method, substituting the residual scale into the Cauchy kernel function to calculate a weighted factor sub-; performing weighted least squares optimization on the matching point pairs according to the weighting factor, and eliminating abnormal matching point pairs among the multiple matching point pairs; step three, comprising: executing step one on the target point set and the source point set after spatial transformation to obtain an iterative hybrid enhanced feature point set, and performing global constraint optimization based on the iterative hybrid enhanced feature point set; extracting significant geometric features of the target point set and the source point set after spatial transformation, and performing local constraint optimization based on the significant geometric features; fusing the global constraint optimization with the local constraint optimization to perform joint constraint optimization; step four, comprising: iteratively executing step three until the residual scale is less than a preset scale threshold or the number of executions of step three reaches a preset number threshold, wherein the residual scale decreases as the number of executions of step three increases.

[0005] Based on the same inventive concept, the present disclosure also provides a point set registration device based on kernel function and feature constraint, including: a feature extraction module, used to perform step one on the source point set and the target point set, the step one including: extracting feature points to obtain an initial feature point set; screening out an enhanced feature point set from the initial feature point set based on local curvature and local density; randomly sampling the initial feature point set to obtain a supplementary feature point set, mixing the supplementary feature point set with the enhanced feature point set in a preset ratio to obtain a mixed enhanced feature point set; a coarse registration module, used to perform step two, including: based on the mixed enhanced feature point set, performing spatial transformation on the source point set, including performing nearest neighbor matching on the source point set and the target point set to obtain multiple matching point pairs; dynamically estimating the residual scale of the matching point pair based on the median absolute deviation method, and substituting the residual scale into the Cauchy kernel function to calculate Obtain a weighting factor; perform weighted least squares optimization on the matching point pairs according to the weighting factor, and eliminate abnormal matching point pairs among the multiple matching point pairs; a fine alignment module is used to execute step three, including: executing step one on the target point set and the source point set after spatial transformation to obtain an iterative hybrid enhanced feature point set, and performing global constraint optimization based on the iterative hybrid enhanced feature point set; extracting significant geometric features of the target point set and the source point set after spatial transformation, and performing local constraint optimization based on the significant geometric features; fusing the global constraint optimization with the local constraint optimization to perform joint constraint optimization; an iteration module is used to execute step four, including: iteratively executing step three until the residual scale is less than a preset scale threshold or the number of executions of step three reaches a preset number threshold, wherein the residual scale decreases with an increase in the number of executions of step three.

[0006] Based on the same inventive concept, the present disclosure also provides an electronic device, comprising: a processor; a memory for storing executable instructions; wherein the processor is used to read the executable instructions from the memory and execute the executable instructions to implement any one of the methods described.

[0007] Based on the same inventive concept, the present disclosure also provides a computer-readable storage medium on which a computer program is stored. The storage medium stores the computer program, and when the computer program is executed by a processor, the processor implements any one of the methods described.

[0008] Compared with the existing technology, the technical solution provided by the present invention has the following advantages: the method provided by the present invention can effectively reduce errors, enhance robustness and convergence speed in point set registration tasks with occlusion, missing, noise interference and large initial value deviation, and achieve high-precision point set registration, which has good engineering application value and promotion prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and, together with the description, serve to explain the principles of the present disclosure.

[0010] In order to more clearly illustrate the embodiments of the present disclosure or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0011] Figure 1 A schematic diagram of a flow chart of a point set registration method based on kernel function and feature constraints provided in an embodiment of the present disclosure;

[0012] Figure 2 A schematic flow chart of another point set registration method based on kernel function and feature constraints provided in an embodiment of the present disclosure;

[0013] Figure 3 A graph showing the error trend of a registration method based on multiple point sets as the degree of rotation varies according to an embodiment of the present disclosure;

[0014] Figure 4 An error box plot based on a multiple point set registration method provided in an embodiment of the present disclosure;

[0015] Figure 5 A schematic structural diagram of a point set registration device based on kernel function and feature constraints provided in an embodiment of the present disclosure;

[0016] Figure 6 A schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present disclosure. DETAILED DESCRIPTION

[0017] In order to more clearly understand the above-mentioned purposes, features and advantages of the embodiments of the present disclosure, the scheme of the embodiments of the present disclosure will be further described below. It should be noted that the embodiments of the present disclosure and the features therein can be combined with each other in the absence of conflict.

[0018] In the following description, many specific details are set forth to facilitate a full understanding of the embodiments of the present disclosure, but the embodiments of the present disclosure may also be implemented in other ways different from those described herein; it is obvious that the embodiments in the specification are only part of the embodiments of the present disclosure, not all of the embodiments.

[0019] To address the problems existing in the standard ICP algorithm, various improvement methods exist in related technologies. For example, surface normals and feature information are introduced into the standard ICP algorithm to assist in registration; through feature point registration such as PFH and FPFH, corner points and edge points are extracted for matching, thereby reducing sensitivity to the initial pose and strengthening constraints. However, these methods generally have limitations. They mainly rely on edge points or corner points, which leads to sparse or unstable feature points in weakly structured areas (such as flat surfaces and sparse areas), low registration accuracy, and the spatial distribution of feature points is often uneven. In addition, there is also the problem of insufficient feature utilization. Even in the later stages of registration, there is a lack of explicit modeling and constraints on the structural consistency between the extracted feature points, making it difficult to correct local minor deviations remaining from the rough registration. In particular, registration drift is prone to occur in areas with occlusion or missing structure.

[0020] Relevant technicians have also proposed improving the robustness of point set registration through covariance modeling, and further introduced kernel functions to suppress the influence of outliers. However, the effectiveness of kernel functions is highly dependent on the accurate estimation of the scale of outliers. Related methods often use fixed thresholds or estimates based on standard deviations (which are easily distorted by outliers), and lack the ability to robustly adapt to dynamically changing noise and outlier levels.

[0021] In view of this, an embodiment of the present disclosure provides a point set registration method based on kernel function and feature constraints, such as Figure 1 Shown, including:

[0022] Step 1 S1 is performed on the source point set and the target point set to be registered. Step 1 S1 includes: extracting feature points to obtain an initial feature point set; filtering out an enhanced feature point set from the initial feature point set based on local curvature and local density; randomly sampling the initial feature point set to obtain a supplementary feature point set, and mixing the supplementary feature point set with the enhanced feature point set in a preset ratio to obtain a mixed enhanced feature point set. The method provided in the embodiment of the present disclosure uses the local curvature and local density of the point as dual criteria for feature point screening, ensuring that the extracted feature points are not only located in structurally significant areas, but also distributed in information-rich neighborhoods, significantly improving the stability and significance of the feature points. At the same time, in order to avoid registration failure due to too few enhanced feature point sets, the method provided in the embodiment of the present disclosure mixes the filtered feature point set with the randomly sampled point set in proportion, taking into account both registration speed and feature stability, and solving the problem that traditional feature point methods fail in non-significant areas or are unevenly distributed, resulting in weak global constraints.

[0023] Step two S2 includes: performing spatial transformation on the source point set based on the hybrid enhanced feature point set, including performing nearest neighbor matching on the source point set and the target point set to obtain multiple matching point pairs; dynamically estimating the residual scale of the matching point pairs based on the median absolute deviation method, substituting the residual scale into the Cauchy kernel function to calculate the weighting factor; performing weighted least squares optimization on the matching point pairs according to the weighting factor, and eliminating abnormal matching point pairs from the multiple matching point pairs. The embodiment of the present disclosure is based on the scale estimation of the median absolute deviation method (MAD) and the weight calculation of the Cauchy kernel function, which can effectively suppress the influence of abnormal matching in the presence of outliers, noise and occlusion. At the same time, the use of the Cauchy MAD kernel function also takes into account the rapid convergence and high robustness of point set registration.

[0024] Step three S3 includes: executing step one S1 on the target point set and the spatially transformed source point set to obtain an iterative hybrid enhanced feature point set, performing global constrained optimization based on the iterative hybrid enhanced feature point set; extracting salient geometric features from the target point set and the spatially transformed source point set, performing local constrained optimization based on the salient geometric features; and fusing global constrained optimization with local constrained optimization to perform joint constrained optimization. The disclosed embodiment utilizes a two-stage process of coarse registration (i.e., the aforementioned process of eliminating abnormal matching point pairs) and fine registration. In the coarse registration stage, a large-scale coarse alignment is quickly achieved based on the hybrid point set. In the fine registration stage, the hybrid feature weight is increased to strengthen local structure preservation, thereby improving the final registration accuracy and stability.

[0025] Step 4 S4 includes iteratively executing step 3 S3 until the residual scale is less than a preset scale threshold or the number of executions of step 3 reaches a preset number threshold, wherein the residual scale decreases as the number of executions of step 3 increases. The disclosed embodiment gradually reduces the residual scale, allowing a slightly larger tolerance in the early stages of the iteration to maintain stable convergence, and gradually tightening the constraints in the later stages to improve the final registration accuracy and avoid numerical instability.

[0026] The above-mentioned embodiments of the present disclosure provide a coarse-to-fine robust registration method based on kernel functions and feature constraints, which aims to solve the problem of insufficient robustness of traditional ICP algorithms in scenarios with noise, occlusion and large posture deviation. The method first generates a feature-enhanced point set by joint feature extraction and mixed sampling of curvature and density; then, in the coarse registration stage, the residual scale is dynamically estimated based on the median absolute deviation, combined with the Cauchy kernel function to weightedly suppress abnormal matching; finally, in the fine registration stage, feature constraints are introduced to optimize the rigid body transformation parameters to improve the overall registration accuracy and convergence stability. The method provided by the embodiments of the present disclosure can effectively reduce errors, enhance robustness and convergence speed in point set registration tasks with occlusion, missing, noise interference and large initial value deviation, and achieve high-precision point set registration, which has good engineering application value and promotion prospects.

[0027] The point set registration method provided by the embodiments of the present disclosure can be widely used in scenarios such as autonomous driving, intelligent manufacturing, medical imaging, and inverse modeling that require high robustness and high precision point set registration. For example, in the field of autonomous driving, the methods provided by the embodiments of the present disclosure can be used to align continuous point cloud frames acquired by vehicle sensors. Hybrid enhancement features effectively handle complex road structures (buildings, trees) and weakly textured areas (pavement). Cauchy-MAD weighting strongly suppresses dynamic object occlusion and environmental noise, providing strong support for real-time, accurate environmental reconstruction and positioning navigation. In the field of medical imaging disease analysis, the methods provided by the embodiments of the present disclosure can be used to match pre-operative and post-operative patient data frames, scientifically compare conditions, overcome differences in body position and equipment, robustly weighted filter imaging noise, and feature-constrained precise registration ensures precise alignment of key anatomical landmarks, enabling scientific comparison and quantitative assessment of lesions or surgical area changes. In the field of high-precision matching of scanned workpiece point clouds with computer-aided design (CAD) models or standard parts, the methods provided by the embodiments of the present disclosure can be used for feature enhancement to extract key geometric features and achieve uniform coverage. Cauchy-MAD tolerates scanning noise and slight occlusion, and dual-constrained precise registration ensures accuracy in critical dimensional areas, supporting efficient and automated defect detection (deformation, missing parts) and high-fidelity inverse modeling.

[0028] In some embodiments, the process of selecting the enhanced feature point set from the initial feature point set based on local curvature and local density specifically includes:

[0029] Point The local covariance matrix of Perform eigenvalue decomposition to obtain ,extract The smallest eigenvalue in And calculate the local curvature as ,in, is the eigenvector matrix, is the eigenvalue diagonal matrix, And both The diagonal elements of for point The local curvature of And construct the above local curvature index, which can effectively identify the geometric significant structure in the point set and realize the preliminary construction of the feature enhancement point set. The curvature value The larger the value, the more likely it is that the point is located in a feature point or edge area, ensuring that the extracted feature points are located in a structurally significant area.

[0030] The local density is used to characterize the information richness and distribution. In order to avoid the aggregation of feature points, the density index is defined as the k-nearest neighbor average Euclidean distance. The k nearest neighboring points are used to calculate their local statistical information and the local density is calculated as ,in, for point The local density of point For the point The adjacent j-th point, is the neighbor set of k neighboring points, is the set of positive real numbers.

[0031] Retention satisfaction and points, constitute the enhanced feature point set ,in, is the curvature threshold, and only points with curvature greater than this value are retained to ensure that these points are feature points. is the density threshold, and only points with a density less than this value are retained to avoid the stability of the point set being affected by too sparse density. These points constitute the enhanced feature point set, which can ensure the geometric significance of the point set and be used as feature constraints in the subsequent coarse and fine registration stages.

[0032] The initial feature point set is randomly sampled to obtain a supplementary feature point set, and the supplementary feature point set is mixed with the enhanced feature point set according to a preset ratio to obtain a mixed enhanced feature point set, including:

[0033] The supplementary feature point set and the hybrid enhanced feature point set are calculated according to the following formula:

[0034] ;

[0035] in, is a mixed enhanced feature point set, To supplement the feature point set, To enhance the feature point set, is the number of initial feature point sets, The preset ratio.

[0036] The method provided in the embodiment of the present disclosure makes the feature point retain probability , ensuring geometric saliency while simultaneously filling sparse areas with random sampling to accelerate convergence. The resulting hybrid enhanced feature point set balances geometric saliency with spatial distribution completeness. This overcomes the issues of related point set alignment techniques failing in weakly structured areas or experiencing uneven feature point distribution, laying the foundation for subsequent robust registration.

[0037] Specifically, in the above embodiment, randomly sampling the initial feature point set to obtain the supplementary feature point set includes: randomly sampling the initial feature point set to obtain the supplementary feature point set in a uniform grid division manner, thereby ensuring that the selected feature points are evenly distributed.

[0038] In some embodiments, the above-mentioned method based on median absolute deviation to dynamically estimate the residual scale of the matching point pair specifically includes:

[0039] The residual scale is calculated according to the following formula:

[0040] ;

[0041] in, is the residual scale, is the median absolute deviation scaling factor, used to convert the median absolute deviation to the estimated standard deviation level value, is the Euclidean distance between each pair of matching points, Is a computer function that returns the median of given numbers.

[0042] In some embodiments, the weighting factor obtained by substituting the residual scale into the Cauchy kernel function is calculated, specifically including:

[0043] The weighting factor is calculated according to the following formula:

[0044] ;

[0045] in, is the weighting factor, is the residual scale, is the Euclidean distance between each pair of matching points.

[0046] The weighted least squares optimization is performed on the pair errors of the matching points according to the weights, thereby suppressing the influence of abnormal matching points on the robustness of the point set registration, so that the point set registration method provided by the embodiment of the present disclosure can effectively handle high noise and outlier scenes.

[0047] In another specific embodiment, in the coarse registration stage, in order to achieve robust processing of outliers and noise, a link mechanism between the noise scale and the residual scale is established, and this is used to drive the weighted processing of matching point pairs.

[0048] This embodiment proposes to use the median absolute deviation as a robust estimator of the matching residual in the current iteration, and then construct a dynamic adaptive noise scale The scale reflects the distribution range of the "normal error" in the residual of the current point pair.

[0049] The residual of the t-th iteration is defined as:

[0050] ;

[0051] In the coarse registration stage, the fixed or standard deviation-based estimation of the noise and outlier scales is abandoned, and the median absolute deviation (MAD) is used as a robust estimator of the residual dispersion. For each iteration, the median of the Euclidean distance residuals of all matching point pairs is calculated, and then the median absolute deviation of the residuals is constructed with the median as the center:

[0052] ;

[0053] Multiply the MAD by the scaling factor After that, it can be converted into an estimate of the standard deviation dimension, and the noise scale of each iteration is dynamically calculated:

[0054] ;

[0055] in: is the MAD scaling factor, which is used to convert MAD into the estimated standard deviation level value; It is a minimum noise scale constraint to prevent numerical instability in low-noise scenarios; This is the noise scale estimation value used in the embodiment of the present disclosure. Its essence is the adaptive mapping of the residual scale to the outlier suppression capability based on MAD statistics. It can be understood that under the current residual distribution, It is the upper limit of the "normal error range".

[0056] This scale is introduced as a key parameter into the Cauchy kernel function, which assigns a dynamic weight to the residual of each pair of matching points:

[0057] ;

[0058] This function satisfies , which effectively suppresses large residual points. It prunes low-weight matching pairs and retains the valid index set. By dynamically adjusting the matching contribution through weights, the algorithm maintains a high inlier utilization rate despite outlier contamination.

[0059] The noise scale is a boundary value for weighted classification derived from the residual scale. The adaptive criterion of "outlier identification" is constructed through this scale, so that the weight function can be updated according to the current error distribution in each iteration.

[0060] The dynamic estimation Substitute the Cauchy kernel function to calculate the weight. This combination achieves double robustness, that is, MAD robustly estimates the "normal" scale of the current residual, and the Cauchy kernel is based on this scale for large residuals, which is significantly larger than This adaptive mechanism can effectively eliminate abnormal matches, greatly improving the utilization of inliers and the robustness and convergence speed of the algorithm in environments with strong noise and outliers. It avoids the inadaptability caused by the use of fixed standard deviations or empirical thresholds in traditional methods, and significantly enhances the tolerance to occlusion, outliers, and local mismatches during the point set registration process.

[0061] In some embodiments, the above-mentioned global constraint optimization based on the iterative hybrid enhanced feature point set specifically includes:

[0062] Perform global constrained optimization based on the global optimization objective function:

[0063] ;

[0064] in, is the global optimization objective function, is the i-th point in the source point set, Focus on the target point The corresponding point of is the number of points in the source point set, is the weighting factor, R and t are the rotation matrix and translation vector to be solved, that is, the transformation parameters from the source point set to the target point set.

[0065] Local constraint optimization based on salient geometric features includes: dynamically extracting high curvature extreme points (i.e., the aforementioned salient geometric features) from the spatially transformed source and target point sets as enhanced alignment targets, specifically including:

[0066] Perform local constrained optimization based on the local optimization objective function:

[0067] ;

[0068] in, is the local optimization objective function, is the number of significant geometric feature points in the source point set, is the corresponding feature point in the target point set, is the corresponding feature point in the source point set, R and t are the rotation matrix and translation vector to be solved.

[0069] Combine global constraint optimization with local constraint optimization to perform joint constraint optimization, including:

[0070] Joint constraint optimization based on the joint optimization objective function:

[0071] ;

[0072] in, To jointly optimize the objective function, is the feature point constraint weight factor, which is used to control the balance between local constraints and global constraints.

[0073] The above-mentioned embodiments of the present disclosure re-extract features from the transformed source point set, establish a correspondence between the feature points of the source point set and the feature points of the target point set, and simultaneously utilize two types of feature information for constraint optimization, including: in the global structure constraint stage, global optimization is performed based on the hybrid feature enhancement point set generated by the "curvature-density joint screening" and "random-feature complementation" mechanisms, and in the local feature enhancement constraint stage, based on significant geometric features, high curvature extreme points are dynamically extracted from the transformed point set as enhanced alignment targets, and global and local constraints are integrated to jointly optimize the objective function to form a dual-feature guided objective function. Compared with the single feature constraint method in the related art, the above-mentioned embodiments of the present disclosure improve the overall robustness of point set registration and the accuracy of rotation estimation.

[0074] In the fine registration stage, the above-mentioned embodiments of the present disclosure address and correct minor local deviations left over from the coarse registration, especially at occlusions or structural boundaries, significantly improving detail recovery and resistance to local deformation. Specifically, the above-mentioned embodiments of the present disclosure first utilize the data obtained in the coarse registration stage and the feature enhancement step, and simultaneously utilize two types of feature information for constrained optimization. Feature constraints are introduced into the objective function, and combined with a multi-scale noise attenuation strategy, the rotation and translation parameters of the source point set during spatial transformation are optimized, thereby improving point set registration accuracy and local detail recovery capabilities.

[0075] In some embodiments, the residual scale decreases as the number of executions of step 3 increases based on the following formula:

[0076] ;

[0077] in, is the residual scale when step 3 is executed for the kth time, is the residual scale when step 3 is executed for the k-1th time.

[0078] Specifically, during the first execution of step 3, the residual scale is the value at the end of step 2 (i.e., coarse registration). The residual scale is gradually attenuated during the iterations of step 3 (i.e., fine registration), with a minimum threshold of 0.1. This simulates a coarse-to-fine focusing process: initially allowing a slightly larger tolerance to maintain stable convergence, and then gradually tightening the constraints to improve the final registration accuracy while avoiding numerical instability.

[0079] In another embodiment, the residual scale may also be reduced as the number of executions of step 3 increases based on the following formula:

[0080] ;

[0081] Where s is the attenuation scale parameter, is the scaling factor, and k is the number of times step 3 is performed.

[0082] Based on the same inventive concept, one embodiment of the present disclosure also provides a method for evaluating point set registration errors. Error evaluation methods in related technologies generally rely on root mean square error (RMSE) as the primary or even sole evaluation metric. RMSE is sensitive to large local errors and fails to fully reflect the overall robustness, local consistency, and performance of registration results in complex scenarios such as occlusion and non-uniform noise, even in extreme conditions.

[0083] In view of this, the method provided in the above embodiment of the present disclosure includes:

[0084] Using Trimmed RMSE, Tukey kernel function error and Hausdorff distance error as multi-quantitative evaluation indicators of point set registration error can be used to evaluate the robustness and accuracy of the place set registration method in the above embodiment, and then comprehensively and reliably evaluate the quality of point set registration.

[0085] Specifically, Trimmed RMSE calculates the root mean square value after eliminating a certain proportion (e.g., the maximum p%) of error points, suppressing the impact of large local errors on the overall evaluation and more accurately reflecting the registration accuracy of the main area. Trimmed RMSE evaluates the error of point set registration based on the following formula:

[0086] ;

[0087] in: is the number of matching pairs remaining after removing the largest error.

[0088] Specifically, the Tukey kernel function error provides another robustness perspective error metric by performing soft thresholding on larger residual points, which can improve the robustness of point set registration. The Tukey kernel function evaluates the error of point set registration based on the following formula:

[0089] ;

[0090] Where c is the cutoff parameter, which is usually set adaptively according to the data.

[0091] Specifically, the Hausdorff distance error reflects extreme error behavior by measuring the farthest nearest neighbor distance between two point sets. In other words, it defines the maximum value of the nearest neighbor distance of the farthest point between two sets of point sets, which represents the worst matching situation and is suitable for occlusion and non-uniform noise scenes. The Hausdorff distance error is used to evaluate the error of point set registration based on the following formula:

[0092] ;

[0093] Among them, A and B are the source point set and the target point set respectively.

[0094] The above-mentioned embodiments of the present disclosure can comprehensively and systematically reflect the average performance, average accuracy, robustness and extreme matching error characteristics of point set registration through a triple index evaluation system.

[0095] In specific implementation, the method provided in the above embodiment of the present disclosure can be used for image point set processing. Figure 2 As shown in the figure, the specific process includes: 1) input source point set P and target point set Q and preprocess them, including grayscale conversion, edge detection and feature point extraction. 2) Feature point screening and blending, as well as point set feature enhancement, use curvature and density to perform local geometric analysis, extract stable and significant features, and blend them with the overall point set at a set ratio to generate a mixed point set. , this process can refer to step one in the above embodiment. 3) In the coarse registration stage, nearest neighbor matching is performed. Based on the mixed point set, the matching residual is calculated and the noise scale is dynamically adjusted based on MAD. Weighted nearest neighbor registration is used and the Cauchy MAD kernel function adaptive weights and SVD are applied to solve the rigid body transformation to suppress the influence of outliers until convergence. This process can refer to step two in the above embodiment. 4) In the fine registration stage, on the basis of coarse registration, multi-scale enhanced feature constraints are added, and the source point set and the target point set are used for refined alignment to further improve the registration accuracy. In each round of iteration, the residual scale is continuously attenuated, and the transformation parameters are optimized until the number of iterations is reached. This process can refer to steps three and four in the above embodiment. 5) Registration is completed, and the final estimated rigid transformation registration result map is output.

[0096] Figure 3 The error trend chart of each method as the degree of rotation is evaluated on the data using the average error values of Trimmed RMSE, Tukey kernel function error and Hausdorff distance error as error indicators. The experimental design is intended to simulate the common initial pose difference problem in actual point set registration, especially the significant increase in matching difficulty caused by large-angle rotation. The experimental results show that: 1) The traditional ICP algorithm has a high error rate when the rotation angle exceeds The error increases significantly after the rotation, showing the defect of being sensitive to the initial posture; 2) The FICP (Fast Iterative Closest Point, fast ICP variant) algorithm performs well in the small angle range, but lacks stability in the high angle area, and the error fluctuates violently; 3) The FastTrICP (Fast TrustRegion Iterative Closest Point, efficient point set registration algorithm) algorithm has certain robustness at some angles, but the overall error is high and the jitter is obvious; 4) The FeatureICP (ICP ablation algorithm that only introduces feature constraints) algorithm remains stable in the medium and high angle ranges, but the error value is still large; 5) The method provided in the embodiment of the present disclosure maintains the lowest average error in the entire rotation range, and its error curve is smooth without significant jitter, which verifies that the method provided in the embodiment of the present disclosure has superior robustness and accuracy advantages in dealing with a large range of posture differences and initialization uncertainties.

[0097] Figure 4 Error box plots were used to statistically analyze the truncated RMSE distribution of various methods in related technologies across multiple experimental scenarios. The plots were also used to evaluate the average registration error of each algorithm across the entire dataset. The experimental results showed that: 1) the FICP and FastTrICP algorithms exhibited multiple outlier error points, resulting in a high degree of error dispersion; 2) the traditional ICP and FeatureICP algorithms exhibited large error fluctuations, with wide boxes and significant mismatched samples. 3) The proposed method achieved a lower median error across different scenarios; the box was narrower, indicating a lower overall error level; and the number of outliers was significantly reduced, further demonstrating its robust performance advantage.

[0098] It can be understood that some of the unexplained parameter symbols and variable symbols in the above embodiments of the present disclosure are conventional parameter symbols and variable symbols known to those skilled in the art.

[0099] It should be noted that the method of the embodiment of the present application can be performed by a single device, such as a computer or server. The method of this embodiment can also be applied in a distributed scenario, where multiple devices cooperate with each other to complete the method. In such a distributed scenario, one of the multiple devices may only perform one or more steps of the method of the embodiment of the present application, and the multiple devices will interact with each other to complete the above method.

[0100] It should be noted that the above description is limited to some embodiments of the present application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in an order different from that described in the above embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require the specific order or sequential order shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0101] Based on the same inventive concept, corresponding to any of the above embodiments and methods, the present application also provides a point set registration device based on kernel function and feature constraints, such as Figure 5 Shown, including:

[0102] The feature extraction module 10 is used to perform step one on the source point set and the target point set. Step one includes: extracting feature points to obtain an initial feature point set; screening an enhanced feature point set from the initial feature point set based on local curvature and local density; randomly sampling the initial feature point set to obtain a supplementary feature point set, and mixing the supplementary feature point set with the enhanced feature point set according to a preset ratio to obtain a mixed enhanced feature point set.

[0103] The coarse registration module 20 is used to execute step 2, including: performing spatial transformation on the source point set based on the hybrid enhanced feature point set, including performing nearest neighbor matching on the source point set and the target point set to obtain multiple matching point pairs; dynamically estimating the residual scale of the matching point pairs based on the median absolute deviation method, substituting the residual scale into the Cauchy kernel function to calculate a weighting factor; performing weighted least squares optimization on the matching point pairs according to the weighting factor, and eliminating abnormal matching point pairs among the multiple matching point pairs.

[0104] The fine registration module 30 is configured to execute step three, including: executing step one on the target point set and the spatially transformed source point set to obtain an iterative hybrid enhanced feature point set, and performing global constrained optimization based on the iterative hybrid enhanced feature point set; extracting salient geometric features of the target point set and the spatially transformed source point set, and performing local constrained optimization based on the salient geometric features; and fusing the global constrained optimization with the local constrained optimization to perform joint constrained optimization.

[0105] The iterative module 40 is configured to execute step 4, including iteratively executing step 3 until the residual scale is less than a preset scale threshold or the number of executions of step 3 reaches a preset number threshold, wherein the residual scale decreases as the number of executions of step 3 increases.

[0106] The above-mentioned embodiment of the present disclosure provides a coarse-to-fine robust registration device based on kernel functions and feature constraints, which aims to solve the problem of insufficient robustness of traditional ICP algorithms in scenarios with noise, occlusion and large posture deviation. The device first generates a feature-enhanced point set through joint feature extraction and mixed sampling of curvature and density; then, in the coarse registration stage, it dynamically estimates the residual scale based on the median absolute deviation, and combines the Cauchy kernel function to weightedly suppress abnormal matching; finally, in the fine registration stage, it introduces feature constraints to optimize the rigid body transformation parameters to improve the overall registration accuracy and convergence stability. The device provided by the embodiment of the present disclosure can effectively reduce errors, enhance robustness and convergence speed in point set registration tasks with occlusion, missing, noise interference and large initial value deviation, and achieve high-precision point set registration, which has good engineering application value and promotion prospects.

[0107] For the convenience of description, the above devices are described as being divided into various modules according to their functions. Of course, when implementing this application, the functions of each module can be implemented in the same or multiple software and / or hardware.

[0108] The device of the above embodiment is used to implement the corresponding point set registration method based on kernel function and feature constraint in any of the above embodiments, and has the beneficial effects of the corresponding method embodiment, which will not be repeated here.

[0109] Figure 6 A schematic structural diagram of an electronic device provided by an embodiment of the present disclosure is shown.

[0110] like Figure 6 As shown, the electronic device may include a processor 1101 and a memory 1102 storing computer program instructions.

[0111] Specifically, the processor 1101 may include a central processing unit (CPU) or an application specific integrated circuit (ASIC), or may be configured to implement one or more integrated circuits of the embodiments of the present application.

[0112] Memory 1102 may include a large-capacity memory for information or instructions. By way of example and not limitation, memory 1102 may include a hard disk drive (HDD), a floppy disk drive, flash memory, an optical disk, a magneto-optical disk, a magnetic tape, or a universal serial bus (USB) drive, or a combination of two or more of these. Where appropriate, memory 1102 may include removable or non-removable (or fixed) media. Where appropriate, memory 1102 may be internal or external to the integrated gateway device. In a specific embodiment, memory 1102 is a non-volatile solid-state memory. In a specific embodiment, memory 1102 includes a read-only memory (ROM). Where appropriate, the ROM may be a mask-programmed ROM, a programmable ROM (PROM), an erasable PROM (EPROM), an electrically erasable PROM (EEPROM), an electrically alterable ROM (EAROM), or flash memory, or a combination of two or more of these.

[0113] The processor 1101 reads and executes the computer program instructions stored in the memory 1102 to perform the steps of the point set registration method based on kernel function and feature constraint provided in the embodiment of the present disclosure.

[0114] In one example, the electronic device may further include a transceiver 1103 and a bus 1104. Figure 6 As shown, the processor 1101 , the memory 1102 and the transceiver 1103 are connected via a bus 1104 and communicate with each other.

[0115] The bus 1104 may include hardware, software, or both. By way of example, and not limitation, the bus may include an Accelerated Graphics Port (AGP) or other graphics bus, an Extended Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), a HyperTransport (HT) interconnect, an Industrial Standard Architecture (ISA) bus, an InfiniBand interconnect, a Low Pin Count (LPC) bus, a memory bus, a MicroChannel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local Bus (VLB) bus, or other suitable buses, or a combination of two or more of these. Where appropriate, the bus 1104 may include one or more buses. Although embodiments of the present application describe and illustrate a particular bus, this application contemplates any suitable bus or interconnect.

[0116] The following is an embodiment of a computer-readable storage medium provided in an embodiment of the present disclosure. The computer-readable storage medium and the point set registration method based on kernel functions and feature constraints in the above-mentioned embodiments belong to the same inventive concept. For details not fully described in the embodiment of the computer-readable storage medium, please refer to the embodiment of the point set registration method based on kernel functions and feature constraints mentioned above.

[0117] This embodiment provides a storage medium containing computer-executable instructions. When the computer-executable instructions are executed by a computer processor, they are used to perform a point set registration method based on kernel functions and feature constraints.

[0118] Of course, the storage medium containing computer-executable instructions provided in an embodiment of the present disclosure is not limited to the above method operations, and its computer-executable instructions can also execute related operations in the point set registration method based on kernel functions and feature constraints provided in any embodiment of the present disclosure.

[0119] Through the above description of the implementation methods, those skilled in the art can clearly understand that the present disclosure can be implemented with the help of software and necessary general-purpose hardware, and of course it can also be implemented through hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present disclosure, or the part that contributes to the existing technology, can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory (FLASH), hard disk or optical disk, etc., including a number of instructions for enabling a computer cloud platform (which can be a personal computer, server, or network cloud platform, etc.) to execute the point set registration method based on kernel function and feature constraints provided in each embodiment of the present disclosure.

[0120] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises", "includes" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprises a..." do not exclude the presence of other identical elements in the process, method, article or device that includes the above elements.

[0121] The foregoing are merely specific embodiments of the present disclosure, intended to enable those skilled in the art to understand and implement the present disclosure. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present disclosure. Therefore, the present disclosure is not limited to the foregoing embodiments, but is intended to be construed in the broadest manner consistent with the principles and novel features disclosed herein.

Claims

1. A point set registration method based on kernel function and feature constraint, characterized in that: include: Performing step 1 on the source point set and the target point set, the step 1 comprising: extracting feature points to obtain an initial feature point set; screening an enhanced feature point set from the initial feature point set based on local curvature and local density; randomly sampling the initial feature point set to obtain a supplementary feature point set, and mixing the supplementary feature point set with the enhanced feature point set in a preset ratio to obtain a mixed enhanced feature point set; Step 2 includes: based on the hybrid enhanced feature point set, performing spatial transformation on the source point set, including performing nearest neighbor matching on the source point set and the target point set to obtain multiple matching point pairs; dynamically estimating the residual scale of the matching point pairs based on the median absolute deviation method, substituting the residual scale into the Cauchy kernel function to calculate a weighting factor; performing weighted least squares optimization on the matching point pairs according to the weighting factor, and eliminating abnormal matching point pairs from the multiple matching point pairs; Step three includes: performing step one on the target point set and the spatially transformed source point set to obtain an iterative hybrid enhanced feature point set, and performing global constrained optimization based on the iterative hybrid enhanced feature point set; extracting significant geometric features of the target point set and the spatially transformed source point set, and performing local constrained optimization based on the significant geometric features; and fusing the global constrained optimization with the local constrained optimization to perform joint constrained optimization; Step 4 includes iteratively executing step 3 until the residual scale is less than a preset scale threshold or the number of executions of step 3 reaches a preset number threshold, wherein the residual scale decreases as the number of executions of step 3 increases.

2. The method according to claim 1, characterized in that The step of selecting an enhanced feature point set from the initial feature point set based on local curvature and local density includes: Point The local covariance matrix of Perform eigenvalue decomposition to obtain ,extract The smallest eigenvalue in And calculate the local curvature as ,in, is the eigenvector matrix, is the eigenvalue diagonal matrix, And both The diagonal elements of for point The local curvature of Select and point The local density is calculated based on the nearest k neighboring points: ,in, for point The local density of point For the point The adjacent j-th point, is the neighbor set of k neighboring points, is the set of positive real numbers; Retention satisfaction and points, constituting the enhanced feature point set ,in, is the curvature threshold, is the density threshold; The randomly sampling the initial feature point set to obtain a supplementary feature point set, and mixing the supplementary feature point set with the enhanced feature point set according to a preset ratio to obtain a mixed enhanced feature point set, comprising: The supplementary feature point set and the hybrid enhanced feature point set are calculated according to the following formula: ； in, is the hybrid enhanced feature point set, is the supplementary feature point set, is the enhanced feature point set, is the number of the initial feature point set, is the preset ratio.

3. The method according to claim 1 or 2, characterized in that The randomly sampling the initial feature point set to obtain a supplementary feature point set includes: The supplementary feature point set is obtained by randomly sampling the initial feature point set in a uniform grid division manner.

4. The method according to claim 1, wherein The dynamically estimating the residual scale of the matching point pair based on the median absolute deviation method includes: The residual scale is calculated according to the following formula: ； in, is the residual scale, is the median absolute deviation scaling factor, is the Euclidean distance between each pair of matching points, is a function that returns the median of given values.

5. The method according to claim 4, characterized in that Substituting the residual scale into the Cauchy kernel function to calculate the weighting factor includes: The weighting factor is calculated according to the following formula: ； in, is the weighting factor, is the residual scale, is the Euclidean distance between each pair of matching points.

6. The method according to claim 5, characterized in that The performing global constraint optimization based on the iterative hybrid enhanced feature point set includes: The global constrained optimization is performed based on the global optimization objective function: ； in, is the global optimization objective function, is the i-th point in the source point set, Focus on the target point The corresponding point of is the number of points in the source point set, is the weighting factor, R and t are the rotation matrix and translation vector to be solved; The performing local constrained optimization based on the significant geometric features includes: The local constrained optimization is performed based on the local optimization objective function: ； in, is the local optimization objective function, is the number of significant geometric feature points in the source point set, is the corresponding feature point in the target point set, is the corresponding feature point in the source point set, R and t are the rotation matrix and translation vector to be solved; The fusing of the global constraint optimization and the local constraint optimization to perform joint constraint optimization includes: The joint constraint optimization is performed based on the joint optimization objective function: ； in, is the joint optimization objective function, and λ is the feature point constraint weight factor.

7. The method according to claim 1, characterized in that The residual scale decreases as the number of executions of step 3 increases based on the following formula: ； in, is the residual scale when step 3 is executed for the kth time, is the residual scale when step 3 is executed for the k-1th time.

8. A point set registration device based on kernel function and feature constraint, characterized in that: include: A feature extraction module is configured to perform step 1 on the source point set and the target point set, wherein step 1 includes: extracting feature points to obtain an initial feature point set; screening an enhanced feature point set from the initial feature point set based on local curvature and local density; randomly sampling the initial feature point set to obtain a supplementary feature point set, and mixing the supplementary feature point set with the enhanced feature point set in a preset ratio to obtain a mixed enhanced feature point set; A coarse registration module is configured to execute step 2, comprising: performing spatial transformation on the source point set based on the hybrid enhanced feature point set, including performing nearest neighbor matching on the source point set and the target point set to obtain a plurality of matching point pairs; dynamically estimating the residual scale of the matching point pairs based on a median absolute deviation method, substituting the residual scale into a Cauchy kernel function to calculate a weighting factor; performing weighted least squares optimization on the matching point pairs according to the weighting factor, and eliminating abnormal matching point pairs from the plurality of matching point pairs; A fine registration module is configured to execute step three, including: executing step one on the target point set and the spatially transformed source point set to obtain an iterative hybrid enhanced feature point set, and performing global constrained optimization based on the iterative hybrid enhanced feature point set; extracting significant geometric features of the target point set and the spatially transformed source point set, and performing local constrained optimization based on the significant geometric features; and fusing the global constrained optimization with the local constrained optimization to perform joint constrained optimization; An iterative module is configured to execute step 4, comprising: iteratively executing step 3 until the residual scale is less than a preset scale threshold or the number of executions of step 3 reaches a preset number threshold, wherein the residual scale decreases as the number of executions of step 3 increases.

9. An electronic device, characterized in that: include: processor; A memory for storing executable instructions; wherein the processor is used to read the executable instructions from the memory and execute the executable instructions to implement the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: The storage medium stores a computer program, and when the computer program is executed by a processor, the processor implements the method according to any one of claims 1 to 7.

Images