A multi-scale fusion point cloud feature extraction and adaptive matching method

Abstract

The present invention relates to a multi-scale fusion point cloud feature extraction and adaptive matching method. The present invention collects, filters, and voxelizes point clouds; determines the normal vector of each point in the point cloud; screens effective feature points according to the normal vectors of the point cloud points; performs multi-scale feature calculation on the retained effective feature points, and performs secondary feature point screening; classifies the effective feature points screened for the second time into concave-convex feature points and strip feature points; further screens the strip feature points according to the degeneration direction; determines the corresponding feature sequence of the concave-convex feature points and the remaining strip feature points; calculates the similarity of the feature sequence for candidate points in the source point cloud and the target point cloud, performs point pair matching according to the similarity, and traverses and screens out all matching points; and uses adaptive weights to control the contribution of three types of matching points, namely, featureless point cloud points, concave-convex feature points, and strip feature points, in the source point cloud and the target point cloud to the point cloud registration residual, and performs registration.

Description

Technical Field

[0001] The present invention relates to the technical field of positioning and mapping, and in particular to a multi-scale fusion point cloud feature extraction and adaptive matching method. Background Art

[0002] Point cloud registration is a key issue in SLAM. Its goal is to align point clouds in different coordinate systems by calculating a transformation matrix. The registration process is typically divided into two stages: coarse registration and fine registration. Coarse registration involves rapidly estimating an initial transformation matrix when the relative positions of two point clouds are completely unknown. Its core goal is to minimize the initial pose discrepancy and provide a good starting point for subsequent fine registration. Methods such as feature-based, probability-based, and deep learning-based methods are used. Fine registration, based on the initial transformation provided by coarse registration, iteratively minimizes the geometric error or probability distribution difference between point clouds to achieve high-precision alignment. Numerous methods have been used to achieve fine registration. Among them, the iterative closest point (ICP) algorithm proposed by Besl et al. is the most widely used. The ICP algorithm iteratively updates the transformation matrix by minimizing a predefined cost function until the cost function converges. The original ICP algorithm is slow and prone to local optima, leading to the introduction of various variants such as point-to-point, point-to-line, and point-to-plane. In recent years, several excellent registration methods have emerged. Dellenbach et al. proposed a continuous-time iterative closest point (ICP) method, which achieves smoother pose transformations by adding continuous-time constraints to the optimization function. Ferrari et al. constructed a kd-tree structure based on principal component analysis for point cloud registration, using six parameters to achieve good performance under different operating conditions. However, when located in a long and narrow tunnel, the environment along the tunnel's direction of travel is similar, and many point cloud registration methods are no longer applicable. The side walls of the mine have uneven textures, and when faced with degradation, the inner wall features are reduced. This forces the traditional iterative closest point registration process to rely on a large number of point clouds without significant geometric attributes, reducing the proportion of valid feature points in the point cloud registration and leading to increased registration errors. Although feature-driven methods such as LeGo-LOAM improve robustness through edge / plane point classification and matching, they use single-line scanning roughness indicators for feature extraction, which cannot effectively capture the concave and convex features in mine tunnels that span multiple lines and have different directions. On the other hand, they only divide the point cloud into edge points and plane points, and do not perform similarity matching on feature points. Edge points are easily mistakenly matched to other edge points. Summary of the Invention

[0003] In order to solve the above technical problems or at least partially solve the above technical problems, the present invention provides a multi-scale fusion point cloud feature extraction and adaptive matching method.

[0004] In a first aspect, the present invention provides a multi-scale fusion point cloud feature extraction and adaptive matching method, comprising:

[0005] Collect the point cloud to be adaptively matched, and pre-process the point cloud to be adaptively matched by filtering and voxelization;

[0006] Determine the local neighborhood of each point in the point cloud and use the local neighborhood of the point cloud points to calculate the normal vector of each point;

[0007] Screen effective feature points based on the normal vectors of point cloud points;

[0008] Perform multi-scale feature calculation on the retained valid feature points and perform secondary feature point screening;

[0009] Classify the effective feature points screened for the second time into concave-convex feature points and strip feature points;

[0010] The strip feature points are further screened out according to the degradation direction;

[0011] Determine the corresponding feature series of the concave and convex feature points and the remaining strip feature points; the feature series of the concave and convex feature points includes: the average value of the sum of the normal vector angles corrected for direction consistency within the normalized annular neighborhood of the concave and convex feature points, the radius of the normalized neighborhood with the highest feature significance, and the curvature characteristics of the concave and convex feature points; the feature series of the strip feature points includes: the average value of the sum of the normal vector angles corrected for direction consistency within the normalized annular neighborhood of the concave and convex feature points, the radius of the normalized neighborhood with the highest feature significance, linearity, and directional consistency measurement;

[0012] For the candidate points in the source point cloud and the target point cloud, the similarity of the feature sequence is calculated, point pairs are matched based on the similarity, and all matching points are traversed and filtered out;

[0013] Adaptive weights are used to control the contribution of three types of matching points, namely, featureless point cloud points, concave-convex feature points, and strip feature points, in the source point cloud and target point cloud to the point cloud registration residuals for registration.

[0014] Furthermore, the process of calculating the normal vector of each point using the local neighborhood of the point cloud includes:

[0015] Compute the covariance matrix of each point using its neighboring points in its local neighborhood:

[0016] ;

[0017] Compute the eigenvalues ​​and eigenvectors of the covariance matrix:

[0018] ;

[0019] in, Point cloud point The local neighborhood of The center of mass of the point cloud, the coordinates of the center of mass are in the local neighborhood The average of the coordinates of the midpoints, Point cloud point The covariance matrix of Represents the covariance matrix The eigenvalues ​​and eigenvectors of

[0020] The eigenvector corresponding to the minimum eigenvalue is taken as the normal vector of the point cloud point.

[0021] Furthermore, the screening of effective feature points according to the normal vectors of the point cloud points includes: obtaining the normal vectors of the point cloud points and the neighborhood points corresponding to each point cloud point;

[0022] Perform directional consistency correction on the normal vectors of point cloud points;

[0023] Calculate the average of the sum of the direction consistency corrected normal vectors of each point cloud point and the angles between the direction consistency corrected normal vectors of the neighboring points corresponding to each point cloud point:

[0024] ;

[0025] in, Point cloud point The local neighborhood of Point cloud points and Directional consistency correction normal vector;

[0026] When the point cloud When the point cloud points are less than the set threshold, the point cloud points are filtered out. When the value is greater than the set threshold, the point cloud point is retained as a valid feature point.

[0027] Furthermore, multi-scale feature calculation is performed on the retained valid feature points, and feature points are screened out twice, including:

[0028] For each valid feature point, a multi-scale annular neighborhood is divided using circles of different radii with the valid feature point as the center.

[0029] Calculate the average of the sum of the angles between the directional consistency-corrected normal vectors in different annular neighborhoods:

[0030] ;

[0031] Where x represents the index of the x-th ring neighborhood, Here it is represented by a dot The neighborhood corresponding to the xth ring centered at is, Point cloud points and Directional consistency correction normal vector;

[0032] The average value of the normal vector angle is corrected by the direction consistency within each ring, and the neighborhood with the highest feature significance is selected;

[0033] Traversing each valid feature point, there is a neighborhood with the highest feature significance of any valid feature point. When the neighborhood point is larger, the point cloud point is deleted, and the remaining valid feature points are the second screening valid feature points, and the second screening valid feature points are retained. and the radius of the neighborhood with the highest feature significance .

[0034] Furthermore, according to the linearity The effective feature points screened for the second time are classified into concave-convex feature points and strip feature points: the linearity is calculated as follows:

[0035] ;

[0036] in, is the maximum eigenvalue, The linearity is less than the maximum eigenvalue and not less than the minimum eigenvalue. When the linearity threshold is greater than the set value, the effective feature points are classified as bar feature points in the second screening. When the linearity is less than the set threshold, the effective feature points are classified as concave and convex feature points after the second screening.

[0037] Furthermore, the method of further screening out the strip feature points according to the degradation direction includes:

[0038] Calculate the directional consistency measure of the strip feature point direction and the degenerate direction: ,

[0039] in, Indicates the maximum eigenvalue of the bar feature point The corresponding eigenvector represents the direction of the bar feature point. represents the degenerate direction unit vector, Represents the dot product operation. When the direction consistency measure is less than the direction consistency measure threshold, it means that the strip feature point is basically consistent with the degradation direction, and the strip feature is a degradation-sensitive feature point and is screened out.

[0040] Furthermore, the point cloud registration residual is minimized by using three types of point pairs, namely, featureless point cloud points, concave-convex feature points, and strip feature points, in the source point cloud and the target point cloud:

[0041] ;

[0042] Among them, R is the rotation matrix in the registered transformation relationship, and t is the translation matrix; Represents the featureless point cloud points in the source point cloud, represents the concave and convex feature points in the source point cloud, Represents the strip feature points in the source point cloud; Represents the paired featureless point cloud points in the target point cloud, represents the paired concave and convex feature points in the target point cloud, Represents paired bar feature points in the target point cloud;

[0043] are the corresponding adaptive weights for featureless point cloud points, concave-convex feature points, and strip feature points respectively;

[0044] Be the minimum value and avoid division by 0;

[0045] Update using the Levenberg-Marquardt algorithm , repeat the iteration until Convergence, registration completed.

[0046] Furthermore, adaptive weight is calculated as follows:

[0047] ;

[0048] Adaptive weights It is based on the number of points in each type of point cloud The smaller the number of categories, the higher the adaptive weight will be, which can avoid one type of point cloud dominating the optimization process and balance the contribution of different point clouds.

[0049] In a second aspect, the present invention provides a multi-scale fusion point cloud feature extraction and adaptive matching device, comprising: at least one processing unit, the processing unit being connected to a storage unit via a bus unit, the storage unit storing a computer program, and the processing unit implementing the multi-scale fusion point cloud feature extraction and adaptive matching method by running the computer program stored in the storage unit.

[0050] In a third aspect, the present invention provides a computer-readable storage medium, which stores a computer program. When the computer program is executed, it implements the multi-scale fusion point cloud feature extraction and adaptive matching method.

[0051] The above technical solution provided by the embodiment of the present invention has the following advantages compared with the prior art:

[0052] This application proposes a multi-scale fusion point cloud feature extraction and adaptive matching method, which uses the attribute filtering of point cloud points to extract concave-convex feature points in the concave and convex areas of the mine point cloud and the strip feature points in the crack and gradient fault areas, and divides them by type. The strip feature points are further filtered considering the influence of the degradation direction; and common features are extracted by multi-scale local feature extraction and different features are extracted for different types of feature points. The contribution of concave-convex feature points and strip feature points in the registration process is improved by combining adaptive weights, and the original scarce structural feature points are used to improve the robustness of the registration, strengthen the robustness of point cloud registration in environments such as mines, and improve the positioning and mapping accuracy in complex environments. At the same time, the present invention has no special requirements for hardware equipment, has a wide range of applications, and can cope with a variety of complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0054] In order to more clearly illustrate the embodiments of the present invention 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 paying any creative labor.

[0055] Figure 1 A flowchart of a multi-scale fusion point cloud feature extraction and adaptive matching method provided by an embodiment of the present invention;

[0056] Figure 2 A flowchart for calculating the normal vector of each point in the local neighborhood of a point cloud point provided by an embodiment of the present invention;

[0057] Figure 3 A flowchart for screening effective feature points based on the normal vectors of point cloud points provided by an embodiment of the present invention;

[0058] Figure 4 A flowchart of performing multi-scale feature calculation on retained valid feature points and performing secondary feature point screening provided by an embodiment of the present invention;

[0059] Figure 5 Schematic diagram of a multi-scale fusion point cloud feature extraction and adaptive matching device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0060] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0061] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or apparatus comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or apparatus comprising the element.

[0062] Example 1

[0063] See Figure 1 As shown, the multi-scale fusion point cloud feature extraction and adaptive matching method provided by this application includes:

[0064] Step 1: Collect the point cloud to be adaptively matched and pre-process it through filtering and voxelization. Filtering removes noise points from the point cloud, using methods such as statistical filtering or radius filtering. Voxelization downsamples the point cloud to make the point cloud data more uniform and smaller in dimension, reducing computational effort.

[0065] Step 2: Determine the local neighborhood of each point in the point cloud, and use the local neighborhood of the point cloud point to calculate the normal vector of each point, such as Figure 2 As shown, the process includes:

[0066] Compute the covariance matrix of each point using its neighboring points in its local neighborhood:

[0067] ;

[0068] Compute the eigenvalues ​​and eigenvectors of the covariance matrix:

[0069] ;

[0070] in, Point cloud point The local neighborhood of The center of mass of the point cloud, the coordinates of the center of mass are in the local neighborhood The average of the coordinates of the midpoints, Point cloud point The covariance matrix of Represents the covariance matrix The eigenvalues ​​and eigenvectors of .

[0071] Take the eigenvector corresponding to the minimum eigenvalue as the normal vector of the point cloud point The eigendecomposition of the covariance matrix will obtain three eigenvalues ​​and corresponding eigenvectors, where the eigenvector corresponding to the minimum eigenvalue is the normal vector.

[0072] Step 3: Screen effective feature points based on the normal vector of the point cloud point, such as Figure 3 As shown, including:

[0073] Get the normal vector of the point cloud point and the neighborhood points corresponding to each point cloud point;

[0074] Perform directional consistency correction on the normal vector of the point cloud point to obtain the directional consistency corrected normal vector: normal vector The directional consistency correction normal vector is ;

[0075] Calculate the average of the sum of the direction consistency corrected normal vectors of each point cloud point and the angles between the direction consistency corrected normal vectors of the neighboring points corresponding to each point cloud point:

[0076] ;

[0077] when When it is less than the set threshold, it means that the direction consistency correction normal vector around the point does not change much, the neighborhood is flat, and it is screened out. When the value is greater than the set threshold, it indicates that the normal vector around the point has changed significantly due to directional consistency correction, which may be a groove, a protrusion, a crack, or an edge. It is retained as a valid feature point. By eliminating flat areas, the subsequent calculation amount is reduced, while focusing on areas with significant geometric changes, the efficiency and accuracy of point cloud matching are improved.

[0078] Step 4: Perform multi-scale feature calculation on the retained valid feature points and perform secondary feature point screening, such as Figure 4 As shown, including:

[0079] For each valid feature point, a multi-scale annular neighborhood is divided using circles of different radii with the valid feature point as the center. The initial circle radius is 3-5 times the average spacing of the point cloud points, and the radius of the circle dividing the annular neighborhood increases by 2 times.

[0080] Calculate the average of the sum of the angles between the directional consistency-corrected normal vectors in different annular neighborhoods:

[0081] ;

[0082] Where x represents the index of the x-th ring neighborhood, Here it is represented by a dot The neighborhood corresponding to the xth ring is the center.

[0083] The average value of the normal vector angle is corrected by the direction consistency in each ring, and the neighborhood with the highest feature significance is selected: as the number of rings increases, when the corresponding When it changes from large to small, the previous ring neighborhood is used as the neighborhood with the highest feature significance.

[0084] Traversing each valid feature point, there is a neighborhood with the highest feature significance of any valid feature point. When the neighborhood point is larger, the point cloud point is deleted, and the remaining valid feature points are the second screening valid feature points, and the second screening valid feature points are retained. and the radius of the neighborhood with the highest feature significance .

[0085] Step 5: Classify the effective feature points screened for the second time into concave-convex feature points and strip feature points.

[0086] The biggest difference between concave-convex feature points and strip feature points is that the maximum eigenvalue of the strip feature point is much larger than the other eigenvalues, and the directionality is obvious, while the difference between the maximum eigenvalue and the secondary eigenvalue of the concave-convex feature point is relatively small. The effective feature points screened for the second time are classified into concave and convex feature points and strip feature points:

[0087] The linearity is calculated as follows:

[0088] ;

[0089] in, is the maximum eigenvalue, The linearity is less than the maximum eigenvalue and not less than the minimum eigenvalue. When the linearity threshold is greater than the set value, the effective feature points are classified as bar feature points in the second screening. When the linearity is less than the set threshold, the effective feature points are classified as concave and convex feature points after the second screening.

[0090] Step 6: Further filter out the strip feature points according to the degradation direction.

[0091] During the specific implementation process, when the direction of the strip feature point is consistent with the degradation direction, it is easily affected by degradation and needs to be screened out for the third time. The degradation direction refers to the direction of the missing geometric constraints caused by structural repeatability or sensor motion mode in the environment. For example, in tunnel and corridor scenes, the structure along the axis is highly repeated and the horizontal features are sparse. Features along the tangent direction of the circle in the rotationally symmetric structure are difficult to distinguish. Point cloud registration solves the rotation matrix R and translation vector t by minimizing the distance between corresponding points: , are the matching points in the source and target point clouds. When a large number of feature points align with the degenerate direction, the Jacobian matrix has a null space, resulting in unobservable translational components along the degenerate direction and unobservable rotational angles around the degenerate direction. Registration errors in the degenerate direction accumulate with iterations, causing error propagation. This manifests as a continuous increase in positioning error along the degenerate direction, resulting in displacement drift, and incorrectly matched points being misidentified as correct matches due to their consistent orientation, resulting in mismatches.

[0092] Methods for further filtering out strip feature points according to the degradation direction include:

[0093] Calculate the directional consistency measure of the strip feature point direction and the degenerate direction: ,

[0094] in, Indicates the maximum eigenvalue of the bar feature point The corresponding eigenvector represents the direction of the bar feature point. represents the degenerate direction unit vector, Represents the dot product operation. When the direction consistency measure is less than the direction consistency measure threshold, it means that the strip feature point is basically consistent with the degradation direction. The strip feature point is a degradation-sensitive feature point and is screened out.

[0095] Step 7: Determine the corresponding feature sequence of the concave and convex feature points and the remaining strip feature points. This includes:

[0096] The characteristic sequence of concave and convex feature points includes: the average value of the sum of the normalized direction consistency correction normal vector angles within the annular neighborhood of the concave and convex feature points , the radius of the neighborhood with the highest normalized feature significance And the curvature characteristics of concave and convex feature points.

[0097] For concave and convex feature points, the curvature feature is calculated as follows:

[0098] ;

[0099] in, is the minimum eigenvalue

[0100] The average value of the sum of the angles of the normal vectors corrected for directional consistency within the annular neighborhood of the normalized concave and convex feature points as follows:

[0101] ;

[0102] The radius of the neighborhood with the highest normalized feature significance as follows:

[0103] ;

[0104] Therefore, the feature sequence corresponding to the concave and convex feature points is ;

[0105] The characteristic series of the strip feature points include: the average value of the sum of the normalized angles of the direction consistency corrected normal vectors in the annular neighborhood of the concave and convex feature points , the radius of the neighborhood with the highest normalized feature significance , linearity and directional consistency measures.

[0106] The characteristic sequence of the bar feature points is .

[0107] Step 8: For the candidate points in the source point cloud and the target point cloud, calculate the dot product of the feature sequence divided by the product of the modulus length to obtain the similarity. The similarity calculation can be obtained by the following formula:

[0108] ;

[0109] in, Represents the feature sequence corresponding to the candidate points in the source point cloud and the target point cloud, represents the dot product operation, Indicates the modulus length corresponding to the feature sequence. When it is greater than the threshold, it is represented as a pair of matching points. If the current point cloud has a similarity of more than the threshold number of groups that are greater than the threshold, the corresponding points with the highest similarity are taken to form a matching point pair, and all matching points are traversed and filtered out.

[0110] Step 9: Use adaptive weights to control the contribution of three types of matching points in the source point cloud and the target point cloud, namely, featureless point cloud points, concave-convex feature points, and strip feature points, to the point cloud registration residuals for registration.

[0111] The point cloud registration residual is calculated by combining the adaptive weights and three types of matching points in the source point cloud and the target point cloud: featureless point cloud points, concave-convex feature points, and strip feature points:

[0112] ;

[0113] Among them, R is the rotation matrix in the registered transformation relationship, and t is the translation matrix; Represents the featureless point cloud points in the source point cloud, represents the concave and convex feature points in the source point cloud, Represents the strip feature points in the source point cloud; Represents the paired featureless point cloud points in the target point cloud, represents the paired concave and convex feature points in the target point cloud, Represents paired bar feature points in the target point cloud;

[0114] Indicates the adaptive weights corresponding to the three types of point cloud points; adaptive weights It is based on the number of points in each type of point cloud The optimization is performed by calculating the inverse of , so that smaller categories have higher adaptive weights, preventing a certain type of feature points from dominating the optimization process. The contributions of different point clouds are balanced, especially when a certain type of point cloud has a large number of points, its adaptive weight is reduced to make the registration more balanced.

[0115] Be the minimum value and avoid division by 0;

[0116] The overall optimization goal is to minimize the weighted point cloud registration residual, and the Levenberg-Marquardt algorithm is used to update the transformation relationship. , repeat the iteration until Convergence, registration completed.

[0117] Example 2

[0118] See Figure 5 As shown, an embodiment of the present invention provides a multi-scale fusion point cloud feature extraction and adaptive matching device, including: at least one processing unit, the processing unit connected to a storage unit via a bus unit, the storage unit being a computer-readable storage medium that can be used to store software programs, computer executable programs, and modules, such as the software programs, computer executable programs, and modules corresponding to a multi-scale fusion point cloud feature extraction and adaptive matching method in an embodiment of the present invention. The processing unit implements the multi-scale fusion point cloud feature extraction and adaptive matching method by running the software programs, computer executable programs, and modules stored in the storage unit, including:

[0119] Collect the point cloud to be adaptively matched, and pre-process the point cloud to be adaptively matched by filtering and voxelization;

[0120] Determine the local neighborhood of each point in the point cloud and use the local neighborhood of the point cloud points to calculate the normal vector of each point;

[0121] Screen effective feature points based on the normal vectors of point cloud points;

[0122] Perform multi-scale feature calculation on the retained valid feature points and perform secondary feature point screening;

[0123] Classify the effective feature points screened for the second time into concave-convex feature points and strip feature points;

[0124] The strip feature points are further screened out according to the degradation direction;

[0125] Determine the corresponding feature series of the concave and convex feature points and the remaining strip feature points; the feature series of the concave and convex feature points includes: the average value of the sum of the normal vector angles corrected for direction consistency within the normalized annular neighborhood of the concave and convex feature points, the radius of the normalized neighborhood with the highest feature significance, and the curvature characteristics of the concave and convex feature points; the feature series of the strip feature points includes: the average value of the sum of the normal vector angles corrected for direction consistency within the normalized annular neighborhood of the concave and convex feature points, the radius of the normalized neighborhood with the highest feature significance, linearity, and directional consistency measurement;

[0126] For the candidate points in the source point cloud and the target point cloud, the similarity of the feature sequence is calculated, point pairs are matched based on the similarity, and all matching points are traversed and filtered out;

[0127] Adaptive weights are used to control the contribution of three types of matching points, namely featureless point cloud points, concave-convex feature points and strip feature points, in the source point cloud and target point cloud to the point cloud registration residual, and the point cloud registration residual is minimized for registration.

[0128] Of course, the computer program stored in the storage unit of the multi-scale fusion point cloud feature extraction and adaptive matching device provided in an embodiment of the present invention is not limited to the method operations described above, and can also perform related operations in the multi-scale fusion point cloud feature extraction and adaptive matching method provided in any embodiment of the present invention.

[0129] Example 3

[0130] An embodiment of the present invention provides a computer-readable storage medium storing a computer program. When the computer program is executed, the method for extracting and adaptively matching point cloud features using multi-scale fusion is implemented, including:

[0131] Collect the point cloud to be adaptively matched, and pre-process the point cloud to be adaptively matched by filtering and voxelization;

[0132] Determine the local neighborhood of each point in the point cloud and use the local neighborhood of the point cloud points to calculate the normal vector of each point;

[0133] Screen effective feature points based on the normal vectors of point cloud points;

[0134] Perform multi-scale feature calculation on the retained valid feature points and perform secondary feature point screening;

[0135] Classify the effective feature points screened for the second time into concave-convex feature points and strip feature points;

[0136] The strip feature points are further screened out according to the degradation direction;

[0137] Determine the corresponding feature series of the concave and convex feature points and the remaining strip feature points; the feature series of the concave and convex feature points includes: the average value of the sum of the normal vector angles corrected for direction consistency within the normalized annular neighborhood of the concave and convex feature points, the radius of the normalized neighborhood with the highest feature significance, and the curvature characteristics of the concave and convex feature points; the feature series of the strip feature points includes: the average value of the sum of the normal vector angles corrected for direction consistency within the normalized annular neighborhood of the concave and convex feature points, the radius of the normalized neighborhood with the highest feature significance, linearity, and directional consistency measurement;

[0138] For the candidate points in the source point cloud and the target point cloud, the similarity of the feature sequence is calculated, point pairs are matched based on the similarity, and all matching points are traversed and filtered out;

[0139] Adaptive weights are used to control the contribution of three types of matching points, namely featureless point cloud points, concave-convex feature points and strip feature points, in the source point cloud and target point cloud to the point cloud registration residual, and the point cloud registration residual is minimized for registration.

[0140] An embodiment of the present invention provides a computer-readable storage medium, in which the computer program stored is not limited to the method operations described above, but can also execute related operations in a multi-scale fusion point cloud feature extraction and adaptive matching method provided by any embodiment of the present invention.

[0141] In the embodiments provided by the present invention, it should be understood that the disclosed structures and methods can be implemented in other ways. For example, the structural embodiments described above are merely illustrative. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interface, structure or unit, which can be electrical, mechanical or other forms.

[0142] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0143] In addition, the functional units in the various embodiments of the present invention may be integrated into a single processing unit, each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0144] The foregoing description is intended only to provide specific embodiments of the present invention, which will enable those skilled in the art to understand and implement the present invention. 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 invention. Therefore, the present invention is not intended to be limited to the embodiments shown herein, but is intended to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A multi-scale fusion point cloud feature extraction and adaptive matching method, characterized in that: include: Collect the point cloud to be adaptively matched, and pre-process the point cloud to be adaptively matched by filtering and voxelization; Determine the local neighborhood of each point in the point cloud and use the local neighborhood of the point cloud points to calculate the normal vector of each point; Screen effective feature points based on the normal vectors of point cloud points; Perform multi-scale feature calculation on the retained valid feature points and perform secondary feature point screening; Classify the effective feature points screened for the second time into concave-convex feature points and strip feature points; The strip feature points are further screened out according to the degradation direction; Determine the corresponding feature sequence of the concave and convex feature points and the remaining strip feature points; The feature sequence of concave and convex feature points includes: the average value of the sum of the normal vector angles corrected for directional consistency within the normalized annular neighborhood of the concave and convex feature points, the normalized radius of the neighborhood with the highest feature significance, and the curvature characteristics of the concave and convex feature points; the feature sequence of strip feature points includes: the average value of the sum of the normal vector angles corrected for directional consistency within the normalized annular neighborhood of the concave and convex feature points, the normalized radius of the neighborhood with the highest feature significance, linearity, and directional consistency metrics; For the candidate points in the source point cloud and the target point cloud, the similarity of the feature sequence is calculated, point pairs are matched based on the similarity, and all matching points are traversed and filtered out; Adaptive weights are used to control the contribution of three types of matching points, namely featureless point cloud points, concave-convex feature points and strip feature points, in the source point cloud and target point cloud to the point cloud registration residual, and the point cloud registration residual is minimized for registration.

2. The multi-scale fusion point cloud feature extraction and adaptive matching method according to claim 1, characterized in that: The process of calculating the normal vector of each point using the local neighborhood of the point cloud point includes: Compute the covariance matrix of each point using its neighboring points in its local neighborhood: ； Compute the eigenvalues ​​and eigenvectors of the covariance matrix: ； in, Point cloud point The local neighborhood of The center of mass of the point cloud, the coordinates of the center of mass are in the local neighborhood The average of the coordinates of the midpoints, Point cloud point The covariance matrix of Represents the covariance matrix The eigenvalues ​​and eigenvectors of The eigenvector corresponding to the minimum eigenvalue is taken as the normal vector of the point cloud point.

3. The multi-scale fusion point cloud feature extraction and adaptive matching method according to claim 1, characterized in that: The screening of effective feature points according to the normal vectors of the point cloud points includes: obtaining the normal vectors of the point cloud points and the neighborhood points corresponding to each point cloud point; Perform directional consistency correction on the normal vectors of point cloud points; Calculate the average of the sum of the direction consistency corrected normal vectors of each point cloud point and the angles between the direction consistency corrected normal vectors of the neighboring points corresponding to each point cloud point: ； in, Point cloud point The local neighborhood of Point cloud points and Directional consistency correction normal vector; When the point cloud When the point cloud points are less than the set threshold, the point cloud points are filtered out. When the value is greater than the set threshold, the point cloud point is retained as a valid feature point.

4. The multi-scale fusion point cloud feature extraction and adaptive matching method according to claim 1, characterized in that: Perform multi-scale feature calculation on the retained valid feature points and perform secondary feature point screening, including: For each valid feature point, a multi-scale annular neighborhood is divided using circles of different radii with the valid feature point as the center. Calculate the average of the sum of the angles between the directional consistency-corrected normal vectors in different annular neighborhoods: ； Where x represents the index of the x-th ring neighborhood, Here it is represented by a dot The neighborhood corresponding to the xth ring centered at is, Point cloud points and Directional consistency correction normal vector; The average value of the normal vector angle is corrected by the direction consistency within each ring, and the neighborhood with the highest feature significance is selected; Traversing each valid feature point, there is a neighborhood with the highest feature significance of any valid feature point. When the neighborhood point is larger, the point cloud point is deleted, and the remaining valid feature points are the second screening valid feature points, and the second screening valid feature points are retained. and the radius of the neighborhood with the highest feature significance .

5. The multi-scale fusion point cloud feature extraction and adaptive matching method according to claim 1, characterized in that: According to linearity The effective feature points screened for the second time are classified into concave-convex feature points and strip feature points: the linearity is calculated as follows: ； in, is the maximum eigenvalue, The linearity is less than the maximum eigenvalue and not less than the minimum eigenvalue. When the linearity threshold is greater than the set value, the effective feature points are classified as bar feature points in the second screening. When the linearity is less than the set threshold, the effective feature points are classified as concave and convex feature points after the second screening.

6. The multi-scale fusion point cloud feature extraction and adaptive matching method according to claim 1, characterized in that: Methods for further filtering out strip feature points according to the degradation direction include: Calculate the directional consistency measure of the strip feature point direction and the degenerate direction: , in, Indicates the maximum eigenvalue of the bar feature point The corresponding eigenvector represents the direction of the bar feature point. represents the degenerate direction unit vector, Represents the dot product operation. When the direction consistency measure is less than the direction consistency measure threshold, it means that the strip feature point is basically consistent with the degradation direction, and the strip feature is a degradation-sensitive feature point and is screened out.

7. The multi-scale fusion point cloud feature extraction and adaptive matching method according to claim 1, characterized in that: The point cloud registration residual is calculated by combining the adaptive weights and three types of matching points in the source point cloud and the target point cloud: featureless point cloud points, concave-convex feature points, and strip feature points: ； Among them, R is the rotation matrix in the registered transformation relationship, and t is the translation matrix in the registered transformation relationship; Represents the featureless point cloud points in the source point cloud, represents the concave and convex feature points in the source point cloud, Represents the strip feature points in the source point cloud; Represents the paired featureless point cloud points in the target point cloud, represents the paired concave and convex feature points in the target point cloud, Represents paired bar feature points in the target point cloud; are the corresponding adaptive weights for featureless point cloud points, concave-convex feature points, and strip feature points respectively; To avoid dividing by zero; Update the transformation relationship using the Levenberg-Marquardt algorithm , repeat the iteration until Convergence, registration completed.

8. The multi-scale fusion point cloud feature extraction and adaptive matching method according to claim 7, characterized in that: Adaptive weights is calculated as follows: ； Adaptive weights It is based on the number of points in each type of point cloud The smaller the number of categories, the higher the adaptive weight will be, which can avoid one type of point cloud dominating the optimization process and balance the contribution of different point clouds.

9. A multi-scale fusion point cloud feature extraction and adaptive matching device, comprising: At least one processing unit, the processing unit being connected to a storage unit via a bus unit, characterized in that the storage unit stores a computer program, and the processing unit implements the multi-scale fusion point cloud feature extraction and adaptive matching method according to any one of claims 1 to 8 by running the computer program stored in the storage unit.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed, it implements the multi-scale fusion point cloud feature extraction and adaptive matching method according to any one of claims 1 to 8.

Images