A low overlap point cloud registration method combining dimensionality reduction projection and feature matching

By combining dimensionality reduction projection and image feature matching methods, the difficulty of feature point extraction and mismatch problems in low overlap rate point cloud registration are solved, and efficient point cloud registration is achieved, suitable for large-scale point cloud data.

CN114494368BActive Publication Date: 2024-11-08SOUTHEAST UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111558350.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-12-20
Publication Date
2024-11-08
Estimated Expiration
2041-12-20

AI Technical Summary

Technical Problem

Existing point cloud registration algorithms are difficult to effectively deal with low overlap rate point clouds, especially in the problems of high computational complexity, time-consuming and mismatch, and cannot be directly applied to low overlap rate point cloud registration.

Method used

Combining the dimensionality reduction projection and image feature matching methods, the feature points are extracted, described and matched in the image space after dimensionality reduction, and filtered through the voting mechanism and the distance invariance of rigid body transformation, and iterative optimization of the ICP algorithm to obtain an accurate transformation matrix .

Benefits of technology

It improves the accuracy and efficiency of low overlap rate point cloud registration, solves the problems of difficulty in extracting feature points, mismatch and iteration time-consuming, and is suitable for large-scale point cloud data.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114494368B_ABST
    Figure CN114494368B_ABST
Patent Text Reader

Abstract

A low overlap rate point cloud registration method combining dimensionality reduction projection and feature matching includes the following steps: 1) taking the mean normal vector direction of the point cloud as the initial dimensionality reduction direction; 2) performing direction correction on the initial dimensionality reduction direction to obtain the final dimensionality reduction direction; 3) performing projection along the final dimensionality reduction direction to generate a depth image after dimensionality reduction; 4) generating a multi-resolution image according to the grid size; 5) performing two-step screening according to the voting mechanism and the distance invariance of the rigid body transformation to obtain reliable matching point pairs; 6) performing two-step screening using the voting mechanism and the distance invariance of the rigid body transformation; 7) solving the rotation matrix for the final target point pair, and using the ICP algorithm for iterative optimization to obtain the final accurate transformation matrix. The method of the present invention solves the problems that the feature matching of low overlap rate point clouds is prone to errors and the iterative search time is long, and improves the accuracy and efficiency of registration.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of point cloud registration applications in three-dimensional vision, and in particular to a low-overlap point cloud registration method combining dimensionality reduction projection and feature matching. Background Art

[0002] The purpose of point cloud registration is to solve the transformation matrix of point clouds with different postures under the same coordinate system, and use the transformation matrix to unify the multi-view scanning point clouds under the same coordinate system, and finally obtain a complete three-dimensional model and scene. In recent years, point cloud registration has been widely studied, but the performance of traditional point cloud registration algorithms is limited by good initial positions and large overlap rates. Among them, the overlap rate is defined as the proportion of the number of points in the overlapping area in the complete point cloud. When the overlap rate between the source point cloud and the target point cloud is less than 60%, the degree of overlap between the two point clouds is low. Therefore, a fast and accurate registration algorithm for point clouds with low overlap rates is still a difficult point in point cloud registration.

[0003] There are many point cloud registration algorithms currently, which can be roughly divided into two categories: global search-based and local feature-based registration algorithms.

[0004] The registration algorithm based on global search is to find corresponding points from the point cloud to be matched according to the constraints of geometric position relationship, calculate the transformation matrix, and iterate multiple times to obtain the optimal transformation matrix. The classic registration algorithm based on global search is the Iterative Closest Point (ICP) algorithm, but this algorithm has high requirements on the initial position of the point cloud. According to the ICP algorithm, a variety of improved forms have been derived, such as PLICP, GICP, etc. Such improved algorithms have improved the accuracy, speed and noise resistance of point cloud registration to a certain extent, but they still cannot complete the registration of low overlap rate point clouds well; the registration method based on probability statistics is also based on the global unit, using the probability density function to estimate the point cloud distribution, and the representative one is the Normal distributions transform (NDT) algorithm proposed by Biber et al.; there are other registration algorithms based on global search, such as the algorithm based on random sample consensus (RANSAC), the algorithm based on four-points congruent sets (4PCS) proposed by Mellado et al. and its improved form D4PCS algorithm, Super4PCS algorithm, etc. However, the registration algorithms based on global search generally have the characteristics of high computational complexity and take a long time; at the same time, for the registration of low overlap rate point clouds, the search of non-overlapping areas does not work for registration, but increases the amount of calculation and reduces the overall registration efficiency.

[0005] The main work of the registration algorithm based on local features is to define feature descriptors and establish the corresponding relationship between the features of point clouds. Rusu et al. proposed the point feature histogram (PFH) that counts the angle between the normal vectors of two adjacent points as a descriptor, and its improved form, the fast point feature histogram (FPFH); Daniel F. Huber et al. proposed the spin image (SI) feature quantity that converts the point cloud into a grid form; Tombari et al. combined the ideas of point signature and point feature histogram and proposed the SHOT feature descriptor; Andrea et al. proposed the three-dimensional shape context (3DSC), which uses histograms to describe shape contour features in the logarithmic polar coordinate system, which is an extension of the two-dimensional shape context; Li et al. obtained a small-scale clustered point cloud with obvious features by clustering the target three-dimensional point cloud, which effectively improved the registration efficiency and accuracy; In addition, there are common local feature descriptors such as scale-invariant feature transform (SIFT) and rotation-invariant feature transform (RIFT). Since the local feature extraction algorithm of point cloud is highly complex and prone to problems such as mismatching, it cannot be directly used for the registration of low overlap point clouds. At the same time, local feature extraction is easily disturbed by clutter and occlusion, that is, when there is clutter and occlusion in the three-dimensional point cloud, the registration effect will be greatly reduced.

[0006] Traditional registration algorithms, whether based on global search or local features, cannot be directly applied to low-overlap point cloud registration. To solve this problem, many scholars have proposed a two-step registration strategy, which is to first extract the target points and overlapping areas, and then iteratively search within the overlapping area to obtain the optimal solution. Wang et al. proposed an overlapping region extraction method based on region segmentation, which can still complete the extraction of overlapping regions when the point cloud acquisition viewing angles are greatly different, thereby improving the registration efficiency; Li et al. proposed an improved TrICP algorithm based on contribution factors. The algorithm automatically calculates the degree of overlap and can reliably and efficiently automatically register laser and image reconstruction point clouds containing a lot of noise, partial overlap, and non-homologous sources; Lu et al. divided the point cloud into multiple regions, first performed D4PCS registration on each region, and then magnified the role of the overlapping region to complete the global registration, but the algorithm is time-consuming and not suitable for large-scale point cloud data; Zhang Yuan et al. combined regional blocking and convex optimization to extract overlapping regions and optimize the corresponding relationship respectively, and finally used the ICP algorithm to complete the registration; although the above algorithms have improved the accuracy of low overlap point cloud registration to a certain extent, they all calculate and extract feature points in three-dimensional space, and the algorithms have problems such as high computational complexity and long time consumption. Summary of the invention

[0007] In order to address the deficiencies in the prior art, the purpose of the present invention is to provide a low overlap point cloud registration method that combines dimensionality reduction projection and feature matching. Aiming at the problem of low overlap point cloud registration, by combining dimensionality reduction projection with image feature matching, feature point extraction, description and matching are completed in the reduced dimensionality image space, thereby solving the problems of difficulty in extracting feature points of low overlap point clouds, easy mismatching, and long iteration time.

[0008] To achieve the above object, the present invention provides a low overlap point cloud registration method combining dimensionality reduction projection and feature matching, comprising the following steps:

[0009] 1) The mean normal vector direction of the point cloud is used as the initial dimension reduction direction;

[0010] 2) performing direction correction on the initial dimension reduction direction to obtain a final dimension reduction direction;

[0011] 3) Projecting along the final dimension reduction direction to generate a reduced-dimensional depth image;

[0012] 4) Generate multi-resolution images according to the grid size;

[0013] 5) Perform two-step screening based on the voting mechanism and the distance invariance of rigid body transformation to obtain reliable matching point pairs;

[0014] 6) Two-step screening using voting mechanism and distance invariance of rigid body transformation;

[0015] 7) Solve the rotation matrix for the final target point pair, use the ICP algorithm to iteratively optimize, and obtain the final accurate transformation matrix.

[0016] Furthermore, the step 1) further comprises:

[0017] The normal vector of the discrete point is solved by principal component analysis, p i The normal vector is the eigenvector corresponding to the minimum eigenvalue of the matrix:

[0018]

[0019] Eξ i =λ i ξ i , i = {1, 2, 3},

[0020] Where R is the neighborhood radius, N is the number of points in the area, and p i represents the three-dimensional coordinate vector of the i-th point in the domain, p′ represents the three-dimensional coordinate vector of the center of mass of the domain, d i represents the Euclidean distance from the i-th point to the centroid in the domain, ξ i is the eigenvector of matrix E, i is an integer greater than or equal to 1;

[0021] p i The point normal vector takes the eigenvector corresponding to the minimum eigenvalue of the matrix, denoted by n = (x i ,y i , z i ), then the initial projection direction of dimensionality reduction is determined as the mean normal vector of each point in the point cloud P:

[0022]

[0023] Furthermore, the step 2) further comprises:

[0024] The spatial neighborhood is evenly divided along the initial projection direction, and projected along multiple neighborhood directions to obtain different dimensionality reduction images;

[0025] The hash algorithm is used to obtain the hash fingerprint of the reduced-dimensional image, and the Hamming distance is compared. The image with the smallest distance value is regarded as the image with the highest similarity, and the corresponding direction is regarded as the corrected direction.

[0026] Rotate the original point cloud so that the Z-axis direction of the rotated point cloud space coordinate system is aligned with the projection direction:

[0027]

[0028] in,

[0029]

[0030] Furthermore, the step 3) also includes,

[0031] Calculate the maximum depth value max within each grid c D (c), and then define the point set P′(c) as the outermost point set in the grid:

[0032] P′(c)={p∈P||max D (c)-p z |<δ1};

[0033] Perform Gaussian weight interpolation on the Z coordinate of the midpoint of P′(c) to calculate the pixel value f(c) corresponding to the grid:

[0034]

[0035] Among them, the normalization parameter W is expressed as:

[0036]

[0037] The Gaussian function g is expressed as:

[0038]

[0039] Interpolation is performed to calculate the pixel value corresponding to each grid to obtain a depth image after dimensionality reduction.

[0040] Furthermore, the step 4) also includes,

[0041] Establish the relationship between the grid size d and the mean spacing D of the discrete points in the point cloud:

[0042] d = l·D,

[0043] Among them, l is the coefficient;

[0044] The resolution is expressed as:

[0045]

[0046] Furthermore, the step 5) also includes,

[0047] The ORB operator is used to perform ORB matching on n different layers. The image feature point pairs corresponding to each layer are expressed as:

[0048] matches[i], i∈[0,n-1],

[0049] Among them, matches[0] corresponds to the layer with the highest resolution under l0;

[0050] Count the number of times the matching relationship in matches[0] is repeated in other layers. The vote count increases by one when the following equation holds:

[0051] l0×(matches[0][j].X, matches[0][j].Y)

[0052] = l i ×(matches[i][k].X, matches[i][k].Y)

[0053] Among them, matches[0][j] represents the j-th pair of matching points in the first layer, and matches[i][k] represents the k-th pair of matching points in the i+1th layer.

[0054] Furthermore, the step 6) also includes,

[0055] The vote threshold δ2 is determined according to the number of layers. Matching points greater than δ2 are considered reliable matching points and mapped to the original point cloud to obtain Matches.

[0056] Matches any two points (p i ,q i ) and (p j ,q j)If it is a correct pair of matching points, then according to the distance invariance of rigid body transformation, there is a relationship dist(p i , p j ) = dist(q i ,q j ); choose any two pairs of points, select δ3>0, so that:

[0057]

[0058] Among them, the numerator represents the difference in distance between matching points, and the denominator is to reduce the sensitivity of the distance constraint to the point cloud scale. The final matching point pair that satisfies the distance constraint is determined when the above formula is satisfied.

[0059] Furthermore, the step 7) further includes,

[0060] Using the quaternion method, the initial transformation matrix of the final matching point pair is obtained, and the point cloud to be registered is transformed to a similar position;

[0061] The ICP algorithm is applied to perform iterative calculations to obtain the precise transformation matrix.

[0062] To achieve the above-mentioned purpose, the present invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program running on the processor, and when the processor runs the computer program, the steps of the low overlap point cloud registration method combining dimensionality reduction projection and feature matching as described above are executed.

[0063] To achieve the above objectives, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when run, executes the steps of the low overlap point cloud registration method combining dimensionality reduction projection and feature matching as described above.

[0064] Compared with the prior art, the low overlap point cloud registration method combining dimensionality reduction projection and feature matching of the present invention has the following beneficial effects: for the problem of low overlap point cloud registration, by combining dimensionality reduction projection with image feature matching, feature point extraction, description and matching are completed in the image space after dimensionality reduction, which not only solves the problems of difficulty in extracting feature points of low overlap point clouds, easy mismatching, and long iteration time, but also improves the accuracy and efficiency of registration and adaptability to low overlap rates; as the scale of the point cloud increases, the speed advantage becomes more significant.

[0065] Other features and advantages of the present invention will be set forth in the description which follows, and in part will be apparent from the description, or may be learned by practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0067] Figure 1 A flow chart of a method for low overlap point cloud registration combining dimensionality reduction projection and feature matching according to the present invention;

[0068] Figure 2 A schematic diagram of point cloud dimensionality reduction according to the present invention;

[0069] Figure 3 A schematic diagram of direction correction according to the present invention;

[0070] Figure 4 Schematic diagram of image effects after dimension reduction with different coefficients according to the present invention;

[0071] Figure 5 A flow chart of feature matching and position calculation according to the present invention;

[0072] Figure 6 This is a diagram of a direction correction experiment according to the present invention;

[0073] Figure 7 It is a dimensionality reduction image hash fingerprint according to the present invention;

[0074] Figure 8 A feature matching and screening result diagram according to the present invention;

[0075] Fig. 9 A schematic diagram comparing the registration results obtained according to the present invention and other registration methods;

[0076] Fig.10 Schematic diagram of error curve and speed curve according to the present invention. DETAILED DESCRIPTION

[0077] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0078] Example 1

[0079] Figure 1 The following is a flow chart of a method for low overlap point cloud registration combining dimensionality reduction projection and feature matching according to the present invention. Figure 1 , the low overlap point cloud registration method combining dimensionality reduction projection and feature matching of the present invention is described in detail.

[0080] First, in step 101, the normal vector information of the point cloud is obtained according to the principal component analysis method, and the mean normal vector is used as the initial projection direction for dimensionality reduction.

[0081] In the embodiment of the present invention, the normal vector of the discrete point is solved by the principal component analysis method, p i The normal vector is the eigenvector corresponding to the minimum eigenvalue of the matrix:

[0082]

[0083] Eξ i =λ i ξ i , i = {1, 2, 3},

[0084] Among them, R is the neighborhood radius, N is the number of points in the area, and p i represents the three-dimensional coordinate vector of the i-th point in the domain, p′ represents the three-dimensional coordinate vector of the center of mass of the domain, d i represents the Euclidean distance from the i-th point to the centroid in the domain, ξ i is the eigenvector of matrix E, i is an integer greater than or equal to 1;

[0085] p i The point normal vector takes the eigenvector corresponding to the minimum eigenvalue of the matrix, denoted by n = (x i ,y i , z i ), then the initial projection direction of dimensionality reduction is determined as the mean normal vector of each point in the point cloud P:

[0086]

[0087] In step 102, the direction is corrected by dividing the initial projection direction into neighborhoods and using a hash algorithm.

[0088] In the embodiment of the present invention, in order to reduce the dimension of the point cloud overlap area along the approximate direction to increase the image feature matching points, the initial projection direction obtained in step 101 is corrected:

[0089] 1) After determining the initial projection direction, the spatial neighborhood is evenly divided along the direction, and projected along multiple neighborhood directions to obtain different dimensionality reduction images;

[0090] 2) Using the hash algorithm to obtain the hash fingerprint of the above image, and compare its Hamming distance, the image with the smallest distance value is regarded as the image with the highest similarity, and the direction corresponding to this image is used as the corrected direction;

[0091] 3) Rotate the original point cloud so that the Z-axis direction of the rotated point cloud space coordinate system is aligned with the projection direction:

[0092]

[0093] in,

[0094]

[0095] In step 103, the discrete points in the projection grid are interpolated by a Gaussian function to obtain pixel values ​​corresponding to the grid.

[0096] In the embodiment of the present invention, since the pixel grid may contain multiple discrete points, taking the depth of one point as the pixel value cannot accurately represent the information of all points in the area, so it is considered to use the Gaussian weight interpolation method to obtain the pixel value.

[0097] 1) First calculate the maximum depth value max within each grid c D (c), and then define the point set P′(c) as the outermost point set in the grid:

[0098] P′(c)={p∈P||max D (c)-p z |<δ1}

[0099] 2) Perform Gaussian weight interpolation on the Z coordinate of the midpoint of P′(c) to calculate the pixel value f(c) corresponding to the grid:

[0100]

[0101] Among them, the normalization parameter W is expressed as:

[0102]

[0103] The Gaussian function g is expressed as:

[0104]

[0105] Interpolation is performed to calculate the pixel value corresponding to each grid to obtain a depth image after dimensionality reduction.

[0106] Figure 2 is a schematic diagram of point cloud dimensionality reduction according to the present invention, as shown in Figure 2 As shown in the figure, firstly, the projection direction is determined by the normal vector information and direction correction of the point cloud, and the original point cloud data is rotated and translated to align the mean normal vector with the z-axis direction; then the point cloud data is projected onto the xoy plane, and the resolution is determined according to the mean spacing of the point cloud and rasterized; finally, the Z coordinate in each grid is interpolated to obtain the pixel value, and the three-dimensional point cloud is converted into a two-dimensional depth map.

[0107] Figure 3 It is a schematic diagram of direction correction according to the present invention, as shown in Figure 3As shown, for partially overlapping point clouds, the correct matching relationships mainly concentrate in the overlapping area. For the image space after dimensionality reduction of such point clouds, the prerequisite for obtaining these correct matching relationships is that the overlapping area is projected along approximately the same direction. In the dimensionality reduction process of the present invention, this direction is obtained through direction correction.

[0108] The neighborhood space of the initial direction of the target point cloud is evenly divided into k parts, and a direction with an angle θ (which can be 30° or 45°) to the initial direction is taken as a candidate.

[0109] As Figure 3 shown, n1, n2,..., n k are normal vector uniformly distributed in the neighborhood and with an angle θ to . Taking n k as an example, assuming its Z-axis component is Δz, then Then the conversion relationship between n k and can be solved

[0110] Figure 4 is a schematic diagram of the image effect after dimensionality reduction according to different coefficients l of the present invention. As Figure 4 shown, the image effects obtained by dimensionality reduction with different coefficients l, Figure 4 in (a): l = 1; Figure 4 in (b): l = 1.5; Figure 4 in (c): l = 2; Figure 4 in (d) l = 3; It can be seen from Figure 4 that when l ≤ 1.2, some points (abnormal points) with pixel value 0 are unevenly distributed in the image. This is because the grid is small and no projection points fall into it, and in this case, more false feature points will be added in feature extraction; when l ≥ 3, obvious serrations appear at the edge of the image, and at the same time, fine features are lost. This is because the grid is large and the grid contains more points, resulting in detail loss after interpolation, and the feature points that can be extracted will be greatly reduced. When 1.2 < l < 3, the image resolution is more appropriate. The dimensionality reduction process can not only reduce the loss of point information but also accurately express the point cloud information. In summary, multiple groups of l are selected in the interval (1.2, 3) to generate multi-resolution dimensionality reduction images for further feature matching.

[0111] In step 104, grids of different sizes are selected according to the mean spacing of the point cloud to generate dimensionality reduction images with different resolutions.

[0112] In the embodiment of the present invention, point clouds acquired by different devices have different point set sizes and resolutions. In order to make the registration method adaptive to the point cloud scale, the relationship between the grid size d and the mean spacing D of discrete points of the point cloud is established in the dimensionality reduction process:

[0113] d=l·D

[0114] The grid size d determines the number of discrete points in the grid, which also determines the resolution of the image after dimensionality reduction. The resolution can be expressed as:

[0115]

[0116] In step 105, the layers are divided according to the resolution, and ORB feature matching is performed between different layers respectively, and the number of votes for repeated matching relationships is counted.

[0117] In the embodiment of the present invention, after the multi-resolution image is generated by step 104, in the image feature matching, the present invention uses the ORB operator to perform ORB matching on n different layers, and the image feature point pairs corresponding to each layer are represented as matches[i], i∈[0, n-1], where matches[0] corresponds to the layer with the highest resolution under l0. The number of repetitions of the matching relationship in matches[0] in other layers is counted, and the number of votes is increased by one when the following formula is true:

[0118] l0×(matches[0][j].X, matches[0][j].Y)

[0119] = l i ×(matches[i][k].X, matches[i][k].Y)

[0120] Where matches[0][j] represents the j-th pair of matching points in the first layer, and matches[i][k] represents the k-th pair of matching points in the i+1th layer.

[0121] In step 106, two screenings are performed based on the voting mechanism and the distance invariance of the rigid body transformation to obtain a reliable matching relationship.

[0122] In the embodiment of the present invention, the vote threshold δ2 is determined according to the number of layers, and matching points greater than δ2 are regarded as reliable matching points and mapped to the original point cloud to obtain Matches. i ,q i ) and (p j ,q j )If it is a correct pair of matching points, then according to the distance invariance of rigid body transformation, there is a relationship dist(p i , p j ) = dist(qi ,q j ), so choose any two pairs of points and select appropriate δ3>0 so that:

[0123]

[0124] Among them, the numerator represents the difference in distance between matching points, and the denominator is to reduce the sensitivity of the distance constraint to the point cloud scale. The final matching point pair that satisfies the distance constraint is determined when the above formula is satisfied.

[0125] In step 107, the quaternion method is applied to the final matching point to obtain its position conversion matrix.

[0126] In the embodiment of the present invention, the final matching point pair is obtained through step 106, and the initial transformation matrix R, T is obtained through the quaternion method, and the point cloud to be registered is transformed to a similar position based on this.

[0127] In step 108, the ICP algorithm is used to iteratively calculate and obtain an accurate solution for the registration.

[0128] In the embodiment of the present invention, after the quaternion method is used for the initial registration, after the registered point cloud is converted to a similar position, the ICP algorithm is applied to the registered point cloud, and an accurate transformation matrix can be obtained with only a small number of iterative calculations.

[0129] Figure 5 For the feature matching and position solving flow chart according to the present invention, the following will refer to Figure 5 , the feature matching and position solution process of the present invention is further described.

[0130] In step 501, feature matching is performed on the two-dimensional reduced-dimensional image: the reduced-dimensional image of the point cloud to be registered is regarded as a layer at the same coefficient, and image feature matching is performed on different layers respectively;

[0131] In step 502, feature point screening is performed to obtain three-dimensional matching point pairs: the number of times the same matching point pair appears repeatedly in different layers is counted, and matching point pairs with votes less than a threshold are considered unreliable matches and are removed; the retained matching point pairs are mapped back to the point cloud to obtain three-dimensional matching point pairs, and the final target point pairs are screened according to the distance invariance of the rigid body transformation;

[0132] In step 503, the transformation matrix is ​​calculated: the positional relationship between the target point pairs is solved using the quaternion method, thereby completing the initial registration of the point cloud; finally, the ICP is used to refine the registration to obtain an accurate result.

[0133] Figure 6 This is a diagram of the direction correction experiment according to the present invention, as shown in FIG. Figure 6 As shown, Figure 6 (a) is the candidate direction; Figure 6(b) is the directional dimension reduction image; Figure 6 (c) is the directional dimension reduction image; Figure 6 (d) is a directional dimension reduction image. The initial direction angles of the point clouds to be registered are quite different. The images obtained by dimension reduction along the initial directions are not suitable for feature matching and are prone to many incorrect matching relationships. Therefore, direction correction is required.

[0134] Figure 7 is the dimensionality reduction image hash fingerprint according to the present invention, such as Figure 7 As shown, pHash algorithm is used to solve q0 to q k Similarity with p. Figure 7 (a)-(d) in the figure correspond to the hash fingerprints of p, q0, q1, and q2, where the Hamming distance is calculated to obtain Figure 7 (a) and (d) have the highest similarity, so q2 corresponding to (d) is taken as the reduced-dimensional image after correction in the Q direction.

[0135] Figure 8 The feature matching and screening result diagram according to the present invention is shown as follows: Figure 8 As shown, after determining the dimensionality reduction direction, the value is changed to generate a multi-resolution image, and the ORB algorithm is used to match feature points in each layer, and the feature point pairs with lower votes are screened out to obtain the image matching result; then the matching feature points are mapped to the point cloud, and a second screening is performed according to the distance invariance of the rigid body transformation, and the matching result is shown in the figure below.

[0136] Table 1 shows the time consumption of each step of the registration method. The time consumption of this registration algorithm is relatively high in feature matching and direction correction. This is because feature matching is done through multiple matches on different layers, and direction correction includes multiple dimensionality reduction operations of point clouds, which is a separate operation.

[0137] Table 1 Time consumption of each step of the registration method

[0138] step Time consumed (ms) percentage(%) Dimensionality reduction transformation 379 14.87 Direction Correction 726 28.49 Feature Matching 941 36.93 Quaternion solution 95 3.72 ICP Iteration 407 15.97 total 2548 100

[0139] Fig. 9 FIG. 4 is a schematic diagram comparing the registration results obtained according to the present invention and other registration methods. Fig. 9 As shown, combined Fig. 9 Compared with Table 2, it can be seen that directly applying the ICP algorithm to the point cloud with low overlap rate will lead to registration failure; the Super4PCS algorithm is not suitable for the registration of point clouds with low overlap rate and large scale; and the method of the present invention can obtain better registration effect for point clouds with overlap rate as low as 30%.

[0140] Table 2 Comparison of registration results

[0141]

[0142] Fig.10 is a schematic diagram of an error curve and a speed curve according to the present invention, as shown in Fig.10 As shown, the low overlap point cloud registration method combining dimensionality reduction projection and feature matching of the present invention is suitable for low overlap point cloud registration. Compared with other listed algorithms, the present invention has advantages in both speed and accuracy, and the speed advantage becomes more significant as the point cloud scale increases.

[0143] In one embodiment of the present invention, there is also provided an electronic device, comprising a memory and a processor, wherein the memory stores a computer program running on the processor, and when the processor runs the computer program, the steps of the memory management method applicable to the back-end design of large-scale integrated circuits as described above are executed.

[0144] In one embodiment of the present invention, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed, the steps of the memory management method applicable to the back-end design of large-scale integrated circuits as described above are executed.

[0145] The memory management method applicable to the back-end design of large-scale integrated circuits of the present invention adopts virtual memory technology, memory binary storage technology and dirty page marking technology, uniformly realizes an efficient memory management method applicable to the back-end design of large-scale integrated circuits, can quickly store and load data, and utilizes memory binary direct storage combined with dirty page marking technology to naturally realize dirty page preservation technology to adapt to the future technical development trend of distributed cloudification; at the same time, by utilizing the Swap mechanism implied by the virtual memory file of the Linux system, cold data can be placed on the hard disk to break through the limitation of the actual physical memory capacity.

[0146] Those skilled in the art can understand that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention is described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions recorded in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A low overlap point cloud registration method combining dimensionality reduction projection and feature matching, comprising the following steps: Step 1) taking the mean normal vector direction of the point cloud as the initial dimension reduction direction; Step 2) performing direction correction on the initial dimension reduction direction to obtain a final dimension reduction direction; Step 3) Projecting along the final dimension reduction direction to generate a reduced-dimensional depth image; Step 4) generating a multi-resolution image according to the grid size; Step 5) Divide the layers according to the resolution, perform ORB feature matching between different layers, and count the number of votes for repeated matching relationships; Step 6) Perform two-step screening based on the voting mechanism and the distance invariance of rigid body transformation to obtain reliable matching point pairs; Step 7) Solve the rotation matrix for the final target point pair, use the ICP algorithm to iteratively optimize, and obtain the final accurate transformation matrix; The step 5) further comprises: The ORB operator is used to perform ORB matching on n different layers. The image feature point pairs corresponding to each layer are expressed as: matches[i], i∈[0,n-1], in, matches[0] corresponds to l 0 Select the layer with the highest resolution. Count the number of times the matching relationship in matches[0] is repeated in other layers. The vote count increases by one when the following equation holds: L 0 ×(matches[0][j]X,matches[0][j].Y) =l i ×(matches[i][k].X,matches[i][k].Y) Among them, matches[0][j] represents the j-th pair of matching points in the first layer, and matches[i][k] represents the k-th pair of matching points in the i+1th layer; The step 6) further comprises: Determine the vote threshold δ based on the number of layers 2 , greater than δ 2 The matching points are regarded as reliable matching points and mapped to the original point cloud to obtain Matches; Matches any two points (p i ,q i ) and (p j ,q j )If it is a correct pair of matching points, then according to the distance invariance of rigid body transformation, there is a relationship dist(p i , p j ) = dist(q i ,q j ); choose any two pairs of points and select δ 3 >0, so that: Among them, the numerator represents the difference in distance between matching points, and the denominator is to reduce the sensitivity of the distance constraint to the point cloud scale. The final matching point pair that satisfies the distance constraint is determined when the above formula is satisfied.

2. According to claim 1, the low overlap rate point cloud registration method combining dimensionality reduction projection and feature matching, It is characterized in that The step 1) further comprises: The normal vector of the discrete point is solved by principal component analysis, p i The normal vector is the eigenvector corresponding to the minimum eigenvalue of the matrix: Ex i =λ i x i ,i={1,2,3}, Where R is the neighborhood radius, N is the number of points in the area, and p i represents the three-dimensional coordinate vector of the i-th point in the domain, p′ represents the three-dimensional coordinate vector of the center of mass of the domain, d i represents the Euclidean distance from the i-th point to the centroid in the domain, ξ i is the eigenvector of matrix E, i is an integer greater than or equal to 1; p i The point normal vector takes the eigenvector corresponding to the minimum eigenvalue of the matrix, denoted by n = (x i ,y i , z i ), then the initial projection direction of dimensionality reduction is determined as the mean normal vector of each point in the point cloud P:

3. The low overlap point cloud registration method combining dimensionality reduction projection and feature matching according to claim 1, It is characterized in that The step 2) further comprises: The spatial neighborhood is evenly divided along the initial projection direction, and projected along multiple neighborhood directions to obtain different dimensionality reduction images; The hash algorithm is used to obtain the hash fingerprint of the reduced-dimensional image, and the Hamming distance is compared. The image with the smallest distance value is regarded as the image with the highest similarity, and the corresponding direction is regarded as the corrected direction. Rotate the original point cloud so that the Z-axis direction of the rotated point cloud space coordinate system is aligned with the projection direction: in, 4. The low overlap point cloud registration method combining dimensionality reduction projection and feature matching according to claim 1, It is characterized in that The step 3) also includes, Calculate the maximum depth value max within each grid c D (c), and then define the point set P′(c) as the outermost point set in the grid: P′(c)={p∈P||max D (c)-p z |<d 1 }; Perform Gaussian weight interpolation on the Z coordinate of the midpoint of P′(c) to calculate the pixel value f(c) corresponding to the grid: Among them, the normalization parameter W is expressed as: The Gaussian function g is expressed as: Interpolation is performed to calculate the pixel value corresponding to each grid to obtain a depth image after dimensionality reduction.

5. The low overlap point cloud registration method combining dimensionality reduction projection and feature matching according to claim 1, It is characterized in that The step 4) further comprises: Establish the relationship between the grid size d and the mean spacing D of the discrete points in the point cloud: d = l·D, Among them, l is the coefficient; The resolution is expressed as:

6. The low overlap rate point cloud registration method combining dimensionality reduction projection and feature matching according to claim 1, It is characterized in that The step 7) further comprises: Using the quaternion method, the initial transformation matrix of the final matching point pair is obtained, and the point cloud to be registered is transformed to a similar position; The ICP algorithm is applied to perform iterative calculations to obtain the precise transformation matrix.

7. An electronic device, It is characterized in that It comprises a memory and a processor, wherein the memory stores a computer program running on the processor, and the processor executes the steps of the low overlap rate point cloud registration method combining dimensionality reduction projection and feature matching as described in any one of claims 1 to 6 when running the computer program.

8. A computer-readable storage medium having a computer program stored thereon, It is characterized in that When the computer program is running, the steps of the low overlap point cloud registration method combining dimensionality reduction projection and feature matching described in any one of claims 1 to 6 are executed.