A registration method for low overlap rate point clouds

Abstract

This invention proposes a registration method for point clouds with low overlap rates, effectively improving upon traditional point cloud registration methods which suffer from slow speed, low accuracy, and poor stability. The method is based on variable data sequence length least truncated squares (LTS). For the complete sequence obtained by ascending ordering the Euclidean distances of matching point pairs, since pose changes after each iteration cause variations in the distance sequence of the matching point pairs, in each iteration, using the right endpoint of the complete sequence as a reference, a dynamically changing local sequence is selected from the complete sequence based on the point cloud overlap rate and the current iteration number. This changes the position of the LTS truncated sequence on the complete sequence. When the iteration termination condition is met, the maximum length of the local sequence is the same as the length of the complete sequence. As the iteration progresses, the length of the local sequence continuously increases, thus implementing a variable data sequence length strategy during registration. Compared with traditional point cloud registration methods, this invention offers better speed, accuracy, and stability, especially when registering point clouds with low overlap rates, where the performance improvement is particularly significant.

Description

Technical Field

[0001] This invention relates to the field of point cloud registration technology, specifically a registration method for point clouds with low overlap rates. Background Technology

[0002] Point cloud registration is a key technology in fields such as 3D reconstruction, medical image analysis, autonomous driving, and intelligent manufacturing.

[0003] Point cloud registration refers to the process of transforming the coordinate systems of multiple point cloud datasets to align them within the same coordinate system. From an implementation perspective, common point cloud registration methods can be broadly categorized into registration methods based on local features, registration methods based on probability statistics, and registration methods based on global search strategies. Among these, registration methods based on global search strategies achieve better registration results and are therefore more widely used.

[0004] The main idea behind global search-based registration methods is to find the optimal solution by exploring possible registration transformations in the global search space. Interpolation-based registration (ICP) is currently the most commonly used registration method based on a global search strategy, and it has high registration accuracy. However, ICP registration requires a good initial pose; otherwise, it is prone to getting trapped in local optima. Moreover, ICP does not perform well in registering point clouds with low overlap rates.

[0005] The TrICP method introduces Least Truncation Squares (LTS). By sorting the Euclidean distances of matched point pairs in ascending order, LTS is used to truncate a certain length of the matched point pair sequence for pose transformation estimation. This effectively reduces the adverse effects of low overlap and noisy points, significantly improving the shortcomings of ICP, especially for registration of low overlap point clouds, where it exhibits superior performance. However, because TrICP truncates a fixed length of matched point pair sequence at a fixed position using LTS, it ignores the role of ideal matched point pairs with larger distances in pose transformation estimation when the initial pose deviation of low overlap point clouds is large. This leads to slow convergence speed, low registration accuracy, and poor stability during the registration process. Summary of the Invention

[0006] To address the shortcomings of existing point cloud registration methods, such as slow speed, low accuracy, and poor stability when registering point clouds with low overlap rates, this invention proposes a registration method for point clouds with low overlap rates, which can effectively improve the registration effect of point clouds with low overlap rates.

[0007] Analysis of the point cloud registration process based on point-to-point mapping reveals that when the point cloud pose deviation is large, matching point pairs with larger distances have a more significant convergence effect, significantly improving convergence speed and accuracy. Conversely, when the point cloud pose deviation is small, matching point pairs with smaller distances contain more ideal matching point pairs, thus contributing more to improving convergence accuracy. However, in LTS applications, matching point pairs with larger distances are always located to the right of the matching point pair distance sequence. Therefore, to improve convergence speed and accuracy, this invention employs a variable data sequence length strategy during the registration iteration process, still using the LTS data truncation method to process each obtained local data sequence. This ensures that the local sequence processed by LTS contains as many ideal matching point pairs as possible. Especially when the point cloud pose deviation is large, ideal matching point pairs with larger distances facilitate rapid convergence, thereby improving the overall efficiency of registration.

[0008] To achieve the above technical objectives, the technical solution of the present invention is as follows:

[0009] A registration method for point clouds with low overlap rate, the method comprising the following steps:

[0010] S1. Input the source point cloud and the target point cloud;

[0011] S2. Determine matching point pairs based on the nearest neighbor principle;

[0012] S3. Sort the matching point pairs in ascending order of Euclidean distance to obtain the complete matching point pair distance sequence L (complete sequence);

[0013] S4. Calculate the local matching point pair distance sequence L1 (local sequence) in the complete sequence L based on the point cloud overlap rate and the current iteration number;

[0014] S5. Apply least truncation squares (LTS) to the current local sequence L1 to obtain the truncation matching point pair distance sequence L2 (the truncation sequence);

[0015] S6. Calculate the pose transformation matrix from the source point cloud to the target point cloud using the current truncated sequence L2;

[0016] S7. Perform pose transformation on the source point cloud using the pose transformation matrix;

[0017] S8. Algorithm convergence judgment.

[0018] Preferably, in step S4, the length l of the complete sequence depends on the source point cloud size, and the length l2 of the truncated sequence depends on the overlap rate ξ, both remaining constant during the registration process. The right endpoint L1 of the local sequence... right Always with the right endpoint L of the complete sequence right The local sequence overlaps, and the initial length l1 of the local sequence is the same as the length l2 of the LTS truncated sequence. In each iteration, the left endpoint L1 of the local sequence...left The system moves leftward on the complete sequence according to a dynamic step size α. When the iteration termination condition is met, the maximum value that the local sequence length l1 can achieve is the same as the complete sequence length l. As the iteration progresses, the local sequence length l1 continuously increases, thus implementing a variable data sequence length strategy during registration. α depends on the point cloud overlap rate and the current iteration number. α and L1... right and L1 left The calculation formula is as follows:

[0019]

[0020] L1 right =L right

[0021] L1 left =L1′ left -(α×l)

[0022] Where n is the current iteration number, ξ is the point cloud overlap rate, and L1′ left L1 of the previous iteration left .

[0023] Preferably, in step S5, at each iteration, LTS uses the local sequence L1 obtained in step S4 as the processing sequence, and truncates the interval [N] on the local sequence L1. start N end The matching point pairs within the frame constitute a truncation sequence L2, with a length of l2, used for pose transformation estimation. The position of the truncation sequence on the complete sequence changes with the length of the local sequence. The length of the truncation sequence l2 depends on the overlap rate ξ and remains constant during registration. The truncation interval can be calculated by the following formula:

[0024] N start =N′ start -(α×l)

[0025] N end =N start +l2

[0026] In the interval [N] start N end The sum of squared distances between point pairs can be obtained using the least squares method, as expressed by the following formula:

[0027]

[0028] Among them, D i N′ represents the Euclidean distance between the matched point pairs. start N represents the previous iteration start When N start When = 0, the local sequence L1 coincides with the complete sequence L.

[0029] Preferably, assuming an active point cloud P, a target point cloud Q, and an overlap rate ξ, the flow of the method of the present invention is as follows:

[0030] ① Let the interval [N] be the cutoff value. start N end N end =l,N start = (1-ξ)×l2, l2=ξ×l;

[0031] ② For all points P in the source point cloud P i Find its nearest neighbor Q in the target point cloud Q. i Calculate the square of its Euclidean distance. And for all Sort in ascending order;

[0032] ③ Calculate the cutoff interval [N] start N end S is the sum of squared distances between matching pairs within the same region. TS ,have

[0033]

[0034] If the iteration stopping condition is met, then exit; otherwise, let S′ TS =S TS Meanwhile, let N′ start =N start The left endpoint of the extracted interval is retained.

[0035] ④ Apply the singular value decomposition (SVD) method to the truncated sequence obtained in step ③ and solve for the pose transformation matrix (R, T) according to the following formula;

[0036]

[0037] ⑤ Use the pose transformation matrix (R, T) obtained in step ④ to perform spatial pose transformation on the source point cloud P;

[0038] P = RP + T

[0039] ⑥ Calculate [N] start N end Change the local sequence length l1, and then proceed to step ②:

[0040]

[0041] N start =N′ start -(α×l)

[0042] N end =N start +l2

[0043] Where n is the current iteration number, N′ start N represents the previous iteration start During the iteration process, if N start When <0, take N. start =0.

[0044] In step ③, the iteration terminates if any of the following conditions are met:

[0045] ① The number of iterations is greater than the set threshold N iter ;

[0046] ② The mean square error (MSE) value e of the point-to-point distance is less than a given threshold, where: e = S TS / l2;

[0047] ③ The MSE of two adjacent iterations |e i+1 -e i | less than a given threshold, where i represents the iteration order.

[0048] Compared with the prior art, the advantages of the method of the present invention are as follows:

[0049] This invention proposes a registration method for point clouds with low overlap rates. It utilizes LTS to sort matching point pairs in ascending order based on their Euclidean distance. In each iteration, the position of the LTS truncation sequence on the complete sequence is changed according to the point cloud overlap rate and the current iteration number. This allows for the maximum utilization of ideal matching point pairs for pose transformation estimation throughout the registration process, effectively improving the accuracy, speed, and stability of point cloud registration. The effect is particularly significant when registering point clouds with low overlap rates. Attached Figure Description

[0050] Figure 1 This is a flowchart of the method of the present invention;

[0051] Figure 2 This is a schematic diagram of the LTS truncation sequence change in the method of the present invention;

[0052] Figure 3 This is a registration result of the method of the present invention on a Bunny point cloud with a high overlap rate (ξ=0.7);

[0053] Figure 4 This is a registration result of the method of the present invention on a Bunny point cloud with a low overlap rate (ξ=0.5);

[0054] Figure 5 This is a registration result of the method of the present invention on a high overlap (ξ=0.7) Bunny point cloud with added Gaussian noise (σ=0.001);

[0055] Figure 6This is a registration result of the method of the present invention on a low overlap (ξ=0.5) Bunny point cloud with added Gaussian noise (σ=0.001).

[0056] Figure 7 REMS comparison of Bunny point clouds registered with the method of this invention, ICP, and TrICP. Detailed Implementation

[0057] To make the problems and technical solutions related to this invention clearer, the technical solutions of this invention will be described in detail below with reference to the accompanying drawings.

[0058] It is understandable to those skilled in the art that some well-known details may be omitted from the accompanying drawings.

[0059] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent.

[0060] Figure 2 The implementation process of the variable data sequence length strategy is explained. Here, L represents the complete sequence containing the distances of all matched point pairs, with length l; L1 represents the local sequence, with length l1; L2 represents the LTS truncation sequence, with length l2 (determined by the point cloud overlap rate); l and l2 remain constant; N... start With N end These represent the left and right endpoints of the LTS interval, respectively, and N. start The initial value is l-l2, N′ start N in the previous iteration start α is the dynamic step size of the local sequence.

[0061] The local sequence L1 expands to the left from the right endpoint of the complete sequence L. During the first iteration, the local sequence L1 is used as the initial cut sequence for LTS, at which point l1 = l2, meaning the initial length of the local sequence is equal to the length of the cut sequence for LTS.

[0062] As the number of iterations increases, the local sequence L1 expands to the left according to the dynamic step size α, where α is determined by the overlap rate ξ and the current iteration number n. When the initial pose deviation of the point cloud is large, the value of α is small, and the length of the local sequence l1 changes slowly. As the number of iterations increases, the value of α increases continuously, accelerating the expansion of the local sequence L1 to the left. α and l1 are calculated according to the following formulas:

[0063]

[0064] l1=l′1-α×l

[0065] Where l′1 is the length l1 of the local sequence in the previous iteration.

[0066] The current local sequence L1 is used as the data sequence processed by the current LTS. A truncated sequence L2 is obtained by truncating L1 as needed. During the iteration process, the length of the truncated sequence l2 remains constant, while the length of the local sequence l1 continuously increases with each iteration. This causes the position of the truncated sequence L2 on the complete sequence L to continuously shift to the left during registration, thus implementing a variable data sequence length strategy. When the iteration termination condition is met, the maximum value that the local sequence length l1 can achieve is equal to the length l of the complete sequence. At this point, the left endpoint of the truncated sequence L2 coincides with the left endpoint of the complete sequence L.

[0067] according to Figure 3 It can be seen that the method of the present invention can achieve effective registration for point cloud data with an overlap rate of 0.7.

[0068] according to Figure 4 It can be seen that the method of the present invention can achieve effective registration for point cloud data with a low overlap rate of 0.5.

[0069] according to Figure 5 It can be seen that the method of the present invention can achieve effective registration of point cloud data with an overlap rate of 0.7 and Gaussian noise (σ = 0.001) added.

[0070] according to Figure 6 It can be seen that the method of the present invention can achieve effective registration for low overlap point cloud data with an overlap rate of 0.5 and Gaussian noise (σ = 0.001) added.

[0071] Figures 3 to 6 The registration experiment verified the effectiveness and stability of the method of the present invention.

[0072] The root mean square error (REMS) reflects the accuracy of point cloud registration; the smaller the value, the higher the registration accuracy. According to... Figure 7 It can be known that:

[0073] In the first 5 iterations, the REMS curve of the method of the present invention decreased the fastest, indicating that the registration speed of the method of the present invention is better than that of ICP and TrICP at this time.

[0074] During the entire registration process, ICP underwent 13 iterations, the method of this invention underwent 37 iterations, and TrICP underwent 51 iterations. This shows that although the overall registration speed of the method of this invention is slower than that of ICP, it is significantly better than that of TrICP.

[0075] Throughout the registration process, the REMS value of the method of the present invention is less than ICP and TrICP at the end of the iteration, indicating that the method of the present invention has higher registration accuracy.

[0076] Figure 7 The analysis results verify the superiority of the method of the present invention.

[0077] The above description is merely one embodiment of the present invention and is not intended to limit the invention. Those skilled in the art will recognize that the present invention can have various modifications and variations. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A registration method for point clouds with low overlap rate, characterized by: S1. Input the source point cloud and the target point cloud; S2. Determine matching point pairs based on the nearest neighbor principle; S3. Sort the matching point pairs in ascending order of Euclidean distance to obtain the complete matching point pair distance sequence, i.e., the complete sequence L; S4. Calculate the local matching point pair distance sequence, i.e., the local sequence L1, in the complete sequence L based on the point cloud overlap rate and the current iteration number; S5. Apply least-truncation squares to the current local sequence L1 to obtain the truncation matching point pair distance sequence, i.e., the truncation sequence L2; S6. Calculate the pose transformation matrix from the source point cloud to the target point cloud using the current truncated sequence L2; S7. Perform pose transformation on the source point cloud using the pose transformation matrix; S8. Algorithm convergence judgment; In step S4, the complete sequence length l depends on the source point cloud size, and the truncated sequence length l2 depends on the overlap rate ξ, both of which remain unchanged in the registration process; the right end point L1 of the local sequence right is always coincident with the right end point L of the complete sequence tight , and the initial length l1 of the local sequence is the same as the minimum truncated least square truncated sequence length l2; in each iteration, the left end point L1 of the local sequence left is moved left on the complete sequence according to a dynamic moving step α, and when the iteration termination condition is met, the maximum value of the local sequence length l1 that can be achieved is the same as the complete sequence length l; as the iteration proceeds, the local sequence length l1 is constantly growing, so that the variable data sequence length strategy is realized in the registration process; the α depends on the point cloud overlap rate and the current iteration number, and the calculation formulas of α, L1 right and L1 left are as follows: L1 right = L right L1 left = L1' left - (a x l) in, n is the current iteration number, ξ is the point cloud overlap rate value, L1 left L1 of the previous iteration left .

2. The registration method for point clouds with low overlap rate according to claim 1, characterized in that, In step S5, at each iteration, the local sequence L1 resulting from step S4 is processed by the least trimmed squares and the matching point pairs within the interval [N start , N end ] are extracted from the local sequence L1 to form the trimmed sequence L2 of length l2 for pose transformation estimation; the position of the trimmed sequence on the complete sequence varies with the length of the local sequence, the length of the trimmed sequence l2 depends on the overlap ratio ξ and remains constant during the registration process; the trimmed interval can be computed by: N start =N′ start -(α×l) N end =N start +l2 In the interval [N] start N end The sum of squared distances between point pairs can be obtained using the least squares method, as expressed by the following formula: Among them, D i N′ represents the Euclidean distance between the matched point pairs. start N represents the previous iteration start During the iteration process, if N start <0, take N start =0, the local sequence L1 coincides with the complete sequence L.

Images