A Robust 3D Point Cloud Feature Matching Method Based on Progressive Consensus Voting

Abstract

The present invention provides a robust 3D point cloud feature matching method based on progressive consistency voting. The method involves acquiring 3D point cloud data containing a source point cloud and a target point cloud; performing feature extraction and feature matching on the 3D point cloud data to obtain an initial matching set; evaluating the compatibility of each matching correspondence in the initial matching set with candidate correspondences in the current voting set, and voting on the results. Progressive consistency voting (PCV) is used to assign a confidence score to each correspondence. The confidence scores of a voting set and its corresponding confidence scores are then iteratively optimized. The Otsu thresholding method (OTSU) is then used to select correspondences from the final scoring results to form a transformation set. The RanSAC algorithm is used to calculate the rotation matrix of the transformation set, and the source point cloud is converted to the target point cloud according to the rotation matrix. Therefore, the present invention can achieve more accurate and robust feature matching of 3D point clouds.

Description

Technical Field

[0001] The present invention belongs to the technical field of computer vision optimization, and in particular relates to a robust three-dimensional point cloud feature matching method based on progressive consistency voting. Background Art

[0002] With the increasing level of automation, the requirements for computer vision systems are becoming increasingly stringent in fields such as remote sensing, robotic navigation, autonomous driving, and cultural relic restoration. Object recognition based on two-dimensional images is no longer sufficient for everyday tasks. Therefore, 3D object recognition technology based on three-dimensional point cloud feature matching has become a research hotspot in computer vision in recent years. The continuous development of point cloud acquisition technology and the emergence of a large number of affordable point cloud acquisition devices have greatly stimulated enthusiasm for point cloud data processing research.

[0003] At present, the research methods for feature matching of three-dimensional point clouds mainly include calculating binary labels and calculating confidence: the method of calculating binary labels is to assign binary labels (i.e. true or false) to the correspondence. Ma et al. (Ma J, Zhao J, Jiang J, et al. Locality preserving matching [J]. International Journal of Computer Vision, 2019, 127 (5): 512-531.) proposed a locality preserving matching (LPM) model to identify the correct correspondence, thereby completing the feature matching of point clouds. Tombari and Stefano proposed a 3D Hough voting (3DHV) method (Tombari F, Di Stefano L. Object recognition in 3D scenes with occlusions and clutter by hough voting [C] / / 2010 Fourth Pacific-Rim Symposium on Image and Video Technology. IEEE, 2010: 349-355.). This method uses the local reference frame (LRF) of each key point to project each corresponding item into the 3D Hough space. The projected points that form a cluster in the Hough space are identified as internal points, thereby completing the feature matching of the point cloud. Methods that calculate confidence include: Rodol et al. (Rodolà E, Albarelli A, Bergamasco F, et al. A scale-independent selection process for 3D object recognition in cluttered scenes [J]. International journal of computer vision, 2013, 102 (1): 129-145.) Use the game theory matching (GTM) model to select correspondences that satisfy the global geometric consistency constraint, thereby completing the feature matching of the point cloud.

[0004] Existing methods suffer from low matching accuracy during the feature matching process of 3D point clouds. This is due to the following reasons: First, low accuracy. Due to the complex surface structure of the object being measured, there may be many false matches during the feature matching process, resulting in poor matching accuracy. Second, they suffer from low robustness. Existing methods are subject to certain influences on point cloud data, such as Gaussian noise, changes in data resolution, clutter and occlusion, and changes in data modality, making it difficult to obtain satisfactory matching results. Summary of the Invention

[0005] In order to solve the above problems existing in the prior art, the present invention provides a robust 3D point cloud feature matching method based on progressive consistency voting. The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0006] The present invention provides a robust three-dimensional point cloud feature matching method based on progressive consistency voting, which includes:

[0007] Step 1: Obtain 3D point cloud data including source point cloud and target point cloud;

[0008] Step 2: Perform feature extraction and feature matching on the three-dimensional point cloud data to obtain an initial matching set;

[0009] The initial matching set includes at least one corresponding relationship describing the matching between the source point cloud and the target point cloud;

[0010] Step 3: For each matching relationship in the initial matching set, perform a compatibility evaluation between the corresponding relationship and the candidate corresponding relationships in the current voting set to obtain a compatibility score between each matching relationship in the initial matching set and the candidate corresponding relationships;

[0011] Wherein, the current voting set consists of the corresponding relations with the highest voting scores;

[0012] Step 4: Vote for the compatibility scores of each correspondence in the initial matching set and the candidate correspondence to obtain the voting score of each correspondence in the initial matching set;

[0013] Step 5: Update the current voting set according to the voting score of each corresponding relationship in the initial matching set, and repeat steps 3 to 4 until the maximum number of iterations is reached, and use the last voting score as the final scoring result;

[0014] Step 6: Using the Otsu threshold method (OTSU), select corresponding relationships from the corresponding relationships of the final scoring results to form a transformation set;

[0015] Step 7: Use the RanSAC algorithm to calculate the rotation matrix of the transformation set, and complete the transformation from the source point cloud to the target point cloud according to the rotation matrix.

[0016] Optionally, before step 2, the robust three-dimensional point cloud feature matching method based on progressive consistency voting further includes:

[0017] The three-dimensional point cloud data is downsampled in a voxelized manner.

[0018] Optionally, step 3 includes:

[0019] Step 3-1: For each matching correspondence in the initial matching set, calculate the rigid constraint of the Euclidean distance between the corresponding correspondence and the candidate correspondence in the current voting set;

[0020] Step 3-2: Based on the rigid constraints, calculate the corresponding relationship of each match in the initial matching set, and calculate the compatibility score of the corresponding relationship with the candidate corresponding relationships in the current voting set.

[0021] Optionally, the rigid constraint of the Euclidean distance is:

[0022]

[0023] in, and are two arbitrary source point clouds, and are two arbitrary target point clouds;

[0024] The compatibility score between each correspondence in the initial matching set and the candidate correspondence is:

[0025]

[0026] Where c1 represents the correspondence in the initial matching set, c2 represents the candidate correspondence in the current voting set, τ is the distance parameter, τ = 10pr;

[0027]

[0028] Among them, P i Represents a point cloud collection Any point in general, P nn It's P i The nearest neighbor point of pr represents the point cloud resolution.

[0029] Optionally, step 4 includes:

[0030] For each correspondence in the initial matching set, all compatibility scores of the correspondence and the candidate correspondences are integrated, and the integrated result is used as the voting score of the correspondence to obtain the voting score of each correspondence in the initial matching set.

[0031] Optionally, the voting score of each corresponding relationship in the initial matching set is:

[0032]

[0033] Among them, c represents the corresponding relationship in the initial matching set, c j represents the corresponding relationship of the jth candidate in the current voting set, represents the voting set of the i-th iteration.

[0034] Optionally, in step 5, updating the current voting set according to the voting score of each corresponding relationship in the initial matching set includes:

[0035] Sort each corresponding relationship in the initial matching set in descending order according to the voting score;

[0036] The current voting set is updated according to the initial matching set sorted in descending order.

[0037] Among them, the updated current voting set is:

[0038]

[0039] in, is the initial matching set of the i-1th iteration, c j represents the jth candidate correspondence in the current voting set, and k is the current voting set of the i-th iteration

[0040] Optionally, step 6 includes:

[0041] Step 6-1: Based on the voting score distribution of the final scoring result, use the Otsu threshold method (OTSU) to set the adaptive threshold;

[0042] Step 6-2: Select the corresponding relations with voting scores greater than the adaptive threshold from the corresponding relations of the final scoring results to form a transformation set.

[0043] Optionally, in step 7:

[0044] Step 7-1: Calculate the nearest target point cloud corresponding to each source point cloud in the transformation set;

[0045] Step 7-2: Find the rigid transformation that minimizes the average distance between each source point cloud and the corresponding nearest target point cloud in the transformation set, and obtain the translation matrix and rotation matrix;

[0046] Step 7-3: Use the translation matrix and rotation matrix to translate and rotate each source point cloud in the transformation set to obtain a new transformation point set;

[0047] Step 7-4: Determine whether the new transformation set and the target point cloud set satisfy the objective function. If not, repeat steps 7-1 to 7-3 until the new transformation set and the target point cloud set satisfy the objective function.

[0048] Among them, the target point cloud set is composed of target point clouds corresponding to the source point clouds.

[0049] Beneficial effects of the present invention:

[0050] (1) The present invention provides a robust three-dimensional point cloud feature matching method based on progressive consistency voting, and proposes a progressive optimization selection voting scheme. The voting set in PCV is gradually optimized using the scoring results of the previous iteration, providing better pairing performance and improving the robustness of pairing.

[0051] (2) The present invention provides a robust 3D point cloud feature matching method based on progressive consistency voting, proposes an adaptive cardinality to cardinally transform the voting set to handle corresponding sets of different scales and internal ratios, and provides better accuracy performance.

[0052] (3) The present invention provides a robust three-dimensional point cloud feature matching method based on progressive consistency voting. The proposed PCV defines the correctness of the correspondence based on global geometric consistency and has a high-quality voting set. It can better avoid the influence of Gaussian noise, data extraction, clutter and occlusion and data modal changes on feature matching to achieve the best performance. In addition, the PCV algorithm has high time efficiency and can effectively improve the point cloud registration performance.

[0053] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 This is a flowchart of a robust three-dimensional point cloud feature matching method based on progressive consistency voting provided by the present invention;

[0055] Figure 2 is a schematic diagram of an algorithm for measuring compatibility between matches provided by an embodiment of the present invention;

[0056] Figure 3 Schematic diagram of obtaining an adaptive threshold using the OTSU algorithm to determine the voting set cardinality according to an embodiment of the present invention;

[0057] Figure 4 This is the registration result diagram of the solution of the present invention. DETAILED DESCRIPTION

[0058] The present invention will be further described in detail below with reference to specific examples, but the embodiments of the present invention are not limited thereto.

[0059] like Figure 1 As shown, the present invention provides a robust three-dimensional point cloud feature matching method based on progressive consistency voting, which includes:

[0060] Step 1: Obtain 3D point cloud data including source point cloud and target point cloud;

[0061] It is worth noting that 3D point cloud information is a collection of scattered points that contain information such as normal vectors. These scattered points can reflect the general structure of the scene, but too many points can lead to redundancy and matching errors, resulting in poor rendering effects.

[0062] The present invention can perform downsampling processing on the three-dimensional point cloud data in a voxelized manner.

[0063] It is worth noting that the present invention requires that the density of the point cloud be evenly distributed in each spatial coordinate to prevent data density from affecting computing efficiency. In order to reduce calculation errors caused by excessive point density or uneven distribution, the principle of the point cloud point cloud reduction method is to grid the first-dimensional space, also known as voxelization. After gridding, each grid is called a voxel (Voxel). The basic idea of point cloud reduction is to check each voxel. If there is a point in the voxel, a point is used to replace the point set in the voxel. The coordinates of this point can be the center coordinates of the voxel or the expected coordinates of the point set. Using voxel grids to achieve point reduction can reduce the number of points while ensuring the shape characteristics of the point cloud, improve the speed of algorithms such as registration, surface reconstruction, and shape recognition, and ensure optimal accuracy.

[0064] Step 2: Perform feature extraction and feature matching on the three-dimensional point cloud data to obtain an initial matching set;

[0065] The initial matching set includes at least one corresponding relationship describing the matching between the source point cloud and the target point cloud;

[0066] It is worth noting that: since the feature values of point clouds on the same or similar surfaces are very similar and robust to point cloud noise, this step extracts the features of each point cloud and then matches the local key points for feature matching to obtain the initial matching set C initial .

[0067] Step 3: For each matching relationship in the initial matching set, perform a compatibility evaluation between the corresponding relationship and the candidate corresponding relationships in the current voting set to obtain a compatibility score between each matching relationship in the initial matching set and the candidate corresponding relationships;

[0068] Wherein, the current voting set consists of the corresponding relations with the highest voting scores;

[0069] It is worth noting that the present invention will have the highest s f (p s , p t ) value, is defined as the initial voting set use Voting. Among them, p s , p tare two key points in the source point cloud and the target point cloud respectively. initial and voting set C voting The correctness of most candidate matches in the set is evaluated. Matches with high correctness receive high voting scores. The voting set is then updated based on the voting scores, and the compatibility of each correspondence with all candidates in the predefined voting set is repeatedly checked in an iterative manner. Because only correct correspondences are geometrically compatible with each other, the calculation of the voting score relies on the evaluation of the pairwise compatibility of two correspondences. The specific process is:

[0070] Step 3-1: For each matching correspondence in the initial matching set, calculate the rigid constraint of the Euclidean distance between the corresponding correspondence and the candidate correspondence in the current voting set;

[0071] Step 3-2: Based on the rigid constraints, calculate the corresponding relationship of each match in the initial matching set, and calculate the compatibility score of the corresponding relationship with the candidate corresponding relationships in the current voting set.

[0072] The present invention calculates the pairwise compatibility of two corresponding relationships using the Euclidean distance constraint (L2) method. The Euclidean distance constraint between the source point cloud and the target point cloud is as follows: Figure 2 shown.

[0073] The rigid constraints of the Euclidean distance are:

[0074]

[0075] in, and are two arbitrary source point clouds, and are two arbitrary target point clouds;

[0076] By using the Euclidean distance constraint, only the key point positions already provided in the key point detection stage, that is, the Euclidean distance between the source point cloud and the target point cloud, can be used to compare their geometric constraints, and performance optimization can be achieved by gradually optimizing the voters.

[0077] Therefore, the present invention defines the compatibility score between each correspondence in the initial matching set and the candidate correspondence based on the rigid constraint of the Euclidean distance as:

[0078]

[0079] Where c1 represents the correspondence in the initial matching set, c2 represents the candidate correspondence in the current voting set, τ is the distance parameter, τ = 10pr;

[0080]

[0081] Among them, P i Represents a point cloud collection Any point in general, P nn It's P i The nearest neighbor point of pr represents the point cloud resolution.

[0082] Obtaining the compatibility score F(c1, c2) between point clouds facilitates comparison of the correctness of matching between point clouds and improves the correctness of subsequent voting score results.

[0083] Step 4: Vote for the compatibility scores of each correspondence in the initial matching set and the candidate correspondence to obtain the voting score of each correspondence in the initial matching set;

[0084] The present invention integrates all compatibility scores of each corresponding relationship in the initial matching set with the candidate corresponding relationships, uses the integrated result as the voting score of the corresponding relationship, and obtains the voting score of each corresponding relationship in the initial matching set.

[0085] It is worth noting that after evaluating the compatibility, the present invention needs to calculate the voting score of each compatibility score. In each iteration of PCV (for example, the i-th iteration), the voting score s i (c) will be assigned to each corresponding c.s i (c) is the sum of c and the voting set C in the current iteration initial The integral of the compatibility scores of each candidate in the ballot. More specifically, s i (c) is defined as:

[0086]

[0087] Among them, s i (c) is the voting score of each corresponding relationship in the initial matching set, c represents the corresponding relationship in the initial matching set, c j represents the corresponding relationship of the jth candidate in the current voting set, represents the voting set of the i-th iteration.

[0088] From the above equation, we can see that the voting score is independent of feature similarity, which is sensitive to clutter, occlusion, and limited overlap. Therefore, the present invention repeats the voting score calculation process once and uses the voting score from the last iteration as the final scoring result, thereby improving the reliability of the voting score.

[0089] Step 5: Update the current voting set according to the voting score of each corresponding relationship in the initial matching set, and repeat steps 3 to 4 until the maximum number of iterations is reached, and use the last voting score as the final scoring result;

[0090] The present invention can sort each corresponding relationship in the initial matching set in descending order according to the voting score; and update the current voting set according to the initial matching set sorted in descending order.

[0091] It is worth mentioning that after the voting scores are determined, the definition and update of the voting set are carried out. The definition and update of the voting set are important components of the PCV method. The basic goal of this step is to initial Find a subset with a high internal ratio as the voting set C voting , so voters will be more persuasive and voting results will be easier to distinguish. However, each corresponding label is unknown, and the label is also the target of feature matching. To solve this problem, we propose the following method to gradually optimize the voting set.

[0092] In the first iteration of PCV, we define the voting set as s f (p s ,p t ) corresponding to the set with the highest value of Voting. It is worth noting that The definition of has relatively little impact on the performance of PCV, and PCV can even be used in a system composed of randomly selected pairs of Achieve good results.

[0093] Because s f (p s ,p t ) are sensitive to clutter, occlusion, and limited overlap, It may contain many outliers. The present invention can update the score based on the last iteration result This approach is reasonable because it has been previously verified that the confidence scores of correspondences can be optimized within the consensus voting framework. In addition, we experimentally observed that the quality of the correspondence scoring results will improve during the iteration process. Therefore, in the i-th iteration of PCV, Can be updated as follows:

[0094]

[0095] in, is the current voting set for the i-th time, is the current voting set after the i-1-th update, is the initial matching set of the i-1th iteration, c j represents the jth candidate correspondence in the current voting set, and k is the current voting set of the i-th iteration

[0096] Step 6: Using the Otsu threshold method (OTSU), select corresponding relationships from the corresponding relationships of the final scoring results to form a transformation set;

[0097] In the above formula (5), the parameter k needs special attention because its goal is to Introduce as many inliers as possible while rejecting outliers. If we fix k, then a small value of k will not be able to include all potential inliers. If C initial If the number of inline items in is less than k, then large values of k will be dragged into outliers. To overcome this problem, we propose a method to make the value of k adaptive. Specifically, we Apply Otsu threshold method to obtain an adaptive threshold τ otsu Then, the present invention defines k as the number of votes greater than τ otsu The corresponding number of , because the voting score calculated using geometric constraints has a certain degree of discrimination. Get the adaptive threshold τ otsu The adaptive strategy to determine the voting set cardinality is as follows Figure 3 Therefore, the voting set of the PCV method uses the scoring results of the previous iteration for progressive optimization on the one hand, and has an adaptive cardinality to cope with initial matching sets with different scales and higher ratios on the other hand. The process is:

[0098] Step 6-1: Based on the voting score distribution of the final scoring result, use the Otsu threshold method (OTSU) to set the adaptive threshold;

[0099] Step 6-2: Select the corresponding relations with voting scores greater than the adaptive threshold from the corresponding relations of the final scoring results to form a transformation set.

[0100] Step 7: Use the RanSAC algorithm to calculate the rotation matrix of the transformation set, and complete the transformation from the source point cloud to the target point cloud according to the rotation matrix.

[0101] It is worth noting that the rotation matrix is calculated to obtain the registration relationship between the two point cloud data. Point cloud registration is essentially to transform the point cloud from one coordinate system to another. Point cloud registration usually requires two point cloud data. The first type of point cloud data is called the original point cloud. The second type of point cloud data is called the target point cloud. Point cloud registration is to display the original point cloud on the coordinates of the target point cloud. We can determine the transformation relationship of the coordinates by finding point clouds with similar features in the point cloud. Assume that the rigid body transformation occurs from the original point cloud to the target point cloud, that is, the original point cloud can obtain the target point cloud by rotation and translation. The rotation and translation process here is represented by the rotation transformation matrix R and the translation transformation matrix T. The iterative closest point algorithm (ICP) calculates the rotation matrix and R and translation matrix T. The specific steps are as follows:

[0102] Step 7-1: Calculate the transformation set The target point cloud set corresponding to each source point cloud The nearest target point cloud;

[0103] Step 7-2: Find the transformation set A rigid transformation with the minimum average distance between each source point cloud and the corresponding nearest target point cloud is used to obtain the translation matrix and the rotation matrix;

[0104] Step 7-3: Use the translation matrix and rotation matrix to translate and rotate each source point cloud in the transformation set to obtain a new transformation point set.

[0105] Step 7-4: Determine the new transformation set Set with target point cloud Whether the objective function is satisfied. If not, repeat 7-1 to 7-3 until the new transformation set and the target point cloud set satisfy the objective function.

[0106] Among them, the target point cloud set is composed of target point clouds corresponding to the source point clouds.

[0107] Objective function:

[0108] Among them, min is a given threshold, which makes the two point set data meet the optimal matching under a certain measurement criterion.

[0109] refer to Figure 4 , Figure 4 This is the result of the three-dimensional scene point cloud registration of the present invention. Figure 4 The effect display in the figure shows that the matching accuracy is high.

[0110] During actual operation, the algorithm of the present invention is as follows:

[0111] Requirements: Source point cloud and target point cloud

[0112] make sure: and The consistent features between them.

[0113] 1: Generate C by matching local key point descriptors initial ;

[0114] 2: Initialize iteration index i = 0 and initial transformation set

[0115] 3: Based on s f (p s ,p t ) Define the initial voting set

[0116] 4: When i <N iter When N iter For the maximum number of iterations, do the following loop;

[0117] 5: Use formula (2) to calculate c∈C initial and Compatibility scores between;

[0118] 6: Use formula (4) to calculate the voting score, c∈C initial ;

[0119] 7: Update using formula (5)

[0120] 8: i = i + 1;

[0121] 9: End the loop;

[0122] 10: C is evaluated based on the score of the last iteration. initial Sort by

[0123] 11: Push the correspondence with the highest score into C inlier ;

[0124] 12: Return to C inlier .

[0125] The present invention has at least the following advantages:

[0126] (1) The present invention proposes a progressive optimization voting scheme, which uses the scoring results of the previous iteration to gradually optimize the voting set in PCV, providing better pairing performance and improving the robustness of pairing.

[0127] (2) The present invention proposes an adaptive cardinality to cardinally transform the voting set to handle corresponding sets of different scales and internal ratios, thereby providing better accuracy performance.

[0128] (3) The PCV of the present invention defines the correctness of the correspondence based on global geometric consistency and has a high-quality voting set, which can better avoid the influence of Gaussian noise, data extraction, clutter and occlusion and data modal changes on feature matching to achieve optimal performance. In addition, the PCV algorithm has high time efficiency and can effectively improve the point cloud registration performance.

[0129] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of the technical features being referred to. Thus, a feature identified as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, "plurality" means two or more, unless otherwise specifically defined.

[0130] Although the present application is described herein with reference to various embodiments, those skilled in the art will be able to understand and implement other variations of the disclosed embodiments in practicing the claimed application by examining the drawings, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality.

[0131] The above is a further detailed description of the present invention in conjunction with specific preferred embodiments, and the specific implementation of the present invention should not be considered to be limited to these descriptions. For those skilled in the art of the present invention, without departing from the concept of the present invention, several simple deductions or substitutions can be made, which should be considered to fall within the scope of protection of the present invention.

Claims

1. A robust 3D point cloud feature matching method based on progressive consistency voting, characterized by: include: Step 1: Obtain 3D point cloud data including source point cloud and target point cloud; Step 2: Perform feature extraction and feature matching on the three-dimensional point cloud data to obtain an initial matching set; The initial matching set includes at least one corresponding relationship describing the matching between the source point cloud and the target point cloud; Step 3: For each matching relationship in the initial matching set, perform a compatibility evaluation between the corresponding relationship and the candidate corresponding relationships in the current voting set to obtain a compatibility score between each matching relationship in the initial matching set and the candidate corresponding relationships; Wherein, the current voting set consists of the corresponding relations with the highest voting scores; Step 4: Vote for the compatibility scores of each correspondence in the initial matching set and the candidate correspondence to obtain the voting score of each correspondence in the initial matching set; Step 5: Update the current voting set according to the voting score of each corresponding relationship in the initial matching set, and repeat steps 3 to 4 until the maximum number of iterations is reached, and use the last voting score as the final scoring result; Step 6: Using the Otsu threshold method (OTSU), select corresponding relationships from the corresponding relationships of the final scoring results to form a transformation set; Step 7: Use the RanSAC algorithm to calculate the rotation matrix of the transformation set, and complete the transformation from the source point cloud to the target point cloud according to the rotation matrix.

2. The robust 3D point cloud feature matching method based on progressive consistency voting according to claim 1, characterized in that: Before step 2, the robust three-dimensional point cloud feature matching method based on progressive consistency voting further includes: The three-dimensional point cloud data is downsampled in a voxelized manner.

3. The robust 3D point cloud feature matching method based on progressive consistency voting according to claim 2, characterized in that: The step 3 comprises: Step 3-1: For each matching correspondence in the initial matching set, calculate the rigid constraint of the Euclidean distance between the corresponding correspondence and the candidate correspondence in the current voting set; Step 3-2: Based on the rigid constraints, calculate the corresponding relationship of each match in the initial matching set, and calculate the compatibility score of the corresponding relationship with the candidate corresponding relationships in the current voting set.

4. The robust 3D point cloud feature matching method based on progressive consistency voting according to claim 3, characterized in that: The rigid constraints of the Euclidean distance are: in, and are two arbitrary source point clouds, and are two arbitrary target point clouds; The compatibility score between each correspondence in the initial matching set and the candidate correspondence is: Where c1 represents the correspondence in the initial matching set, c2 represents the candidate correspondence in the current voting set, τ is the distance parameter, τ = 10pr; Among them, P i Represents a point cloud collection Any point in general, P nn It's P i The nearest neighbor point of pr represents the point cloud resolution.

5. The robust 3D point cloud feature matching method based on progressive consistency voting according to claim 4, characterized in that: The step 4 comprises: For each correspondence in the initial matching set, all compatibility scores of the correspondence and the candidate correspondences are integrated, and the integrated result is used as the voting score of the correspondence to obtain the voting score of each correspondence in the initial matching set.

6. The robust 3D point cloud feature matching method based on progressive consistency voting according to claim 5, characterized in that: The voting score of each corresponding relationship in the initial matching set is: Among them, c represents the corresponding relationship in the initial matching set, c j represents the corresponding relationship of the jth candidate in the current voting set, represents the voting set of iteration i.

7. The robust 3D point cloud feature matching method based on progressive consistency voting according to claim 1, characterized in that: In step 5, updating the current voting set according to the voting score of each corresponding relationship in the initial matching set includes: Sort each corresponding relationship in the initial matching set in descending order according to the voting score; The current voting set is updated according to the initial matching set sorted in descending order.

8. The robust 3D point cloud feature matching method based on progressive consistency voting according to claim 7, characterized in that: The updated current voting set is: in, is the initial matching set of the i-1th iteration, c j represents the jth candidate correspondence in the current voting set, and k is the current voting set of the i-th iteration 9. The robust 3D point cloud feature matching method based on progressive consistency voting according to claim 1, characterized in that: The step 6 comprises: Step 6-1: Based on the voting score distribution of the final scoring result, use the Otsu threshold method (OTSU) to set the adaptive threshold; Step 6-2: Select the corresponding relations with voting scores greater than the adaptive threshold from the corresponding relations of the final scoring results to form a transformation set.

10. The robust 3D point cloud feature matching method based on progressive consistency voting according to claim 1, characterized in that: Step 7: Step 7-1: Calculate the nearest target point cloud corresponding to each source point cloud in the transformation set; Step 7-2: Find the rigid transformation that minimizes the average distance between each source point cloud and the corresponding nearest target point cloud in the transformation set, and obtain the translation matrix and rotation matrix; Step 7-3: Use the translation matrix and rotation matrix to translate and rotate each source point cloud in the transformation set to obtain a new transformation point set; Step 7-4: Determine whether the new transformation set and the target point cloud set satisfy the objective function. If not, repeat 7-1 to 7-3 until the new transformation set and the target point cloud set satisfy the objective function. Among them, the target point cloud set is composed of target point clouds corresponding to the source point clouds.

Images