Point cloud registration method and system based on geometric saliency computation

Abstract

The present application belongs to the technical field of computer graphics image processing, and particularly relates to a point cloud registration method and system based on geometric saliency calculation, comprising the following steps: firstly, a simple and effective method is used to calculate the geometric saliency prior of source point cloud and target point cloud, and all points are divided into high, low and non-salient points. Then, a saliency-aware voting method is used to sort and select the corresponding relationship in the initial corresponding set with pair-wise compatibility constraints. Finally, the confidence value of the selected corresponding is used to further calculate the rigid transformation information by using a robust estimation function. The present application can realize efficient registration of large-scale point cloud based on geometric saliency prior, effectively suppress the influence of false matching and interference points in the point cloud, so as to realize more stable and efficient point cloud alignment.

Description

Technical Field

[0001] This invention belongs to the field of computer graphics and image processing technology, and particularly relates to a point cloud registration method and system based on geometric saliency calculation. Background Technology

[0002] Point cloud registration is a crucial process for aligning 3D point cloud data acquired from multiple perspectives or sensors into a unified coordinate system, and it is widely used in fields such as autonomous driving, robotics, and 3D modeling. Registration allows point clouds acquired from different perspectives to be fused into a complete 3D model. However, the significant geometric differences that may exist in point cloud data from different perspectives present challenges in point cloud registration. This is particularly true for large-scale point cloud registration, where the presence of points with low geometric saliency and numerous outliers often affects the accuracy of the registration results. Although existing 3D keypoint detectors and descriptors have made some progress, significant difficulties remain when processing large-scale point clouds. Therefore, improving the robustness of point cloud registration and reducing the impact of outliers on the results has become a key research issue.

[0003] In point cloud registration, most existing methods rely on geometric information to find correspondences between points. For example, algorithms like RANSAC utilize random sampling consistency to estimate the transformation matrix and eliminate inappropriate correspondences. However, as the number of outliers increases, the computational complexity of these methods also increases, and performance bottlenecks easily occur in the registration of large-scale point clouds. To overcome these problems, researchers have proposed various optimization schemes. Some of these methods attempt to improve the discriminative power of keypoint descriptors through deep learning techniques or design more efficient correspondence selection techniques. However, these methods typically require large amounts of training data and consume significant computational resources, making them difficult to widely implement in practical applications.

[0004] To address the aforementioned problems, this invention proposes a point cloud registration method based on geometric saliency calculation. This method analyzes the geometric saliency of different points in the point cloud, classifying them into high-saliency, low-saliency, and non-saliency points, and uses this saliency information to guide the point cloud registration process. Specifically, we propose a voting and scoring method based on geometric saliency priors to effectively select high-quality correspondences. Furthermore, we combine robust estimation methods to optimize the transformation parameters in the registration process, further improving the accuracy and efficiency of registration. Summary of the Invention

[0005] The main objective of this invention is to propose a point cloud registration method based on point cloud saliency priors, so as to improve the accuracy and efficiency of point cloud registration.

[0006] To achieve the above objectives, this invention provides a point cloud registration method based on point cloud saliency priors, comprising the following steps:

[0007] Step S1: Obtain the two sets of point cloud data to be registered and perform preprocessing;

[0008] Step S2: Classify the point cloud using the geometric saliency calculation framework;

[0009] Step S3: Establish the initial saliency correspondence set;

[0010] Step S4: Selection of voting set based on saliency initial correspondence set and neighborhood consistency measure;

[0011] Step S5: Use the voting set to score all saliency correspondences;

[0012] Step S6: Calculate the rigid registration matrix using the score weighting, and use the rigid registration matrix to achieve point cloud registration.

[0013] Furthermore, in step S2, feature points are first extracted using a geometry-based feature detection method, and then these detected feature points are marked as high saliency points. Next, RANSAC plane segmentation is used to identify points located on the plane region and mark them as non-saliency points. After removing high saliency points and non-saliency points, the remaining points in the point cloud are marked as low saliency points.

[0014] Furthermore, in step S3, let P... 1 and P 2 Representing the source point cloud and the target point cloud respectively, the highly salient point sets are obtained from the source point cloud and the target point cloud respectively using the geometric saliency calculation framework proposed in step S2. and low significance point set Each point is described using a Fast Point Feature Histogram (FPFH). After calculating the FPFH features of the source and target point clouds, a KD-tree based on the feature space is constructed. The nearest neighbor search method is then used to find the most similar point in the target point cloud for each source point cloud point, establishing point-to-point correspondences. Finally, an initial correspondence set C is established on the salient point set, where C = C0. h ∪C l C h It is a set of highly significant points The highly significant correspondence set, C l It is a set of low significance points The low significance correspondence set.

[0015] Furthermore, the formula for calculating the Fast Point Feature Histogram (FPFH) is as follows:

[0016]

[0017] Let the point cloud to be described be point cloud P. For any query point in point cloud P... Calculate its SPFH features with its k nearest neighbors, denoted as . Simultaneous calculation The SPFH features of each nearest neighbor and its k surrounding points are denoted as w k Point and neighboring points The distance between them.

[0018] Furthermore, the Fast Point Feature Histogram (FPFH) is composed of the frequencies of three angle histograms. Each angle histogram has several interval bins, and each interval represents the range of angle information. The histogram is formed by counting the number of points falling in each bin.

[0019] Furthermore, in step S4, λ(c) is used. i ,c j Quantitatively measure a pair of corresponding c i and c j The geometric compatibility between them is expressed by the formula:

[0020]

[0021] Where σ is a specified parameter. and It is a pair of correspondences. and These are the point pairs obtained in step S3 using the nearest neighbor search method; λ(c i ,c j The larger the value, the more likely c is to be affected. i and c j Highly compatible, and vice versa;

[0022] We choose to use highly significant correspondences to build the voting set, and for each highly significant correspondence, we calculate its neighborhood consistency measure:

[0023]

[0024] in It is the highly significant correspondence set C h Inside The KNN neighborhood set; in vector N h ={s i Recorded in} The degree of neighborhood consistency between the corresponding neighborhoods is determined, and then highly significant correspondences with topK scores are selected as voters to construct the voting set V, specifically:

[0025]

[0026] Where topK is a function that extracts the indices of the K largest elements from a given input.

[0027] Furthermore, in step S5, each significance corresponds to c m The rating is:

[0028]

[0029] Wherein, fraction s m Measure the current corresponding c m Geometric compatibility with all highly significant values ​​in the voting set, s m The larger the value, the stronger the geometric compatibility with all corresponding values ​​in the voting set, and vice versa; ultimately, in vector S = {s m The score results corresponding to all saliency are recorded in}.

[0030] Furthermore, based on the saliency correspondence set obtained in step S3... The corresponding score set S = {s} obtained in step S5 m Establish an objective function defined by the weighted distance between corresponding point pairs. By minimizing the objective function, estimate the transformation T consisting of a rotation matrix and a translation vector t. Transform the points in the source point cloud to obtain the transformed source point cloud and align it with the target point cloud.

[0031] Furthermore, the objective function is expressed as:

[0032]

[0033] Where φ(s) m ) is a scoring weighting function:

[0034]

[0035] Where α is a user-specified parameter, and the objective function E(T) is the corresponding point pair The sum of weighted distances between them is solved in a closed form by singular value decomposition.

[0036] The present invention also provides a point cloud registration system based on geometric saliency calculation, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the point cloud registration method based on geometric saliency calculation as described in the above technical solution.

[0037] The beneficial effects of the above-described technical solution of the present invention are as follows:

[0038] This invention provides a point cloud registration method based on geometric saliency priors. It categorizes point clouds into high, low, and non-saliency points using a geometric saliency calculation framework. This classification maximizes the use of high-saliency points during registration while excluding non-saliency points, eliminating interference that could lead to mismatches and reducing the number of mismatches in the initial correspondence set. A voting set is then constructed by selecting reliable correspondences only from those with high saliency, making the correspondences in the voting set more descriptive. The high-confidence voting set is then used to vote and score the entire saliency correspondence set, assigning higher confidence scores to potential interior points in low-saliency correspondences to identify more interior points, while assigning lower scores to outliers in the voting set caused by feature similarity and geometric defects. Finally, the voting scores are used as weights in rigid transformation estimation, penalizing interior points more severely than outliers, thereby significantly mitigating the impact of outliers. Attached Figure Description

[0039] Figure 1 This is a flowchart of the processing of the present invention.

[0040] Figure 2 This is a registration flowchart of the present invention.

[0041] Figure 3 This is a schematic diagram of the classification effect of step S2 of the present invention, wherein (a) is the original point cloud, (b) is a highly significant point, (c) is a non-significant point, and (d) is a low-significant point.

[0042] Figure 4 This is a schematic diagram illustrating the geometric compatibility and neighborhood consistency used in steps S4 and S5 of the present invention.

[0043] Figure 5 This is the registration result of the present invention on the WHU-TLS dataset. Detailed Implementation

[0044] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0045] like Figure 1 The diagram shown is a flowchart of the processing of this invention. The point cloud registration method proposed in this invention, based on geometric saliency calculation, includes the following steps:

[0046] Step S1: Obtain two point clouds to be registered, one of which is the source point cloud and the other is the target point cloud. Perform preprocessing operations by voxel grid downsampling. Specifically, divide the point cloud space into a fixed-size three-dimensional voxel grid, and simplify the point cloud data by calculating the centroid within each voxel so that each voxel retains only one representative point.

[0047] Step S2: It was observed that points in different surface regions contribute significantly differently to the final alignment result. Specifically, points located on surface regions with significant geometric changes, such as sharp edges and corners, may contribute to good alignment results due to their highly expressive geometric features. In contrast, points located within planar regions, due to their high geometric similarity, may tend to decrease the alignment result. Points found in regions not mentioned above are generally considered minor points for registration because their geometric features are difficult to define directly.

[0048] Based on the above observations, such as Figure 2 As shown, this invention designs a geometric saliency calculation framework that divides points into three groups: high, low, and non-saliency points. This classification aims to use as many high-saliency points as possible in registration while excluding non-saliency points. Specifically, the classification results are as follows: Figure 3 As shown, given a point cloud as Figure 3 As shown in (a), feature points are first extracted using a geometry-based feature detection method, and then these detected feature points are marked as highly saliency points, such as... Figure 3 As shown in (b). Next, using RANSAC plane segmentation, we identify points located in the planar region and label them as non-significant points, such as... Figure 3 As shown in (c) above. After removing highly significant and insignificant points, the remaining points in the point cloud are marked as low-significant points, as shown below. Figure 3 As shown in (d).

[0049] Step S3, let P 1 and P 2 Let the source point cloud and the target point cloud be represented respectively. Through the geometric saliency calculation proposed in step S2, we can obtain the highly saliency point sets from the source point cloud and the target point cloud respectively. and low significance point set Each point is described using FPFH (Fast Point Feature Histogram):

[0050]

[0051] Let the point cloud to be described be point cloud P. For any query point in point cloud P... Calculate its SPFH features with its k nearest neighbors, denoted as . Simultaneous calculation The SPFH features of each nearest neighbor and its k surrounding points are denoted as w k Indicates query point and neighboring points The distance between them. The FPFH consists of the frequency of three angular histograms, each with 11 bins. Each bin represents the range of angular information. The histogram is formed by counting the number of points falling into each bin, resulting in a total of 33 bins. After calculating the FPFH features of the source and target point clouds, a KD-tree based on the feature space is constructed. The nearest neighbor search method is used to find the most similar point in the target point cloud for each source point cloud point, establishing point-to-point correspondences. Finally, an initial correspondence set C is established on the salient point set, where C = C0. h ∪C l C h It is a set of highly significant points The highly significant correspondence set, C l It is a set of low significance points The low significance correspondence set.

[0052] Step S4, use λ(c) i ,c j Quantitatively measure a pair of corresponding c i and c j The geometric compatibility between them is expressed by the formula:

[0053] in

[0054] Where σ is a user-specified parameter. and It is a pair of correspondences. and These are the point pairs obtained in step S3 using the nearest neighbor search method. λ(c i ,c j The larger the value, the more likely c is to be affected. i and c j Highly compatible, and vice versa.

[0055] We observe that interior points typically reside in geometrically compatible clusters, and this is particularly evident in interior points selected from highly significant correspondences. Therefore, we choose to utilize highly significant correspondences to construct our voting set, and for each highly significant correspondence, we compute its neighborhood consistency measure:

[0056]

[0057] in It is the highly significant correspondence set C h Inside The KNN neighborhood set, by increasing The size can improve s i While improving accuracy, this increases overall computational complexity. In experiments, the number of adjacent points is typically set to 50 to balance accuracy and efficiency.

[0058] In vector N h ={s i Recorded in} The degree of neighborhood consistency between the corresponding neighborhoods. Then, highly significant correspondences with topK scores are selected as voters for constructing the voting set V, specifically:

[0059]

[0060] Where topK is a function that extracts the indices of the K largest elements from a given input.

[0061] Step S5: Due to feature similarity and geometric defects, some outliers also exist in the voting set of the selected high-significance correspondences. Furthermore, some potential inliers in the low-significance correspondences might improve the alignment results. To identify more inliers by assigning higher confidence scores to some potential inliers in the low-significance correspondences, and to exclude outliers by assigning lower scores to them in the voting set, voting and scoring are performed on all significant correspondences using the voting set V. Each significant correspondence c... m The rating is:

[0062]

[0063] fractions s m Measure the current corresponding c m Geometric compatibility with all highly significant values ​​in the voting set. m The larger the value of , the stronger the geometric compatibility with all corresponding values ​​in the voting set, and vice versa. Ultimately, in vector S = {s} m The score results corresponding to all saliency are recorded in}.

[0064] Step S6 specifically includes the following steps:

[0065] Based on the saliency correspondence set obtained in step S3 The corresponding score set S = {s} obtained in step S5 m Let's establish an objective function defined by the weighted distance between corresponding point pairs. By minimizing the objective function, we can estimate a transformation T consisting of a rotation matrix and a translation vector t, which transforms the points in the source point cloud to align the transformed source point cloud with the target point cloud. Specifically, the objective function is expressed as:

[0066]

[0067] Where φ(s) m ) is a scoring weighting function:

[0068]

[0069] Here, α is a user-specified parameter. The objective function E(T) is the corresponding point pair. The sum of weighted distances between them can be solved in a closed form using singular value decomposition (SVD), just like in classical ICP.

[0070] Figure 5 This paper presents the registration results of this invention and several existing technologies in four example scenarios of the WHU-TLS dataset. Yellow represents the source point cloud, and blue represents the target point cloud. The evaluation metrics used in this invention are MAE (Mean Absolute Error) and ERR (Mean Anisotropy Error). Errors are also calculated for the rotation matrix R and translation vector t composed of the registration results T. A smaller metric indicates higher registration accuracy. Above the results of each method, MAE(R), MAE(t), ERR(R), and ERR(t) are displayed from left to right. The best results are highlighted in bold.

[0071] This invention also provides a point cloud registration system based on geometric saliency calculation, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the point cloud registration method based on geometric saliency calculation as described in the above technical solution.

[0072] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., 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 point cloud registration method based on geometric saliency calculation, characterized in that, Includes the following steps: Step S1: Obtain the two sets of point cloud data to be registered and perform preprocessing; Step S2: Classify the point cloud using the geometric saliency calculation framework; Step S3: Establish the initial saliency correspondence set; In step S3, let and Representing the source point cloud and the target point cloud respectively, the highly salient point sets are obtained from the source point cloud and the target point cloud respectively using the geometric saliency calculation framework proposed in step S2. , and low significance point set , ; Each point is described using a Fast Point Feature Histogram (FPFH). After obtaining FPFH features from the source and target point clouds, a KD tree based on the feature space is constructed. The nearest neighbor search method is used to find the most similar point in the target point cloud for each source point cloud point, and a point-to-point correspondence is established. Finally, an initial correspondence set C is established on the salient point set, where... , It is a set of highly significant points , The highly significant correspondence set, It is a set of low significance points , The set of low-significance correspondences; Step S4: Selection of voting set based on saliency initial correspondence set and neighborhood consistency measure; In step S4, highly significant correspondences are selected to build the voting set, and for each highly significant correspondence, its neighborhood consistency measure is calculated: ； in It is a highly significant correspondence set Inside The KNN neighborhood set, Indicates correspondence and Geometric compatibility between them; in vectors Recorded in The degree of neighborhood consistency between the corresponding neighborhoods is determined, and then highly significant correspondences with topK scores are selected as voters to construct the voting set V, specifically: ； Where topK is a function that extracts the indices of the K largest elements of a given input; Step S5: Use the voting set to score all saliency correspondences; Step S6: Calculate the rigid registration matrix using score weighting, and use the rigid registration matrix to achieve point cloud registration; Based on the saliency correspondence set obtained in step S3 and the correspondence score set obtained in step S5, an objective function defined by the weighted distance between corresponding point pairs is established. By minimizing the objective function, the transformation T, consisting of a rotation matrix and a translation vector t, is estimated. The transformed source point cloud is then aligned with the target point cloud.

2. The point cloud registration method based on geometric saliency calculation as described in claim 1, characterized in that: In step S2, feature points are first extracted using a geometry-based feature detection method, and then these detected feature points are marked as high saliency points. Next, RANSAC plane segmentation is used to identify points located in the planar region and mark them as non-saliency points. After removing high saliency points and non-saliency points, the remaining points in the point cloud are marked as low saliency points.

3. The point cloud registration method based on geometric saliency calculation as described in claim 1, characterized in that: The formula for calculating the Fast Point Feature Histogram (FPFH) is as follows: ； Let the point cloud to be described be point cloud P. For any query point in point cloud P... Calculate its SPFH features with its k nearest neighbors in the neighborhood, denoted as Simultaneous calculation The SPFH features of each nearest neighbor and its k surrounding points are denoted as ; Point and neighboring points The distance between them.

4. The point cloud registration method based on geometric saliency calculation as described in claim 1, characterized in that: The Fast Point Feature Histogram (FPFH) consists of the frequencies of three angle histograms. Each angle histogram has several bins, and each bin represents the range of angle information. The histogram is formed by counting the number of points falling in each bin.

5. The point cloud registration method based on geometric saliency calculation as described in claim 1, characterized in that: In step S4, use Quantitatively measure a pair of corresponding and The geometric compatibility between them is expressed by the formula: ； in It is a specified parameter. and It is a pair of correspondences. and These are the point pairs obtained in step S3 using the nearest neighbor search method; The larger the value, the more likely it is to represent and Highly compatible, and vice versa.

6. The point cloud registration method based on geometric saliency calculation as described in claim 5, characterized in that: In step S5, each saliency corresponds to The rating is: ； Among them, fractions Measure the current corresponding Geometric compatibility with all highly significant values ​​in the voting set. The larger the value, the stronger the geometric compatibility with all corresponding values ​​in the voting set, and vice versa; ultimately, in the vector... The score results corresponding to all saliency are recorded in the record.

7. The point cloud registration method based on geometric saliency calculation as described in claim 1, characterized in that: The objective function is expressed as: ； in It is a scoring weighting function: ； in A user-specified parameter, the objective function are corresponding point pairs The sum of weighted distances between them is solved in a closed form by singular value decomposition.

8. A point cloud registration system based on geometric saliency calculation, characterized in that: It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the point cloud registration method based on geometric saliency calculation as described in any one of claims 1-7.

Images