Point Cloud Processing Method Based on Deep Learning and Adaptive Topology-Preserving Non-Rigid Registration

Abstract

To overcome the problem that it is difficult to ensure the cross-section measurement accuracy due to the severe outliers, missing values, and noise superposition in the cross-section contour measurement point cloud of an irregular cross-section annular component, and that traditional point cloud processing methods are difficult to effectively process the measurement point cloud, the present invention proposes a point cloud processing method based on deep learning and adaptive topology-preserving non-rigid registration. By utilizing the strong function fitting ability and feature learning ability of the PointNet++ network for deep learning, an outlier segmentation point network is constructed, and local fine features are extracted through a hierarchical network architecture and multi-scale grouping to achieve continuous outlier removal of the contour measurement point cloud. Moreover, during non-rigid registration, global topology constraints and local topology constraints are introduced into the objective function established based on the probability density estimation function, solving the complex deformation problems involved in non-rigid registration. Finally, the point cloud of the CAD model after non-rigid registration replaces the measurement point cloud, obtaining a point cloud that can meet the geometric quality assessment accuracy.

Description

Technical Field

[0001] The present invention belongs to the field of complex component measurement, and particularly relates to a point cloud processing method based on deep learning and adaptive topology-preserving non-rigid registration, which is particularly applicable to the geometric quality inspection field of irregular cross-section annular components in high-end equipment such as aviation, aerospace, and nuclear power. Background Art

[0002] Irregular cross-section annular components are one of the key components widely used in high-end equipment such as aviation, aerospace, and nuclear power. Due to their irregular cross-sections and complex cross-sectional shapes, it is difficult to detect them, and thus the forming quality cannot be guaranteed. Taking the superalloy W-shaped seal ring as an example, it has an overall annular closed structure, a small cross-sectional contour size, and typical circular and steep inclined wall features, resulting in its geometric quality inspection relying heavily on destructive sampling final inspection and the lack of quality inspection means during the forming process. In recent years, optical measurement has attracted great interest due to its portability, flexibility, and relatively high measurement accuracy. However, when using a line laser profile sensor to nondestructively measure the cross-sectional contour of the superalloy W-shaped seal ring, the following problems exist:

[0003] 1. Affected by internal and external environmental factors, measurement point clouds usually have serious outliers (especially continuous outliers), missing, and noise superposition problems. Traditional point cloud processing methods (such as denoising and patching) are difficult to handle synergistically, resulting in the inability to perform subsequent geometric quality assessment. For example, optical measurement methods are easily affected by internal and external environmental factors. When the internal and external environments are unsuitable, continuous outliers occur in the measurement point cloud of the irregular cross-sectional contour. Such outliers are different from ordinary outliers (sparse isolated points deviating from the main body of the measurement point cloud), and their manifestation is a large number of continuous pseudo-points deviating from the main body of the measurement point cloud. Traditional outlier removal methods (such as outlier removal methods based on statistical or radius ideas) cannot effectively remove them; local severe missing causes real contour points to be isolated and thus be mis-removed by the denoising algorithm; existing interpolation-based missing patching methods are interfered by continuous outliers and perform wrong patching between continuous outliers and contour points. Multi-view registration measurement is an effective method to solve missing, and denoising is performed on this basis. However, multi-view registration often requires arranging multiple sensors or designing a multi-degree-of-freedom motion measurement platform, which is expensive and complex. Moreover, when measuring the superalloy W-shaped seal ring with a small cross-sectional contour size, the sensor layout and measurement motion planning are restricted, and multi-view registration cannot be performed.

[0004] 2. The density non-uniformity and disorder of measurement point clouds have always been key problems that need to be solved in the field of optical measurement. It heavily relies on resampling and sorting algorithms after fitting. However, the irregular cross-sectional contour is complex, making it difficult to perform the above processing. Measurement point clouds with density non-uniformity and disorder are not conducive to subsequent processing and geometric quality assessment.

[0005] Therefore, it has become a key problem to be solved urgently to process the measured point cloud with serious outliers, missing and noise superposition problems and transform it into an ordered and consistent-density point cloud data in order to achieve accurate cross-section measurement of irregular cross-section annular components. Summary of the Invention

[0006] In order to overcome the technical problems that it is difficult to guarantee the cross-section measurement accuracy due to the serious outliers, missing and noise superposition in the cross-section contour measurement point cloud of irregular cross-section annular components, and the traditional point cloud processing methods are difficult to effectively process the measured point cloud, the present invention proposes a point cloud processing method based on deep learning and adaptive topology-preserving non-rigid registration.

[0007] To achieve the above object, the present invention adopts the following technical solutions:

[0008] A point cloud processing method based on deep learning and adaptive topology-preserving non-rigid registration, which is characterized in that it is used to process the original contour measurement point cloud of an irregular cross-section annular component, and includes the following steps:

[0009] Step 1: Use the PointNet++ network architecture to divide the original contour measurement point cloud into contour points and outlier points, remove the outlier points, and obtain the inner and outer contour point clouds after removing the outlier points;

[0010] Step 2: Stitch the inner and outer contour point clouds after removing the outlier points to obtain the stitched contour measurement point cloud;

[0011] Step 3: Under the condition of adding global and local constraints, adaptively topologically preserve non-rigidly register the CAD model point cloud of the irregular cross-section annular component to the stitched contour measurement point cloud obtained in Step 2 to obtain the CAD model point cloud after non-rigid registration, and use the CAD model point cloud after non-rigid registration to replace the contour measurement point cloud as the point cloud for subsequent geometric quality evaluation of the irregular cross-section annular component.

[0012] Further, the method of using the PointNet++ network architecture to divide the original contour measurement point cloud into contour points and outlier points in Step 1 is:

[0013] First, gradually extract the local fine features of the original contour measurement point cloud along the hierarchical structure in multiple scales;

[0014] Then, gradually propagate the local fine features of the subsampled points back to the original contour measurement point cloud to obtain the scores of each point of the original contour measurement point cloud, and divide the original contour measurement point cloud into contour points and outlier points according to the scores of each point.

[0015] Further, the method of point cloud stitching in Step 2 is:

[0016] Based on the pre - calibration results of the point cloud acquisition system for collecting the original contour measurement point cloud, coordinate transformation is performed to transform the outer contour point cloud after removing outliers into the inner contour point cloud coordinate system to achieve point cloud stitching, or to transform the inner contour point cloud after removing outliers into the outer contour point cloud coordinate system to achieve point cloud stitching.

[0017] Further, the method of adaptive topology - preserving rigid registration in step 3 is as follows:

[0018] First, the problem of registering the CAD model point cloud of the irregular cross - section annular member to the stitched contour measurement point cloud is converted into a probability density estimation function:

[0019]

[0020] Where:

[0021] X N×D represents the stitched contour measurement point cloud;

[0022] N represents the number of points in the stitched contour measurement point cloud;

[0023] D represents the dimension of the stitched contour measurement point cloud;

[0024] x n represents a point in the stitched contour measurement point cloud;

[0025] M represents the number of points in the CAD model point cloud;

[0026] ω is the weight coefficient representing the noise prior level;

[0027] π M+1 = ω;

[0028]

[0029] m = M + 1;

[0030]

[0031] σ 2 is the anisotropic variance;

[0032] is the non - rigid transformation;

[0033] θ is the non - rigid transformation parameter;

[0034] y m represents a point in the CAD model point cloud;

[0035] Then, the anisotropic variance σ in the probability density estimation function is iteratively solved by an improved EM algorithm. 2and non-rigid transformation parameters θ until convergence, so as to obtain an optimally aligned non-rigid transformation matrix At this time, the solution of the probability density estimation function is completed;

[0036] Finally, through the optimally aligned non-rigid transformation matrix non-rigidly register the CAD model point cloud to the stitched contour measurement point cloud;

[0037] The improved EM algorithm is specifically as follows:

[0038] In the E step, calculate the matching probability p n between the point x m in the contour measurement point cloud and the point y old (m|x n ),

[0039] In the M step, introduce global topological constraints and local topological constraints into the objective function established based on the probability density estimation function, and then minimize the objective function after introducing the constraints to find new non-rigid transformation parameters θ and anisotropic variance σ 2 .

[0040] Furthermore, in the M step, add a step of updating the weight coefficient ω representing the noise prior level to achieve the adaptiveness of noise judgment during the registration process.

[0041] The present invention also provides a point cloud processing system based on deep learning and adaptive topology-preserving non-rigid registration, which is characterized in that it includes:

[0042] An outlier segmentation point network module, implemented using the PointNet++ network architecture, for segmenting the contour points and outlier points in the original contour measurement point cloud, and removing the outlier points to obtain the inner and outer contour point clouds after removing the outlier points;

[0043] A point cloud stitching module for stitching the inner and outer contour point clouds after removing the outlier points;

[0044] An adaptive topology-preserving non-rigid registration module for non-rigidly registering the CAD model point cloud of the irregular cross-section ring-shaped member to the contour measurement point cloud obtained after stitching to obtain the non-rigidly registered CAD model point cloud, and the non-rigidly registered CAD model point cloud is used for subsequent geometric quality evaluation of the irregular cross-section ring-shaped member.

[0045] Furthermore, the outlier segmentation point network module includes a first-level combination and a second-level combination arranged in sequence;

[0046] The first-level combination includes multiple sets of sequentially arranged sampling layers, grouping layers, and feature extraction layers. The local fine features of the original contour measurement point cloud are gradually extracted at multiple scales along the hierarchical structure through the sampling layer, grouping layer, and feature extraction layer;

[0047] The second-level combination includes interpolation and cross-level skip connections corresponding to the first-level combination. Through the interpolation and cross-level skip connections, the local fine features of the subsampled points are gradually propagated back to the original contour measurement point cloud along the hierarchical structure to obtain the scores of each point in the original contour measurement point cloud. According to the scores of each point, each point in the original contour measurement point cloud is divided into contour points and outlier points.

[0048] Furthermore, the adaptive topology-preserving non-rigid registration module includes a non-rigid transformation matrix This non-rigid transformation matrix is used to non-rigidly register the point cloud of the CAD model of the irregular cross-section annular member to the stitched contour measurement point cloud; the non-rigid transformation matrix is obtained through the following method:

[0049] First, assume the point cloud of the CAD model as the centroid of the Gaussian mixture model, and assume the stitched contour measurement point cloud as the corresponding data. Transform the problem of registering the point cloud of the CAD model to the stitched contour measurement point cloud into a probability density estimation function:

[0050]

[0051] where, X N×D represents the contour measurement point cloud; N represents the number of points in the contour measurement point cloud; D represents the dimension of the stitched contour measurement point cloud; x n represents the points in the stitched contour measurement point cloud (n = 1,..., N); M represents the number of points in the CAD model point cloud; ω is the weight coefficient representing the noise prior level; π M+1 = ω; m = M + 1; σ 2 is the anisotropic variance; is the non-rigid transformation matrix, θ is the non-rigid transformation parameter; y m represents the points in the CAD model point cloud;

[0052] Then, the anisotropic variance σ 2 and the non-rigid transformation parameter θ in the probability density estimation function are iteratively solved through the improved EM algorithm until convergence. At this time, the solution of the probability density estimation function is completed, and the optimized aligned non-rigid transformation matrix

[0053] The specific improved EM algorithm is as follows:

[0054] In the E-step of the EM algorithm, calculate the matching probability p of the point x in the stitched contour measurement point cloud n and the point y in the CAD model point cloud m ; old (m|x n )

[0055] In the M-step of the EM algorithm, establish the objective function Q(θ,σ 2 ) based on the probability density estimation function, and add global topological constraints and local topological constraints to it, and then minimize the objective function Q1(W,σ 2 ) after adding the constraints to find new parameters:

[0056]

[0057]

[0058] Wherein, is the estimated current number of matching points; p old (m|x n ) is the prior matching probability, α and λ represent two trade-off parameters between two topological constraint terms.

[0059] The present invention also provides a storage medium, on which a computer program is stored; characterized in that: when the computer program is run by a processor, it executes the above method.

[0060] The present invention also provides an electronic device, including a processor and a storage medium; a computer program is stored on the storage medium; characterized in that: when the computer program is run by the processor, it executes the above method.

[0061] The beneficial effects of the present invention are as follows:

[0062] 1. The present invention transforms (rotates, translates, and deforms) the CAD model point cloud to the measured point cloud based on adaptive topology-preserving non-rigid registration. Based on the robustness of non-rigid registration, the CAD model point cloud can automatically identify outliers, missing data, and noise during the transformation to the measured point cloud. When the optimal alignment of the two is achieved, i.e., after non-rigid registration, the transformed CAD model point cloud replaces the measured point cloud, overcoming the problem of the superposition of outliers, missing data, and noise in the point cloud that is difficult to solve by traditional point cloud processing methods. At the same time, the processing result is ordered and density-uniform point cloud data, providing a good data basis for accurately measuring the cross-sectional profile of irregular cross-section annular components and facilitating subsequent geometric quality assessment. Taking six high-temperature alloy seal rings as the measurement objects, the average deviation between the point cloud processed by the present invention and the traditional destructive measurement data is between 8 - 15 μm, with very high precision.

[0063] 2. Traditional non-rigid registration algorithms generally deal with outliers in the measured point cloud by pre-modeling. The pre-modeling method can effectively handle ordinary outliers, thus ensuring registration robustness. However, in the measurement of irregular cross-sectional profiles, the internal and external environments often cause continuous outliers to appear in the measured point cloud. Continuous outliers are very similar to the contour point features. At this time, the pre-modeling method cannot remove continuous outliers, resulting in subsequent non-rigid registration failure. Moreover, the problems solved by previous non-rigid registration do not involve continuous outliers. Therefore, to improve the robustness of the non-rigid registration algorithm against continuous outliers, the present invention utilizes the strong function fitting ability and feature learning ability of PointNet++ network deep learning to construct an outlier segmentation point network, extracts local fine features through a hierarchical network architecture and multi-scale grouping, and realizes the removal of continuous outliers in the measured point cloud of irregular cross-sectional profiles, laying a foundation for subsequent non-rigid registration.

[0064] 3. When the measurement contour differs greatly from the corresponding CAD model, non-rigid registration will involve complex deformations, resulting in registration failure and inability to complete cross-sectional profile measurement. The present invention overcomes this problem of complex deformations by using an improved EM algorithm to solve the probability density estimation function during non-rigid registration. Specifically, global and local topological constraint terms are introduced into the objective function established based on the probability density estimation function to maintain the global and local topology of the point cloud to be registered, effectively dealing with complex deformations.

[0065] 4. The present invention adds a step of updating the weight coefficient in the M step of the EM algorithm, which can realize the self-adaptation of noise during the non-rigid registration process and further improve the processing accuracy of the point cloud. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 It is the flowchart of the method of the present invention.

[0067] Figure 2It is a schematic diagram of the principle of the double-line laser profile sensor alignment measurement system.

[0068] Figure 3 It is a schematic diagram of the inner contour point cloud A and the outer contour point cloud B of an irregular cross-section.

[0069] Figure 4 It is a schematic diagram of the outlier segmentation point network architecture constructed by the present invention.

[0070] Figure 5 It is a schematic diagram of the inner and outer contour point clouds after outlier removal.

[0071] Figure 6 It is a schematic diagram of the contour measurement point cloud after splicing by the present invention.

[0072] Figure 7 It is a schematic diagram of the final processing result of the present invention.

[0073] Explanation of reference numerals:

[0074] 1 - Double-line laser profile sensor alignment measurement system; 2 - W-shaped superalloy seal ring. Detailed implementation manners

[0075] The following further describes the point cloud processing method and the point cloud processing system of the present invention by taking the contour measurement point cloud processing of the irregular cross-section of the W-shaped superalloy seal ring as an example in conjunction with the accompanying drawings.

[0076] See the attached Figure 2 , using a point cloud acquisition system, such as the double-line laser profile sensor alignment measurement system 1, to acquire the original contour measurement point cloud of the irregular cross-section of the W-shaped superalloy seal ring 2, and the contour measurement point cloud includes an inner contour point cloud A and an outer contour point cloud B (see the attached Figure 3 ).

[0077] The inner contour point cloud and the outer contour point cloud are respectively represented as and where: and respectively represent the nth measurement point in the inner contour point cloud and the outer contour point cloud; D represents the point cloud dimension; represents the D-dimensional real vector space; N1 and N2 respectively represent the number of points in the point cloud; A and B respectively represent the measurement coordinate systems of the inner contour point cloud and the outer contour point cloud.

[0078] The method for processing the contour measurement point cloud of the irregular cross-section of the W-shaped superalloy seal ring by using the method of the present invention is specifically as follows:

[0079] Step 1: Outlier segmentation

[0080] Using the PointNet++ network architecture, the original contour measurement point cloud is divided into contour points and outlier points, and the outlier points existing in the original contour measurement point cloud are removed to obtain the inner contour point cloud after outlier removal. and the outer contour point cloud Where: N′1 and N′2 respectively represent the number of points in the inner and outer contour point clouds after outlier removal.

[0081] Step 2: Point cloud stitching

[0082] Stitch the inner contour point cloud after outlier removal and the outer contour point cloud Where, are respectively the coordinate values of point on the x-axis, are respectively the coordinate values of point on the z-axis, and the stitched contour measurement point cloud is obtained x n represents the point in the stitched contour measurement point cloud.

[0083] In this embodiment, the point cloud stitching is performed based on the pre-calibration result of the point cloud acquisition system, such as the line laser contour sensor alignment measurement system (the pre-calibration method is a well-known existing method). The outer contour point cloud after outlier removal is transformed into the coordinate system of the inner contour point cloud to achieve point cloud stitching, and the stitched contour measurement point cloud is obtained

[0084] Step 3: Adaptive topology-preserving non-rigid registration

[0085] Under the condition of adding global and local constraints, the CAD model point cloud of the W-shaped superalloy seal ring is non-rigidly registered to the contour measurement point cloud obtained after stitching in Step 2 to obtain the CAD model point cloud after non-rigid registration Use the CAD model point cloud after non-rigid registration to replace the contour measurement point cloud obtained after stitching as the point cloud for subsequent geometric quality evaluation of the W-shaped superalloy seal ring. Where: y m represents the m-th point in the CAD model point cloud; represents the D-dimensional real vector space; M represents the number of points in the CAD model point cloud.

[0086] In this embodiment, the CAD model point cloud of the W-shaped superalloy seal ring is generated from the corresponding CAD model.

[0087] In addition to the point cloud processing method, the present invention also provides a point cloud processing system based on deep learning and adaptive topology-preserving non-rigid registration. This system is capable of processing the point cloud of irregular cross-section profile measurements of a W-shaped high-temperature alloy sealing ring. The point cloud processing system includes an outlier segmentation point network module, a point cloud splicing module, and an adaptive topology-preserving non-rigid registration module.

[0088] The outlier segmentation point network module is implemented using the PointNet++ network architecture and is used to segment the contour points and outlier points in the original contour measurement point cloud, and remove the outlier points to obtain the inner and outer contour point clouds after removing the outlier points;

[0089] like Figure 4 As shown, the outlier segmentation point network module adopts the PointNet++ network architecture, including the first-level combination and the second-level combination set in sequence.

[0090] The first level combination includes multiple sets of sampling layers, grouping layers, and feature extraction layers, which are set in sequence. Through the sampling layers, grouping layers, and feature extraction layers, local fine features of the input point cloud (the original collected inner and outer contour point cloud) are gradually extracted along the hierarchical structure at multiple scales. Among them:

[0091] The sampling layer selects the local area centroid from the original input point cloud (i.e., the point cloud collected by the point cloud acquisition system) by iterative farthest point sampling. Before the original input point cloud is input into the first level combination, the number of segmentation categories needs to be set to 2 in advance.

[0092] The grouping layer uses the Ball query function to find the multi-scale surrounding neighbors of the local area centroid and establish a local area set of multiple scales.

[0093] The feature extraction layer is a PointNet network unit, which extracts local fine features of the point cloud from local regions at multiple scales. The PointNet network unit consists of a first T-net layer, a second T-net layer, multiple multi-layer perceptrons (MLPs), and a feature fusion layer.

[0094] The second-level combination includes interpolation and cross-level jump links set corresponding to the first-level combination. The local fine features of the sub-sampling points are gradually propagated back to the original input point cloud along the hierarchical structure through interpolation and cross-level jump links to obtain the scores of each point in the original input point cloud. Each point has two scores, namely, the outlier point score and the contour point score. If the outlier point score of a point is high, then the point is an outlier point. If the contour point score of a point is high, then the point is a contour point. According to the two scores of each point, each point in the input point cloud can be divided into contour points and outlier points. The points judged as outliers are removed to obtain the inner contour point cloud after removing the outliers. and outer contour point cloud likeFigure 5 as shown

[0095] The above-mentioned along the hierarchy means dividing the input point cloud into N regions, each region having a center point. The features of each region are extracted through a network to represent the features of the center point. Then, the N center points are further divided into M sub-regions. There are also M center points in these sub-regions. The features of the M sub-regions are continuously extracted through the network to represent the features of the M center points, and so on. The number of divisions depends on the set number of layers of the hierarchical combination.

[0096] The local fine features of the subsampled points refer to the features of the center points in the divided regions.

[0097] A point cloud stitching module for stitching the inner and outer contour point clouds after outlier point removal; in this embodiment, the point cloud stitching module is implemented based on the pre-calibration result of the line laser contour sensor alignment measurement system (the pre-calibration method is a well-known method in the art), including a rotation matrix and a translation vector:

[0098]

[0099] Among them,

[0100]

[0101] In the formula, represents the coordinate transformed to coordinate system A; represents the rotation matrix from the outer contour point cloud coordinate system B to the inner contour point cloud coordinate system A; represents the translation vector from the outer contour point cloud coordinate system B to the inner contour point cloud coordinate system A; θ T represents the angle between the outer contour point cloud coordinate system B and the inner contour point cloud coordinate system A; a and b represent the translation amounts of the outer contour point cloud coordinate system B to the inner contour point cloud coordinate system A in the x-axis and z-axis directions; represents the x-axis coordinate value in the inner contour point cloud coordinate system A, represents the z-axis coordinate value in the inner contour point cloud coordinate system A.

[0102] Through the rotation matrix and the translation vector perform coordinate transformation on the outer contour point cloud, unify it to the inner contour point cloud coordinate system A, and achieve stitching. After stitching, the contour measurement point cloud In other embodiments, the same method can also be used to perform coordinate transformation on the inner contour point cloud, unify it to the outer contour point cloud coordinate system B, and achieve stitching.

[0103] An adaptive topology-preserving non-rigid registration module is used to non-rigidly register the point cloud of the CAD model of the W-shaped superalloy seal ring to the point cloud of the spliced contour measurement, obtaining the point cloud of the CAD model after non-rigid registration. The point cloud of the CAD model after non-rigid registration is used for subsequent geometric quality evaluation of the W-shaped superalloy seal ring.

[0104] The adaptive topology-preserving non-rigid registration module includes a non-rigid transformation matrix This non-rigid transformation matrix is used to non-rigidly register the point cloud of the CAD model of the W-shaped superalloy seal ring to the point cloud of the spliced contour measurement; the non-rigid transformation matrix is obtained by the following method:

[0105] First, assume the point cloud of the CAD model as the centroid of the Gaussian mixture model, assume the point cloud of the spliced contour measurement as the corresponding data, and transform the point cloud of the CAD model to the point cloud of the spliced contour measurement so that the registration problem is converted into a probability density estimation function:

[0106]

[0107] where X N×D represents the point cloud of the contour measurement; N represents the number of points in the point cloud of the contour measurement; D represents the dimension of the point cloud of the spliced contour measurement; x n represents the points in the point cloud of the spliced contour measurement (n = 1,..., N); M represents the number of points in the point cloud of the CAD model; ω is the weight coefficient representing the noise prior level; π M+1 = ω; m = M + 1; σ 2 is the anisotropic variance; is the non-rigid transformation matrix, θ is the non-rigid transformation parameter; y m represents the points in the point cloud of the CAD model (m = 1,..., M).

[0108] Then, the anisotropic variance σ 2 and the non-rigid transformation parameter θ in the probability density estimation function are iteratively solved by the improved EM algorithm (expectation-maximization algorithm) until convergence. At this time, the solution of the probability density estimation function is completed, and the optimized aligned non-rigid transformation matrix

[0109] The EM algorithm is specifically as follows:

[0110] In the E step of the EM algorithm, calculate the points x n in the point cloud of the spliced contour measurement and the points ym The matching probability p old (m|x n );

[0111] In the M-step of the EM algorithm, minimize the objective function Q(θ,σ 2 ) established based on the probability density estimation function to find new parameters:

[0112]

[0113] Wherein, is the estimated current number of matching points; p old (m|x n ) is the prior matching probability,

[0114] The improvement of the improved EM algorithm of the present invention is specifically: in the M-step of the above traditional EM algorithm, the global topology constraint and the local topology constraint are introduced into the objective function Q(θ,σ 2 ) established based on the probability density estimation function. The global local topology is coherent point drift, and the local topology constraint is local linear embedding. Both are introduced in the form of a regularization term, that is, the global topology constraint regularization term E GL and the local topology constraint regularization term E LO are introduced to maintain the global local topological structure of the registered point cloud to cope with the complex deformations generated in non-rigid registration.

[0115] After introducing the global topology constraint and the local topology constraint, the objective function Q(θ,σ 2 ) established based on the probability density estimation function is transformed into:

[0116]

[0117] Among them, α and λ represent two trade-off parameters between two topological constraint terms, and are valued according to experience. In the present invention, it is recommended that α be valued at 100 and λ be valued at 5000000.

[0118] When the EM algorithm iteration stops, the non-rigid transformation matrix of the optimal alignment obtained after the iteration stops is used to non-rigidly register the CAD model point cloud to the stitched contour measurement point cloud to obtain the CAD model point cloud Y′ after non-rigid registration M×D :

[0119] Y′ M×D =Y M×D +GW

[0120] Wherein, G represents the kernel matrix; W represents the coefficient matrix of the kernel; Y′ M×D is the CAD model point cloud after non-rigid registration

[0121] The point cloud of the CAD model after non-rigid registration As shown in the appendix Figure 7 As can be seen from the comparison of the original point cloud in Figure 7 and Figure 3 , the point cloud shown in Figure 7 has removed the outliers in the original point cloud, completed the splicing, missing repair and noise correction of the point cloud data. Therefore, using the point cloud of the CAD model after non-rigid registration to replace the point cloud of the spliced contour measurement can accurately measure the irregular cross-section contour of the W-shaped superalloy seal ring 2.

[0122] As a further optimization, the present invention can also add a link to update the weight coefficient in the M step of the EM algorithm, which can realize the self-adaptation of judging noise during the registration process to improve the processing accuracy of the point cloud.

[0123] On the other hand, the present invention also provides a storage medium, on which a computer program is stored. When the computer program is run by a processor, it can execute the point cloud processing method provided by the present invention.

[0124] On another aspect, the present invention also provides an electronic device, including a processor and a storage medium, wherein a computer program is stored on the storage medium. When the computer program is run by the processor, it can execute the point cloud processing method provided by the present invention.

Claims

1. A point cloud processing method based on deep learning and adaptive topology-preserving non-rigid registration, characterized in that, It is used to process the original contour measurement point cloud of an annular member with an irregular cross-section, including the following steps: Step 1: Using the PointNet++ network architecture, divide the original contour measurement point cloud into contour points and outlier points, remove the outlier points, and obtain the inner and outer contour point clouds after removing the outlier points; Step 2: Stitch the inner and outer contour point clouds after removing the outlier points to obtain the stitched contour measurement point cloud; Step 3: Under the condition of adding global and local constraints, adaptively topologically preserve the non-rigid registration of the CAD model point cloud of the irregular cross-section annular member to the stitched contour measurement point cloud obtained in Step 2 to obtain the non-rigidly registered CAD model point cloud, and use the non-rigidly registered CAD model point cloud to replace the contour measurement point cloud as the point cloud for subsequent geometric quality evaluation of the irregular cross-section annular member; The method of the adaptive topology-preserving non-rigid registration is as follows: First, transform the problem of registering the CAD model point cloud of the irregular cross-section annular member to the stitched contour measurement point cloud into a probability density estimation function: Where: X N×D represents the contour measurement point cloud after splicing; N represents the number of points in the stitched contour measurement point cloud; D represents the dimension of the stitched contour measurement point cloud; x n represents a point within the contour measurement point cloud after splicing; M represents the number of points in the CAD model point cloud; ω is the weight coefficient representing the prior level of noise; π M+1 = ω; m = M + 1; σ 2 is the anisotropic variance; is a non-rigid transformation; θ is the non-rigid transformation parameter; y m represents a point within the point cloud of the CAD model; Then, the anisotropic variance σ in the probability density estimation function is iteratively solved by the improved EM algorithm 2 and the non-rigid transformation parameter θ until convergence, so as to obtain the non-rigid transformation matrix for optimal alignment At this time, the solution of the probability density estimation function is completed; Finally, through the non-rigid transformation matrix of the optimal alignment non-rigidly register the CAD model point cloud to the stitched contour measurement point cloud; The improved EM algorithm is specifically as follows: Calculate the point x in the contour measurement point cloud in the E-step n with the point y in the CAD model point cloud m for the matching probability p old (m|x n ) Introduce the global topological constraint and the local topological constraint into the objective function established based on the probability density estimation function in the M step, and then minimize the objective function after introducing the constraints to find the new non-rigid transformation parameters θ and the anisotropic variance σ 2 .

2. The point cloud processing method based on deep learning and adaptive topology-preserving non-rigid registration according to claim 1, wherein: The method of using the PointNet++ network architecture to divide the original contour measurement point cloud into contour points and outlier points in Step 1 is as follows: First, gradually extract the local fine features of the original contour measurement point cloud along the hierarchical structure in a multi-scale manner; Then, gradually propagate the local fine features of the subsampled points back to the original contour measurement point cloud to obtain the scores of each point in the original contour measurement point cloud, and divide the original contour measurement point cloud into contour points and outlier points according to the scores of each point.

3. The point cloud processing method based on deep learning and adaptive topology-preserving non-rigid registration according to claim 2, characterized in that: The method of point cloud stitching in Step 2 is as follows: Based on the pre-calibration result of the point cloud acquisition system used to acquire the original contour measurement point cloud, perform coordinate transformation to transform the outer contour point cloud after removing the outlier points into the inner contour point cloud coordinate system to achieve point cloud stitching, or transform the inner contour point cloud after removing the outlier points into the outer contour point cloud coordinate system to achieve point cloud stitching.

4. The point cloud processing method based on deep learning and adaptive topology-preserving non-rigid registration according to claim 3, characterized in that: Add a step of updating the weight coefficient ω representing the prior level of noise in the M step to achieve the adaptability of noise judgment during the registration process.

5. A point cloud processing system based on deep learning and adaptive topology-preserving non-rigid registration, characterized in that, Including: An outlier segmentation point network module, implemented using the PointNet++ network architecture, used to segment the contour points and outlier points in the original contour measurement point cloud, and remove the outlier points to obtain the inner and outer contour point clouds after removing the outlier points; A point cloud stitching module, used to stitch the inner and outer contour point clouds after removing the outlier points; An adaptive topology-preserving non-rigid registration module, used to non-rigidly register the CAD model point cloud of the irregular cross-section annular member to the stitched contour measurement point cloud to obtain the non-rigidly registered CAD model point cloud, and the non-rigidly registered CAD model point cloud is used for subsequent geometric quality evaluation of the irregular cross-section annular member; The adaptive topology-preserving non-rigid registration module includes a non-rigid transformation matrix This non-rigid transformation matrix T is used to non-rigidly register the point cloud of the CAD model of the irregular cross-section annular member to the point cloud of the spliced contour measurement; the non-rigid transformation matrix is obtained by the following method: First, assume the point cloud of the CAD model as the centroid of the Gaussian mixture model, and assume the point cloud of the contour measurement after stitching as the corresponding data. Transform the point cloud of the CAD model to the point cloud of the contour measurement after stitching The registration problem is converted into a probability density estimation function: Among them, X N×D represents the contour measurement point cloud; N represents the number of points in the contour measurement point cloud; D represents the dimension of the contour measurement point cloud after splicing; x n represents the points in the contour measurement point cloud after splicing (n = 1,..., N); M represents the number of points in the CAD model point cloud; ω is the weight coefficient representing the noise prior level; π M+1 = ω; m = M + 1; σ 2 is the anisotropic variance; is the non-rigid transformation matrix, θ is the non-rigid transformation parameter; y m represents the points in the CAD model point cloud; Then, the anisotropic variance σ in the probability density estimation function is iteratively solved by the improved EM algorithm 2 and the non-rigid transformation parameter θ until convergence. At this time, the probability density estimation function is solved, and the optimized non-rigid transformation matrix T is obtained; The improved EM algorithm is specifically as follows: Calculate the matching probability p n of the point x m in the concatenated contour measurement point cloud old with the point y n in the CAD model point cloud (m|x ) In the M-step of the EM algorithm, the objective function Q(θ,σ 2 ) is established based on the probability density estimation function, and global topological constraints and local topological constraints are added to it. Then, the objective function Q1(W,σ 2 ) after adding the constraints is minimized to find new parameters: Among them, is the estimated current number of matching points; p old (m|x n ) is the prior matching probability, α and λ represent two trade-off parameters between two topological constraint terms.

6. The point cloud processing system based on deep learning and adaptive topology-preserving non-rigid registration according to claim 5, characterized in that: The outlier segmentation point network module includes a first-level combination and a second-level combination arranged in sequence; The first-level combination includes multiple groups of sampling layers, grouping layers, and feature extraction layers arranged in sequence. The local fine features of the original contour measurement point cloud are gradually extracted at multiple scales along the hierarchical structure through the sampling layer, the grouping layer, and the feature extraction layer; The second-level combination includes interpolation and cross-level skip links arranged corresponding to the first-level combination. The local fine features of the subsampled points are gradually propagated back to the original contour measurement point cloud along the hierarchical structure through the interpolation and cross-level skip links to obtain the scores of each point in the original contour measurement point cloud. According to the scores of each point, each point in the original contour measurement point cloud is divided into contour points and outlier points.

7. Storage medium, on which a computer program is stored; characterized in that: When the computer program is run by a processor, it executes the method according to any one of claims 1-4.

8. An electronic device, comprising a processor and a storage medium; a computer program is stored on the storage medium; characterized in that: When the computer program is run by the processor, it executes the method according to any one of claims 1-4.

Images