A registration method for registering partial measurement data of a vehicle body to an overall CAD model

Abstract

The present invention discloses a registration method for partial body measurement data to the overall CAD model, which acquires the overall body CAD model and the partial body measurement data, obtains the point cloud to be registered and the target point cloud, and then constructs the RSCS set based on the point cloud to be registered, the RSCS set based on the target point cloud, and performs preprocessing; and constructs a Spinnet point cloud neural network to extract descriptors to obtain each descriptor set; uses two-way matching to perform rough registration on each descriptor set, and uses the NICP algorithm based on correcting the normal vector direction for fine registration after rough registration; finally, iteratively obtains a highly accurate matching pair set using the registration error of the matching pair set after fine registration; the present invention realizes an algorithm model from rough registration to fine registration by constructing a descriptor based on deep learning, solves the disadvantage that the manual feature descriptor in the prior art does not have generalization, overcomes the single problem covered by a single sphere, and comprehensively improves the matching accuracy and efficiency on the basis of reducing the calculation amount.

Description

Technical Field

[0001] The present invention relates to the field of digital detection technology, and particularly relates to a method for registering partial measurement data of a vehicle body to an overall CAD model. Background Art

[0002] In traditional manufacturing industries, such as aircraft body manufacturing and automobile body manufacturing, manufacturing errors or assembly errors on the surface have a significant impact on performance indicators such as the overall aerodynamic layout. Therefore, detecting whether the errors are within the allowable range is a crucial part of the product life cycle. The traditional manual fixture detection method is quite time-consuming and laborious, while the current digital twin detection method enables the whole process to achieve high-efficiency automation and intelligence. The digital twin of an object is usually obtained by non-contact three-dimensional scanning of the object's shape, so that the object to be inspected can be twinned in a virtual computer space in the form of a dense three-dimensional point cloud.

[0003] Among them, the key step is to compare the local scanned point cloud with the overall CAD ideal design model, and this process is a point cloud registration problem from local to overall. However, usually for the point cloud of a vehicle body or an aircraft body, the features are not obvious, and it is easy to have incorrect matching point pairs or find multiple matching point pairs during the registration process, resulting in an incorrect final registration result and thus invalidating the detection result.

[0004] The point cloud registration algorithm is divided into two main steps: coarse registration and fine registration. The main problem solved by coarse registration is to pull two point clouds to a direction where the three axes are roughly aligned when the poses of the target point cloud and the point cloud to be registered are quite different, while fine registration fine-tunes the two point clouds to achieve an appropriate registration accuracy.

[0005] In local-to-global point cloud registration, fine registration depends on the corresponding points found in coarse registration. If there are many incorrect matching points, it will surely affect the subsequent fine registration. Therefore, during the coarse registration process, the two point clouds need to be divided into multiple sub-blocks, and their similar sub-blocks are found for alignment. In the process of generating descriptors for sub-blocks (Patches), the most commonly used algorithm is the FPFH (Fast Point Feature Histogram) algorithm. It builds a histogram by weighted statistics of the relationships between adjacent points within a certain range. Each spherical body of a color represents the neighborhood of a key point, and most point cloud block division methods are similar to this. Other mainly used algorithms also include the Shot (Signatures of Histograms of Orientations) descriptor, which is a descriptor generation algorithm based on a local reference coordinate system, and the 4PCS (4-Points Congruent Sets) descriptor based on the RANSAC (Random Sample Consensus) matching framework. They all calculate the features of the point cloud through fixed formulas.

[0006] These existing methods have three obvious disadvantages:

[0007] 1) In terms of key point detection, it is necessary to traverse and calculate all points to finally determine which are the key points that need to be matched. This undoubtedly requires a huge amount of computation, resulting in very poor timeliness of the entire registration and unable to support the online detection requirements in intelligent manufacturing;

[0008] 2) Most of the point cloud descriptors in the existing technical solutions adopt calculation methods of manual features, which will cause unnecessary waste of computing resources, and the descriptors calculated by manual features do not have generalization. When the target point cloud is scanned from an object that is relatively regular, smooth or has no obvious features, there will be too many incorrect point pairs, directly leading to a poor final result;

[0009] 3) When there is a large amount of noise in the point cloud, directly using the ordinary ICP algorithm to reduce the distance between points has a poor effect because the noise points are not the correct corresponding points. Summary of the Invention

[0010] In view of the above deficiencies in the prior art, the present invention provides a registration method for the measurement data of a vehicle body part to the overall CAD model.

[0011] In order to achieve the above invention purpose, the technical solution adopted by the present invention is as follows:

[0012] A registration method for the measurement data of a vehicle body part to the overall CAD model, comprising the following sub-steps:

[0013] S1. Discretize the overall CAD model of the vehicle body to obtain the target point cloud of the vehicle body;

[0014] S2. Collect partial measurement data of the vehicle body to obtain the point cloud to be registered;

[0015] S3. Respectively construct the RSCS set based on the point cloud to be registered and the RSCS set based on the target point cloud according to the point cloud to be registered and the target point cloud;

[0016] S4. Preprocess the RSCS set based on the point cloud to be registered and the RSCS set based on the target point cloud respectively;

[0017] S5. Train the Spinnet point cloud neural network using the triple loss function to obtain the optimized Spinnet point cloud neural network;

[0018] S6. Use the optimized Spinnet point cloud neural network to extract the descriptors in the preprocessed RSCS set based on the point cloud to be registered and the RSCS set based on the target point cloud respectively, and obtain the descriptor set based on the point cloud to be registered and the descriptor set based on the target point cloud respectively;

[0019] S7. Coarsely register each descriptor set in step S6 using bidirectional matching to obtain the set of matched pairs after coarse registration;

[0020] S8. Use the NICP algorithm based on correcting the normal vector direction to finely register the set of matched pairs after coarse registration;

[0021] S9. Obtain the registration error of the set of matched pairs after fine registration, and return to step S3 to reconstruct the RSCS set based on the point cloud to be registered and the RSCS set based on the target point cloud according to the registration error until the preset registration error threshold is met.

[0022] The present invention has the following beneficial effects:

[0023] By collecting partial body measurement data, the point cloud to be registered is obtained. The overall target point cloud of the body is obtained by discretizing the overall CAD model of the body. The RSCS sets based on the point cloud to be registered and the RSCS set based on the target point cloud are constructed respectively, and the sets are preprocessed; the Spinnet point cloud neural network is trained using the triple loss function, and the descriptor is extracted using this network to obtain the descriptor set based on the point cloud to be registered and the descriptor set based on the target point cloud; the bidirectional matching is used to perform rough registration on each descriptor set, and the NICP algorithm based on correcting the normal vector direction is used for fine registration after rough registration; the registration error of the matching pair set after fine registration is obtained to reconstruct the RSCS set based on the point cloud to be registered and the RSCS set based on the target point cloud, and the accurate matching pair set is obtained through iteration; the present invention designs and implements a complete algorithm model for efficiently registering the partial body measurement data point cloud with the overall CAD ideal design model of the body from rough registration to fine registration by constructing a descriptor based on deep learning, solves the drawback that the descriptor established by traditional manual features does not have generalization, the constructed RSCS set overcomes the singularity problem of the classic single sphere coverage, and can greatly reduce the calculation amount compared with the traditional key point detection algorithm, significantly improve the efficiency, and in combination with the NICP algorithm based on correcting the normal vector direction, can comprehensively improve the accuracy and efficiency of the final matching. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is a flowchart of the steps of a registration method for partial body measurement data to the overall CAD model provided by the present invention;

[0025] Figure 2 It is a schematic structural diagram of the optimized Spinnet point cloud neural network in the embodiment of the present invention;

[0026] Figure 3 It is a schematic diagram of the final segmentation and pairing effect of the multi-scale RSCS in the embodiment of the present invention;

[0027] Figure 4 It is an effect diagram of rough registration in the embodiment of the present invention;

[0028] Figure 5 It is an effect diagram of fine registration in the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0029] The following describes the specific embodiments of the present invention to facilitate those skilled in the art to understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions and creations using the concept of the present invention are within the scope of protection.

[0030] As Figure 1 shown, an embodiment of the present invention provides a method for registering partial measurement data of a vehicle body to an overall CAD model, including the following sub-steps:

[0031] S1. Discretize the overall CAD model of the vehicle body to obtain the target point cloud of the vehicle body;

[0032] S2. Collect partial measurement data of the vehicle body to obtain the point cloud to be registered;

[0033] S3. Respectively construct an RSCS set based on the point cloud to be registered and an RSCS set based on the target point cloud according to the point cloud to be registered and the target point cloud;

[0034] In the embodiment of the present invention, the operation of key point detection is replaced by RSCS (Random sphere cover set), and RSCS is a set of randomly covered spheres.

[0035] Preferably, step S3 is specifically:

[0036] Randomly select a point in the point cloud to be registered as the center of the sphere of the sphere, and preset a radius, and construct a single point cloud sphere with the point cloud inside the sphere; select any point outside the single point cloud sphere as the center of the sphere of the sphere, and construct the remaining single point cloud spheres until the point cloud to be processed is traversed to obtain the RSCS set based on the point cloud to be registered; similarly, traverse the target point cloud to obtain the RSCS set based on the target point cloud, where the expression of the preset radius is:

[0037]

[0038] where Radius is the preset radius value, π is a constant value, m is the number of Patches for matching, that is, the number of single point cloud spheres. In a 6-degree-of-freedom space, 3 matches need to be found to calculate a unique rotation matrix, so the first m is set to 3 and a coverage rate of 0.7 is achieved. V is the volume of the minimum bounding sphere of the point cloud to be registered.

[0039] In the embodiments of the present invention, a point is randomly selected as the center of the sphere, and the radius R is set. Calculate the points around the center of the sphere whose distance from the center of the sphere is less than R. These point sets form a single Patch (point cloud sphere). Continue to randomly select a point that does not belong to any Patch as the center of the sphere and repeat the above steps until all the point clouds are covered by a certain number of Patches. These Patches form an RSCS set; taking the point cloud obtained by discretizing the ideal design model of the overall CAD of the vehicle body as the target point cloud and the local measurement point cloud of the real vehicle body as the point cloud to be registered, perform an RSCS generation operation on these two point clouds respectively, and obtain the RSCS sets of the target point cloud and the point cloud to be registered respectively. Let the RSCS of the target point cloud be P = {p1, p2,..., p n}, and the RSCS set of the point cloud to be registered is Q = {q1, q2,..., q n}, where each p or q in the set represents a Patch.

[0040] And set the initial radius as R, and make a rough estimate according to the deformation formula of the sphere volume calculation formula. There are a total of 4 candidate radii, and m are 3, 6, 12, and 24 respectively. Finally, a multi-scale RSCS is formed, which can increase the diversity of the sphere coverage point set.

[0041] S4. Preprocess the RSCS set based on the point cloud to be registered and the RSCS set based on the target point cloud respectively;

[0042] Preferably, step S4 is specifically:

[0043] Preprocess the RSCS set based on the point cloud to be registered and the RSCS set based on the target point cloud according to the preset density of a single point cloud sphere. If the density of a single point cloud sphere in the RSCS set is less than the preset density of a single point cloud sphere, then delete this single point cloud sphere; obtain the preprocessed RSCS set based on the point cloud to be registered and the preprocessed RSCS set based on the target point cloud.

[0044] In the embodiments of the present invention, after the RSCS set is formed, it is necessary to filter according to the density of each Patch and discard some Patches with smaller densities. Here, directly discard the Patches that cover less than 10 points among some Patches, and then obtain the preprocessed RSCS set based on the point cloud to be registered and the preprocessed RSCS set based on the target point cloud.

[0045] S5. Train the Spinnet point cloud neural network using the triple loss function to obtain an optimized Spinnet point cloud neural network;

[0046] Preferably, the triple loss function in step S5 is expressed as:

[0047]

[0048] Among them, L is the value of the triple loss function, is the distance between single point cloud spheres in the same class, is the distance between single point cloud spheres of different classes, is the anchor, is the positive sample, is the negative sample, where the anchor and the positive sample are samples of the same class, N is the number of training samples, and α is a constant not less than 0.

[0049] In the embodiments of the present invention, when training the network, the pre-trained network parameters on the 3DMATCH and KITTI datasets are loaded in advance. Finally, 1000 groups of Patch pairs are made on the ABC (A Big CAD Model Dataset For Geometric Deep Learning) dataset. Each Patch is divided by RSCS, and then the in-class block samples and the out-of-class block samples are manually labeled. Finally, the parameter training of the entire network is completed. After training, each Patch in the RSCS of the target point cloud RSCS and the RSCS of the point cloud to be registered is sent into the network to form the descriptor of each Patch. Then the Patch descriptor sets of the two RSCSs are respectively D = {d1, d2,..., d n} and B = {b1, b2,..., b n}, where the dimension of each descriptor is a 32-dimensional vector, and it is necessary to ensure that the distance between single point cloud spheres in the same class is as small as possible. The optimized Spinnet point cloud neural network structure obtained is as Figure 2 shown.

[0050] S6. Use the optimized Spinnet point cloud neural network to extract the descriptors in the RSCS set based on the point cloud to be registered and the RSCS set based on the target point cloud after preprocessing, and respectively obtain the descriptor set based on the point cloud to be registered and the descriptor set based on the target point cloud;

[0051] S7. Use bidirectional matching to perform rough registration on each descriptor set in step S6 to obtain a set of rough-registered matching pairs;

[0052] In the embodiment of the present invention, in the initial matching stage, a descriptor of a point cloud to be registered (the descriptor is a 32-dimensional vector output by the Patch in RSCS through the spinnet point cloud neural network) may find multiple corresponding descriptors in the descriptor set of the target point cloud, and there may even be a situation where the similarity of non-corresponding points is greater than that of matching points. Therefore, two-way matching is adopted for the matching of descriptors to optimize the above problems. The two-way matching specifically is as follows: find the Patch pair with the smallest Euclidean distance between the descriptors of P in Q (P and Q respectively refer to the target point cloud and the point cloud to be registered, and the Euclidean distance calculates the distance between the descriptors formed by the Patch of P and the Patch of Q through the neural network), and then find the Patch in P with the smallest Euclidean distance of its descriptor in the reverse direction of this Patch in Q at this time. If the Patch found in P is exactly the same, then retain this Patch pair, otherwise discard it. Finally, we will obtain m Patch pairs to form an initial transformation matrix.

[0053] Preferably, step S7 specifically includes the following sub-steps:

[0054] A1. Calculate the Euclidean distance between each descriptor in the descriptor set based on the point cloud to be registered and the descriptor set based on the target point cloud;

[0055] A2. Select the descriptor pair corresponding to the smallest Euclidean distance according to the Euclidean distance of each descriptor to obtain a matching pair;

[0056] In practice, find the descriptor pair corresponding to the smallest Euclidean distance among the descriptors of the target point cloud P in the point cloud Q to be registered, and then find the Patch in the target point cloud P with the smallest Euclidean distance of its descriptor in the reverse direction of this Patch in the point cloud Q to be registered at this time. If the Patch found in the target point cloud P is exactly the same, then retain this Patch pair, otherwise discard it, and finally obtain a matching pair.

[0057] A3. Construct a transformation matrix and combine the matching pair to obtain a set of matching pairs after rough registration.

[0058] In the embodiment of the present invention, the optimal block diagram based on the set of matching pairs after rough registration is as Figure 3 shown.

[0059] Preferably, step A3 is specifically as follows:

[0060] B1. Construct a set of central points of a single point cloud sphere according to the RSCS set based on the point cloud to be registered and the RSCS set based on the target point cloud, and calculate the centroid according to the set of central points;

[0061] B2. Calculate the covariance matrix according to the centroid, where the calculation formula of the covariance matrix is expressed as:

[0062]

[0063] where h is the covariance matrix, is the centroid value of the point set to be registered, is the centroid value of the target point set, x i is the i-th point cloud in the point set to be registered, y i is the i-th point cloud in the target point set, n is the number of point clouds, (.) T is the transpose;

[0064] B3. Perform SVD decomposition on the covariance matrix, and obtain the rotation matrix according to the left singular vector and right singular vector after decomposition;

[0065] In the embodiment of the present invention, perform SVD decomposition on the matrix h, multiply its left singular vector U and right singular vector V to obtain the rotation matrix R, and the rotation matrix is expressed as: R = UV T .

[0066] B4. Construct the translation matrix according to the rotation matrix, and according to the rotation matrix and the translation matrix, perform the rotation and translation of the point cloud to be registered to the target point cloud in the matching pair, and obtain the set of matching pairs after initial registration, where the rotation matrix and the translation matrix are used as transformation matrices.

[0067] In the embodiment of the present invention, the translation matrix is expressed as:

[0068] In the embodiment of the present invention, after rough registration, the enlarged effect diagram of the local measurement data registered to the whole vehicle body, that is, the rough registration effect diagram is as Figure 4 shown.

[0069] In the embodiment of the present invention, construct the target optimization formula according to the initialized translation matrix and the initialized rotation matrix, and obtain the translation matrix and the rotation matrix according to the target optimization formula; where the target optimization formula is expressed as:

[0070]

[0071] where T is the translation matrix, R is the rotation matrix, the translation matrix and the rotation matrix are the transformation matrices, x i is the i-th point cloud in the point set to be registered, y i is the i-th point cloud in the target point set, ||.|| 2 is the square of the vector norm, n is the number of point clouds in the point set to be registered or the target point set;

[0072] For this objective optimization formula, this least squares problem can be solved by singular value decomposition (SVD). First, a point set is formed by the center points of each Patch in each RSCS and the center point set of each Patch to calculate the centroids of the two point sets. Then, a covariance matrix h is calculated by combining the centroids of the two point sets. Next, the matrix h is decomposed by SVD, and the rotation matrix R is obtained by multiplying the left singular vector U and the right singular vector V.

[0073] S8. Use the NICP algorithm based on correcting the normal vector direction to perform fine registration on the set of matching pairs after coarse registration;

[0074] In the embodiment of the present invention, after obtaining the initial transformation matrix, the corresponding point sets need to be finely registered to further reduce the matching error between the local measurement point cloud and the overall target point cloud. This step is based on the NICP algorithm. The traditional ICP algorithm performs rotation and translation by reducing the distance between the nearest point sets of two point clouds. It defaults that the nearest points are corresponding points and continuously iterates this process until the distance between the nearest points is less than a certain set threshold. When the local measurement point cloud contains more noise and different density distributions, the nearest points or the initially calculated corresponding points are not necessarily their corresponding points. Therefore, simply taking the nearest points as corresponding points has poor matching accuracy. Thus, the traditional point-to-point distance is no longer used here, but the NICP algorithm based on correcting the normal vector direction is adopted. It not only requires the distances to be close but also requires the normal vectors and curvatures near the points to be similar, having a certain degree of semantic information. The NICP algorithm based on correcting the normal vector direction is specifically as follows: First, to minimize the distance between the normal vectors, the normal vectors between the points need to be calculated. The calculation formula for the normal vector is also obtained by performing SVD decomposition on the covariance matrix.

[0075] Preferably, step S8 specifically includes the following sub-steps:

[0076] C1. Calculate the eigenvectors according to the set of matching pairs after coarse registration, and select the eigenvector corresponding to the minimum eigenvalue as the normal vector;

[0077] Preferably, step C1 is specifically:

[0078] Construct a covariance matrix according to the set of matching pairs after coarse registration, and perform SVD decomposition on the covariance to obtain the normal vectors between each matching point, where the covariance is expressed as:

[0079]

[0080] where h* is the covariance matrix based on the set of matching pairs after coarse registration, (.) T is the transpose, and p iIt is a neighborhood point of a point in the point set to be registered or the target point set in the set of matching pairs. It is a point in the point set to be registered or the target point set in the set of matching pairs, and n is the number of point clouds in the point set to be registered or the target point set.

[0081] In the embodiments of the present invention, three eigenvalues can be calculated corresponding to three eigenvectors respectively, and the eigenvector corresponding to the minimum eigenvalue is the normal vector.

[0082] C2. Quantify the relationship between the normal vector and the neighborhood points, and correct the direction of the normal vector according to the relationship to obtain the normal vector after correcting the direction.

[0083] Preferably, step C2 is specifically:

[0084] Quantify the relationship between the normal vector and the neighborhood points, and correct the direction of the normal vector according to the relationship. The quantization process of the relationship between the normal vector and the neighborhood points is expressed as:

[0085]

[0086] Among them, r is the relationship between the quantized normal vector and the neighborhood points, n i is the i-th neighborhood normal vector, k is the number of neighborhood points, v is the normal vector, and · is the dot product operator;

[0087] If the relationship between the quantized normal vector and the neighborhood points is less than the preset threshold, the direction of the normal vector is reversed, that is, a negative sign is added, otherwise the current direction of the normal vector remains unchanged, that is, the positive and negative signs of the current normal vector are ensured to remain unchanged; the normal vector after correcting the direction is obtained.

[0088] In the embodiments of the present invention, the direction of the normal vector in the set of matching pairs after fine registration is uncertain as to positive or negative in actual situations. This will cause a huge difference in the included angles of the normal vectors calculated by similar attribute points, resulting in the loss of some important point pairs in the end. Therefore, the direction is corrected by quantifying the relationship between the normal vector and the neighborhood points. Specifically: if the result of the relationship r between the quantized normal vector and the neighborhood points is greater than or equal to 0, then the positive and negative signs of the normal vector remain unchanged, otherwise a negative sign is added.

[0089] C3. Calculate the curvature according to the eigenvectors between the matching points, and screen the matching point pairs that are consistent with the direction of the normal vector and close in distance from the set of matching pairs after rough registration according to the preset curvature value. The calculation formula of the curvature is expressed as:

[0090]

[0091] Among them, α is the curvature, λ1, λ2 and λ3 are the respective eigenvectors, and λ3 is the eigenvector corresponding to the minimum eigenvalue;

[0092] In the embodiments of the present invention, the role of curvature is mainly to eliminate some irrelevant point pairs.

[0093] C4. Construct a minimization objective formula, and use the minimization objective formula to minimize the matching point pairs that are consistent with the normal vector direction and have a similar distance, so as to obtain a set of matching pairs after fine registration.

[0094] In the embodiments of the present invention, the role of the minimization objective formula is to minimize the distance between two center points and the included angle of the normal vectors.

[0095] Preferably, the minimization objective formula in step C4 is expressed as:

[0096]

[0097] Wherein, is the set formed by connecting the center points and normal vectors of the single point cloud spheres that have been mutually matched in the target point cloud, t is the translation matrix, and p i is the set of points to be registered or the target point set, E(T) is the error of the target formula, is the set composed of points and normal vectors in the set of points to be registered or the target point set in the matching pair, and n i is the normal vector in the set of points to be registered or the target point set, and r is the relationship between the quantized normal vector and the domain points.

[0098] S9. Obtain the registration error of the set of matching pairs after fine registration, and return to step S3 according to the registration error to reconstruct the RSCS set based on the point cloud to be registered and the RSCS set based on the target point cloud until the preset registration error threshold is met.

[0099] In the embodiments of the present invention, there is always a problem that the characteristics of the covered point set of a single-scale Patch are not obvious. Therefore, after each complete registration process, the registration error needs to be recorded, and then the radius R is reselected to perform an RSCS operation on the point cloud to be registered and the target point cloud until the preset registration error threshold is met. In this embodiment, it loops four times, and there are 4 candidate radii, where the values of m are 3, 6, 12, and 24 respectively. Finally, a multi-scale RSCS is obtained to increase the diversity of the covered point set of the sphere. During the selection of R, it is necessary to make the characteristics of the point set covered by R as obvious as possible so that higher accuracy can be obtained in the subsequent coarse registration; the effect of fine registration is as Figure 5 shown.

[0100] The present invention is described with reference to the flowcharts and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and combinations of flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to the processors of general purpose computers, special purpose computers, embedded processors, or other programmable data processing devices to produce a machine, such that the instructions executed by the processors of the computer or other programmable data processing devices produce means for implementing the functions specified in one flow Figure 1 one flow or multiple flows and / or blocks Figure 1 or in multiple blocks.

[0101] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a particular manner, such that the instructions stored in the computer-readable memory produce a manufacture including instruction means for implementing the functions specified in one flow Figure 1 one flow or multiple flows and / or blocks Figure 1 or in multiple blocks.

[0102] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operational steps are performed on the computer or other programmable device to produce a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one flow Figure 1 one flow or multiple flows and / or blocks Figure 1 or in multiple blocks.

[0103] Specific embodiments are applied in the present invention to elaborate on the principles and implementation manners of the present invention. The description of the above embodiments is only used to help understand the method and its core idea of the present invention; at the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation manners and application scopes. In summary, the content of this specification should not be construed as a limitation to the present invention.

[0104] Those of ordinary skill in the art will realize that the embodiments described herein are for helping readers understand the principles of the present invention, and it should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those of ordinary skill in the art can make various other specific deformations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed in the present invention, and these deformations and combinations are still within the protection scope of the present invention.

Claims

1. A registration method for registering partial measurement data of a vehicle body to an overall CAD model, characterized in that, It includes the following steps: S1. Discretize the overall CAD model of the vehicle body to obtain the target point cloud of the vehicle body; S2. Collect partial measurement data of the vehicle body to obtain the point cloud to be registered; S3. Respectively construct an RSCS random sphere coverage set based on the point cloud to be registered and an RSCS random sphere coverage set based on the target point cloud according to the point cloud to be registered and the target point cloud. Specifically: Randomly select a point in the point cloud to be registered as the center of the sphere of the sphere, and preset the radius, and construct a single point cloud sphere with the point cloud inside the sphere; select any point outside the single point cloud sphere as the center of the sphere of the sphere, and construct the remaining single point cloud spheres until the point cloud to be processed is traversed to obtain the RSCS set based on the point cloud to be registered; similarly, traverse the target point cloud to obtain the RSCS set based on the target point cloud, where the expression of the preset radius is: wherein, is a preset radius value, is a constant value, m is the number of Patches for matching, that is, the number of single point cloud spheres, V is the volume of the minimum enclosing sphere of the point cloud to be registered; S4. Respectively preprocess the RSCS random sphere coverage set based on the point cloud to be registered and the RSCS random sphere coverage set based on the target point cloud; S5. Train the Spinnet point cloud neural network using the triple loss function to obtain the optimized Spinnet point cloud neural network; S6. Use the optimized Spinnet point cloud neural network to extract the descriptors in the preprocessed RSCS random sphere coverage set based on the point cloud to be registered and the RSCS random sphere coverage set based on the target point cloud respectively, and respectively obtain the descriptor set based on the point cloud to be registered and the descriptor set based on the target point cloud; S7. Use bidirectional matching to roughly register each descriptor set in step S6 to obtain the set of matching pairs after rough registration; S8. Use the NICP algorithm based on correcting the normal vector direction to finely register the set of matching pairs after rough registration; S9. Obtain the registration error of the set of matching pairs after fine registration, and return to step S3 according to the registration error to reconstruct the RSCS random sphere coverage set based on the point cloud to be registered and the RSCS random sphere coverage set based on the target point cloud until the preset registration error threshold is met.

2. The registration method of the body part measurement data to the overall CAD model according to claim 1, characterized in that, Step S4 is specifically: Preprocess the RSCS set based on the point cloud to be registered and the RSCS set based on the target point cloud according to the preset density of a single point cloud sphere. If the density of a single point cloud sphere in the RSCS set is less than the preset density of a single point cloud sphere, then delete the single point cloud sphere; Obtain the preprocessed RSCS set based on the point cloud to be registered and the preprocessed RSCS set based on the target point cloud.

3. The registration method of the body part measurement data to the overall CAD model according to claim 1, characterized in that The triple loss function in step S5 is expressed as: Among them, L is the value of the triple loss function, is the distance between single point cloud spheres in the same class, is the distance between single point cloud spheres of different classes, is the anchor, is the positive sample, is the negative sample, where the anchor and the positive sample are samples of the same class, N is the number of training samples, is a constant.

4. The method for registering the measured data of the vehicle body part to the overall CAD model according to claim 1, characterized in that, Step S7 specifically includes the following sub-steps: A1. Calculate the Euclidean distance between each descriptor in the descriptor set based on the target point cloud and the descriptor set based on the point cloud to be registered; A2. Select the descriptor pair corresponding to the minimum Euclidean distance according to the Euclidean distance of each descriptor to obtain the matching pair; A3. Construct a transformation matrix and combine the matching pair to obtain the set of matching pairs after rough registration.

5. The method for registering the body part measurement data to the overall CAD model according to claim 4, characterized in that, Step A3 is specifically: B1. Construct the central point set of a single point cloud sphere according to the RSCS set based on the point cloud to be registered and the RSCS set based on the target point cloud, and calculate the centroid according to the central point set; B2. Calculate the covariance matrix according to the centroid, where the calculation formula of the covariance matrix is expressed as: Among them, h is the covariance matrix, is the centroid value of the point set to be registered, is the centroid value of the target point set, x i is the i th point cloud in the point set to be registered, y i is the i th point cloud in the target point set, n is the number of point clouds, is the transpose; B3. Perform SVD decomposition on the covariance matrix, and obtain the rotation matrix according to the left singular vector and right singular vector after decomposition; B4. Construct the translation matrix according to the rotation matrix, and according to the rotation matrix and the translation matrix, perform the rotation and translation of the point cloud to be registered to the target point cloud in the matching pair, and obtain the set of matching pairs after initial registration, where the rotation matrix and the translation matrix are used as transformation matrices.

6. The method for registering the body part measurement data to the overall CAD model according to claim 1, wherein Step S8 specifically includes the following sub-steps: C1. Calculate the eigenvector according to the set of matching pairs after rough registration, and select the eigenvector corresponding to the minimum eigenvalue as the normal vector; C2. Quantify the relationship between the normal vector and the neighborhood points, and correct the direction of the normal vector according to the relationship to obtain the normal vector after correcting the direction; C3. Calculate the curvature according to the eigenvectors between each pair of matching points, and screen the matching point pairs that are consistent with the direction of the normal vector and close in distance from the set of matching pairs after rough registration according to the preset curvature value, where the calculation formula of the curvature is expressed as: Among them, is the curvature, , and are respectively the eigenvectors of each feature, is the eigenvector corresponding to the minimum eigenvalue; C4. Construct the minimization objective formula, and use the minimization objective formula to minimize the matching point pairs with the same normal vector direction and close distance to obtain the set of matching pairs after fine registration.

7. The method for registering the measured data of the vehicle body part to the overall CAD model according to claim 6, characterized in that, Step C1 is specifically: Construct the covariance matrix according to the set of matching pairs after rough registration, and perform SVD decomposition on the covariance to obtain the normal vectors between each pair of matching points, where the covariance is expressed as: Among them, is the covariance matrix based on the set of matching pairs after coarse registration, is the transpose, is a neighborhood point of a point in the point set to be registered or the target point set in the set of matching pairs, is a point in the point set to be registered or the target point set in the set of matching pairs, n is the number of point clouds in the point set to be registered or the target point set.

8. The method for registering the measured data of the vehicle body part to the overall CAD model according to claim 6, wherein Step C2 is specifically: Quantify the relationship between the normal vector and the neighborhood points, and correct the direction of the normal vector according to the relationship. The quantization process of the relationship between the normal vector and the neighborhood points is expressed as: Among them, r is the relationship between the quantified normal vector and the domain points, n i is the i th neighborhood normal vector, k is the number of neighborhood points, v is the normal vector, is the dot product operator; If the relationship between the quantified normal vector and the neighborhood points is less than the preset threshold, the direction of the normal vector is reversed, that is, a negative sign is added, otherwise the current direction of the normal vector remains unchanged, that is, the positive and negative signs of the current normal vector are ensured to remain unchanged; obtain the normal vector after correcting the direction.

9. The method for registering the measured data of the vehicle body part to the overall CAD model according to claim 6, characterized in that, The minimization objective formula in step C4 is expressed as: Among them, is a set formed by connecting the center points and normal vectors of individual point cloud spheres that have been mutually matched in the target point cloud, t is a translation matrix, p i is the point set to be registered or the target point set, is the error of the target formula, is a set composed of points and normal vectors in the point set to be registered or the target point set in the matching pair, n i is the normal vector in the point set to be registered or the target point set, r is the relationship between the quantized normal vector and the neighborhood points.

Images