Registration method and device based on probability distribution distance feature description, and storage medium

Abstract

The application relates to the technical field of data processing, and discloses a registration method and device based on probability distribution distance feature description and a storage medium. The method comprises the following steps: generating a source point cloud hybrid probability distribution graph according to a source point cloud, and generating a target point cloud hybrid probability distribution graph according to a target point cloud; performing nonlinear calculation on the source point cloud hybrid probability distribution graph and the target point cloud hybrid probability distribution graph respectively, and taking the calculation results as source point cloud features and target point cloud features; generating a point cloud corresponding relation matrix according to the source point cloud features and the target point cloud features; and determining a registration result of the source point cloud to the target point cloud according to the point cloud corresponding relation matrix. In the foregoing manner, the point cloud corresponding relation matrix is generated according to the source point cloud features and the target point cloud features, and then the registration result of the source point cloud to the target point cloud is determined according to the point cloud corresponding relation matrix, so that the accuracy, stability and generalization ability of the registered point cloud can be effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a registration method based on probability distribution distance feature description, a device and a storage medium. BACKGROUND

[0002] With the development and promotion of depth sensing technology and the metaverse industry, three-dimensional models have been widely used in entertainment, consumption and other daily life fields, and are expected to become the hub of future world digital information flow. The three-dimensional point cloud data based on depth images is usually sampled from multiple perspectives, often accompanied by noise, environmental occlusion and target overlap. Therefore, point cloud registration is the most important data processing link before three-dimensional imaging and its application. The current registration technology is based on the Iterative Closest Point (ICP) algorithm. The point cloud registration needs to meet the requirements of high overlap rate, no noise and data integrity, which restricts the popularity of ICP. In addition, the registration algorithm based on probability model converts the registration problem into solving the maximum estimation of probability density, which is good at noise robustness but prone to registration failure in the case of data missing. The rough registration based on local feature point extraction and feature description, and then the iterative calculation for fine registration, can effectively reduce the calculation cost and is good at data overlap rate and integrity robustness, but is easily disturbed by noise and the point feature lacks global characteristics, which can cause mismatching relationship and finally cause a series of problems such as low accuracy, poor stability and weak generalization ability of the registered point cloud.

[0003] The above content is only used to assist in understanding the technical solutions of the present application and does not represent the acknowledgement of the above content as prior art. SUMMARY

[0004] The main purpose of the present application is to provide a registration method based on probability distribution distance feature description, a device and a storage medium, which aims to solve the technical problems of low accuracy, poor stability and weak generalization ability of the registered point cloud in the prior art.

[0005] To achieve the above purpose, the present application provides a registration method based on probability distribution distance feature description, which comprises the following steps:

[0006] Generating a source point cloud hybrid probability distribution map according to the source point cloud, and generating a target point cloud hybrid probability distribution map according to the target point cloud;

[0007] Respectively performing nonlinear calculation on the source point cloud hybrid probability distribution map and the target point cloud hybrid probability distribution map, and taking the calculation results as source point cloud features and target point cloud features respectively;

[0008] generate a point cloud correspondence relationship matrix according to the source point cloud feature and the target point cloud feature;

[0009] determine a registration result of the source point cloud to the target point cloud according to the point cloud correspondence relationship matrix.

[0010] Optionally, the generating a source point cloud mixture probability distribution map according to the source point cloud and a target point cloud mixture probability distribution map according to the target point cloud comprises:

[0011] obtain an initial source point cloud and an initial point cloud, calibrate the initial source point cloud and the initial point cloud;

[0012] respectively perform quality enhancement on the calibrated initial source point cloud and the initial point cloud to obtain a source point cloud and a target point cloud;

[0013] generate a source point cloud model according to the source point cloud and a target point cloud model according to the target point cloud;

[0014] respectively perform semantic segmentation on the source point cloud model and the target point cloud model by a target proximity algorithm to obtain a source point cloud segmentation result and a target point cloud segmentation result;

[0015] generate a source point cloud mixture probability distribution map according to the source point cloud segmentation result and a target point cloud mixture probability distribution map according to the target point cloud segmentation result.

[0016] Optionally, the generating a source point cloud mixture probability distribution map according to the source point cloud segmentation result and a target point cloud mixture probability distribution map according to the target point cloud segmentation result comprises:

[0017] perform calculation on the source point cloud segmentation result to obtain a source point cloud segmentation part Gaussian distribution and a neighborhood Gaussian distribution of each point of the source point cloud;

[0018] splice the neighborhood Gaussian distribution of each point of the source point cloud and the source point cloud segmentation part Gaussian distribution to which each point of the source point cloud belongs to obtain a source point cloud mixture Gaussian distribution;

[0019] generate a source point cloud mixture probability distribution map according to the source point cloud mixture Gaussian distribution and a preset graph structure;

[0020] perform calculation on the target point cloud segmentation result to obtain a target point cloud segmentation part Gaussian distribution and a neighborhood Gaussian distribution of each point of the target point cloud;

[0021] splice the neighborhood Gaussian distribution of each point of the target point cloud and the target point cloud segmentation part Gaussian distribution to which each point of the target point cloud belongs to obtain a target point cloud mixture Gaussian distribution;

[0022] generate a target point cloud mixture probability distribution map according to the target point cloud mixture Gaussian distribution and a preset graph structure.

[0023] Optionally, the nonlinear calculation is performed on the source point cloud mixture probability distribution map and the target point cloud mixture probability distribution map respectively, and the calculation results are used as the source point cloud feature and the target point cloud feature, including:

[0024] calculating a first probability distribution distance between the target vertex in the source point cloud mixture probability distribution map and each vertex in the target point cloud mixture probability distribution map;

[0025] generating a source vertex vector according to the first probability distribution distance and the number of source vertices;

[0026] calculating a second probability distribution distance between the target vertex in the source point cloud mixture probability distribution map and the source point cloud segmentation part;

[0027] generating a first vertex matrix according to the second probability distribution distance and the number of source vertices;

[0028] calculating a third probability distribution distance between the target vertex in the source point cloud mixture probability distribution map and the target point cloud segmentation part;

[0029] generating a second vertex matrix according to the third probability distribution distance and the number of source vertices;

[0030] extending the target vertex vector to the first vertex matrix and the second vertex matrix respectively;

[0031] generating a source point cloud feature matrix according to the extended first vertex matrix, the second vertex matrix, and the target vertex vector;

[0032] calculating a fourth probability distribution distance between the target vertex in the target point cloud mixture probability distribution map and each vertex in the source point cloud mixture probability distribution map;

[0033] generating a target vertex vector according to the fourth probability distribution distance and the number of target vertices;

[0034] calculating a fifth probability distribution distance between the target vertex in the target point cloud mixture probability distribution map and the target point cloud segmentation part;

[0035] generating a third vertex matrix according to the fifth probability distribution distance and the number of target vertices;

[0036] calculating a sixth probability distribution distance between the target vertex in the target point cloud mixture probability distribution map and the source point cloud segmentation part;

[0037] generating a fourth vertex matrix according to the sixth probability distribution distance and the number of target vertices;

[0038] extending the target vertex vector to the third vertex matrix and the fourth vertex matrix respectively;

[0039] generating a target point cloud feature matrix according to the extended third vertex matrix and the fourth vertex matrix and the target vertex vector;

[0040] obtaining source point cloud features and target point cloud features according to the source point cloud feature matrix and the target point cloud feature matrix.

[0041] Optionally, the obtaining source point cloud features and target point cloud features according to the source point cloud feature matrix and the target point cloud feature matrix comprises:

[0042] respectively performing multi-layer perception machine calculation on the source point cloud feature matrix and the target point cloud feature matrix to obtain source point cloud feature description matrices and target point cloud feature description matrices;

[0043] respectively performing pretreatment on the source point cloud feature description matrices and the target point cloud feature description matrices to obtain source point cloud feature descriptions and target point cloud feature descriptions;

[0044] obtaining source point cloud features and target point cloud features according to the source point cloud feature descriptions and the target point cloud feature descriptions.

[0045] Optionally, the generating a point cloud correspondence relationship matrix according to the source point cloud features and the target point cloud features comprises:

[0046] performing dot product calculation on the source point cloud features and the target point cloud features to obtain a plurality of feature similarities;

[0047] constructing a feature similarity matrix according to the plurality of feature similarities;

[0048] performing row and column calculation on the feature similarity matrix by a sinkhorn algorithm to obtain a point cloud correspondence relationship matrix.

[0049] Optionally, the determining a registration result of the source point cloud to the target point cloud according to the point cloud correspondence relationship matrix comprises:

[0050] obtaining a matched point cloud pair according to the point cloud correspondence relationship matrix;

[0051] performing calculation on the matched point cloud pair by a SVD algorithm according to a plurality of dimensions to obtain a rotation matrix and a translation vector;

[0052] constructing a rigid transformation relationship between the source point cloud and the target point cloud according to the rotation matrix and the translation vector;

[0053] performing rotation and / or translation on each point of the source point cloud according to the rigid transformation relationship to obtain the registration result of the source point cloud to the target point cloud.

[0054] In addition, to achieve the above object, the application further provides a registration device based on probability distribution distance feature description, which comprises:

[0055] a generating module, configured to generate a source point cloud hybrid probability distribution graph according to a source point cloud and a target point cloud hybrid probability distribution graph according to a target point cloud;

[0056] a calculating module, configured to perform nonlinear calculation on the source point cloud hybrid probability distribution graph and the target point cloud hybrid probability distribution graph respectively, and take the calculation results as source point cloud features and target point cloud features;

[0057] the generating module is further configured to generate a point cloud correspondence relationship matrix according to the source point cloud features and the target point cloud features;

[0058] a determining module, configured to determine a registration result of the source point cloud to the target point cloud according to the point cloud correspondence relationship matrix.

[0059] In addition, to achieve the above object, the application further provides a registration device based on probability distribution distance feature description, which comprises: a memory, a processor and a registration program based on probability distribution distance feature description stored on the memory and executable on the processor, and the registration program based on probability distribution distance feature description is configured to implement the registration method based on probability distribution distance feature description as described above.

[0060] In addition, to achieve the above object, the application further provides a storage medium, which stores a registration program based on probability distribution distance feature description, and the registration program based on probability distribution distance feature description implements the registration method based on probability distribution distance feature description as described above when executed by a processor.

[0061] The registration method based on probability distribution distance feature description provided by the application generates a source point cloud hybrid probability distribution graph according to a source point cloud and a target point cloud hybrid probability distribution graph according to a target point cloud, performs nonlinear calculation on the source point cloud hybrid probability distribution graph and the target point cloud hybrid probability distribution graph respectively, takes the calculation results as source point cloud features and target point cloud features, generates a point cloud correspondence relationship matrix according to the source point cloud features and the target point cloud features, and determines a registration result of the source point cloud to the target point cloud according to the point cloud correspondence relationship matrix, so as to effectively improve the accuracy, stability and generalization ability of the registration point cloud. BRIEF DESCRIPTION OF DRAWINGS

[0062] Figure 1 is a structural schematic diagram of a registration device based on probability distribution distance feature description of a hardware running environment related to an embodiment of the present application.

[0063] Figure 2 is a flowchart of a first embodiment of a registration method based on probability distribution distance feature description of the present application.

[0064] Figure 3 is a flowchart of a second embodiment of a registration method based on probability distribution distance feature description of the present application.

[0065] Figure 4 is a functional module schematic diagram of a first embodiment of a registration device based on probability distribution distance feature description of the present application.

[0066] The implementation, functional features and advantages of the present application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0067] It should be understood that the specific embodiments described herein are intended to be illustrative only and not limiting of the present application.

[0068] Reference Figure 1 , Figure 1 is a structural schematic diagram of a registration device based on probability distribution distance feature description of a hardware running environment related to an embodiment of the present application.

[0069] As shown in Figure 1 , the registration device based on probability distribution distance feature description can include a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to realize the connection and communication between these components. The user interface 1003 can include a display, an input unit such as a keyboard, and can also include a standard wired interface, a wireless interface. The network interface 1004 can optionally include a standard wired interface, a wireless interface (such as a wireless fidelity (Wi-Fi) interface). The memory 1005 can be a high-speed random access memory (RAM) memory, or a stable non-volatile memory (NVM), such as a disk memory. The memory 1005 can also be a storage device independent of the aforementioned processor 1001.

[0070] Those skilled in the art can understand that Figure 1The structure shown in the figure does not constitute a limitation on the registration device based on the probability distribution distance feature description, and can include more or fewer components than the figure, or combine certain components, or different component arrangements.

[0071] As shown in Figure 1 The memory 1005 as a storage medium can include an operating system, a network communication module, a user interface module, and a registration program based on the probability distribution distance feature description.

[0072] In Figure 1 The registration device based on the probability distribution distance feature description shown in the figure, the network interface 1004 is mainly used for data communication with the network integration platform workstation; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and the memory 1005 in the registration device based on the probability distribution distance feature description can be arranged in the registration device based on the probability distribution distance feature description, and the registration device based on the probability distribution distance feature description calls the registration program based on the probability distribution distance feature description stored in the memory 1005 through the processor 1001, and executes the registration method based on the probability distribution distance feature description provided by the embodiment of the application.

[0073] Based on the above hardware structure, the registration method based on the probability distribution distance feature description is proposed.

[0074] Referring to Figure 2 , Figure 2 The flowchart of the first embodiment of the registration method based on the probability distribution distance feature description of the application.

[0075] In the first embodiment, the registration method based on the probability distribution distance feature description comprises the following steps:

[0076] Step S10, generating a source point cloud hybrid probability distribution graph according to the source point cloud, and generating a target point cloud hybrid probability distribution graph according to the target point cloud.

[0077] It should be noted that the execution subject of the embodiment is a registration device based on the probability distribution distance feature description, and can also be other devices that can realize the same or similar functions, such as a point cloud registration controller, etc., and the embodiment does not limit this, and in the embodiment, the point cloud registration controller is taken as an example for description.

[0078] It should be understood that the source point cloud hybrid probability distribution graph refers to a graph data structure, and each vertex of the graph is the target point neighborhood probability distribution of the source point cloud and the probability distribution of the segmentation part to which the target point belongs. Similarly, the target point cloud hybrid probability distribution graph refers to a graph data structure, and each vertex of the graph is the target point neighborhood probability distribution of the target point cloud and the probability distribution of the segmentation part to which the target point belongs.

[0079] Step S20, respectively, the source point cloud mixture probability distribution and the target point cloud mixture probability distribution are calculated, and the calculation results are used as source point cloud features and target point cloud features.

[0080] Further, step S20 includes: calculating a first probability distribution distance between the target vertex in the source point cloud mixture probability distribution and each vertex in the target point cloud mixture probability distribution; generating a source vertex vector according to the first probability distribution distance and the number of source vertices; calculating a second probability distribution distance between the target vertex in the source point cloud mixture probability distribution and the source point cloud segmentation part; generating a first vertex matrix according to the second probability distribution distance and the number of source vertices; calculating a third probability distribution distance between the target vertex in the source point cloud mixture probability distribution and the target point cloud segmentation part; generating a second vertex matrix according to the third probability distribution distance and the number of source vertices; extending the target vertex vector to the first vertex matrix and the second vertex matrix respectively; generating a source point cloud feature matrix according to the extended first vertex matrix, the second vertex matrix and the target vertex vector; calculating a fourth probability distribution distance between the target vertex in the target point cloud mixture probability distribution and each vertex in the source point cloud mixture probability distribution; generating a target vertex vector according to the fourth probability distribution distance and the number of target vertices; calculating a fifth probability distribution distance between the target vertex in the target point cloud mixture probability distribution and the target point cloud segmentation part; generating a third vertex matrix according to the fifth probability distribution distance and the number of target vertices; calculating a sixth probability distribution distance between the target vertex in the target point cloud mixture probability distribution and the source point cloud segmentation part; generating a fourth vertex matrix according to the sixth probability distribution distance and the number of target vertices; extending the target vertex vector to the third vertex matrix and the fourth vertex matrix respectively; generating a target point cloud feature matrix according to the extended third vertex matrix and the fourth vertex matrix and the target vertex vector; obtaining source point cloud features and target point cloud features according to the source point cloud feature matrix and the target point cloud feature matrix.

[0081] It should be understood that the number of source point clouds and target point clouds can be N, after obtaining the source point cloud mixture probability distribution map and the target point cloud mixture probability distribution map, the probability distribution distance between the target vertex in the source point cloud mixture probability distribution map and each vertex in the target point cloud mixture probability distribution map is calculated, that is, the first probability distribution distance, then the first probability distribution distance vector is subjected to a SUM operation to obtain a one-dimensional feature of the target vertex, then N vertices form an N*1 vector; the second probability distribution distance between the target vertex and m source point cloud segmentation parts (in the order of small to large label) in the source point cloud mixture probability distribution map is calculated, then a 1*M dimensional vector is formed according to the second probability distribution distance and the number of source vertices, and N vertices form an N*M dimensional first vertex matrix; the third probability distribution distance between the target vertex and m target point cloud segmentation parts (in the order of small to large label) in the source point cloud mixture probability distribution map is calculated, then a 1*M dimensional vector is formed, and then an N*M dimensional second vertex matrix is formed according to the 1*M dimensional vector and N vertices; the N*1 target vertex vector is copied and spliced to expand to the N*M first vertex matrix and the second vertex matrix, and then the expanded first vertex matrix, the second vertex matrix and the target source vertex vector are spliced to form an N*3M dimensional source point cloud feature matrix, and the target point cloud feature matrix is generated in the same way, after obtaining the third vertex matrix and the fourth vertex matrix, the generated target vertex vector is expanded to the third vertex matrix and the fourth vertex matrix, and then the expanded third vertex matrix and the fourth vertex matrix and the target vertex vector generate an N*3M dimensional target point cloud feature matrix.

[0082] It should be noted that the probability distribution distance also becomes a difference measure between probability distributions, which is measured by EMD (Earth Mover's Distance) distance. The EMD distance quantifies the similarity between two probability distributions by minimizing the conversion cost between them. The Gaussian distribution of three-dimensional point cloud data is:

[0083]

[0084]

[0085] Where S represents the source point cloud, T represents the target point cloud, μ1 represents the mean of each dimension variable of the source point cloud, μ2 represents the mean of each dimension variable of the target point cloud, ∑1 represents the covariance matrix of each dimension variable of the source point cloud, ∑2 represents the covariance matrix of each dimension variable of the target point cloud, x represents a point in the point cloud set, and K represents the dimension, which is 3 here.

[0086] Then, the EMD distance can be described as:

[0087] EMD KL(S(x | m1, S1) || T(x | m2, S2)) = (m1 - m2) + Tr(S1 + S2 - 2(S1 1 / 2 S2 S1 1 / 2 ) 1 / 2 )

[0088] Wherein, the KL represents the KL divergence, generally refers to relative entropy.

[0089] Further, the source point cloud feature and the target point cloud feature are obtained according to the source point cloud feature matrix and the target point cloud feature matrix, including: performing multi-layer perception machine calculation on the source point cloud feature matrix and the target point cloud feature matrix respectively to obtain source point cloud feature description matrix and target point cloud feature description matrix; preprocessing the source point cloud feature description matrix and the target point cloud feature description matrix respectively to obtain source point cloud feature description and target point cloud feature description; obtaining the source point cloud feature and the target point cloud feature according to the source point cloud feature description and the target point cloud feature description.

[0090] It can be understood that after obtaining the source point cloud feature matrix and the target point cloud feature matrix, the same number of multi-layer perception machine calculations are performed on the source point cloud feature matrix and the target point cloud feature matrix to obtain the source point cloud feature description matrix and the target point cloud feature description matrix, and the dimensions of the source point cloud feature description matrix and the target point cloud feature description matrix are both N*1024, then the source point cloud feature description matrix and the target point cloud feature description matrix are aggregated to obtain a feature description with strong expression ability, i.e. the source point cloud feature description and the target point cloud feature description.

[0091] It should be noted that the feature description matrix aggregation can be realized in combination with the attention mechanism, and the self-attention mechanism and the cross-attention mechanism are used for intra-graph convolution and cross-graph convolution to effectively aggregate the original feature information into the final feature description matrix.

[0092] Step S30, generating a point cloud correspondence relationship matrix according to the source point cloud feature and the target point cloud feature.

[0093] It should be understood that the point cloud correspondence relationship matrix refers to the relationship matrix between the source point cloud and the target point cloud, and the point cloud correspondence relationship matrix is obtained by alternately performing row and column calculations by the sinkhorn algorithm.

[0094] Further, step S30 includes: performing dot product calculation on the source point cloud feature and the target point cloud feature to obtain a plurality of feature similarities; constructing a feature similarity matrix according to the plurality of feature similarities; and obtaining the point cloud correspondence relationship matrix by performing row and column calculations on the feature similarity matrix by the sinkhorn algorithm.

[0095] It can be understood that the feature similarity refers to the similarity between the source point cloud feature and the target point cloud feature, the number of the feature similarity is multiple, after the point product calculates a plurality of numbers of the feature similarity, a feature similarity matrix is constructed, at this time, in order to improve the accuracy of the point cloud correspondence matrix, the graph matching problem is converted into a linear assignment problem for processing, that is, the feature similarity matrix is alternately calculated by the sinkhorn algorithm, so as to obtain the point cloud correspondence matrix, in addition, the point cloud correspondence matrix will be corrected according to the labels and marks of the segmentation module, when the labels of the segmentation parts where the source point cloud and the target point cloud are located are different, the corresponding weight between the two points is reduced to zero.

[0096] In step S40, the registration result of the source point cloud to the target point cloud is determined according to the point cloud correspondence matrix.

[0097] Further, in step S40, the matching point cloud pair is obtained according to the point cloud correspondence matrix, the rotation matrix and the translation vector are obtained by calculating the matching point cloud pair according to a plurality of dimensions by the SVD algorithm, the rigid transformation relationship between the source point cloud and the target point cloud is constructed according to the rotation matrix and the translation vector, and the registration result of the source point cloud to the target point cloud is obtained by rotating and / or translating each point of the source point cloud according to the rigid transformation relationship.

[0098] It should be understood that the matching point cloud pair refers to the matching point cloud pair between the source point cloud and the target point cloud, and the plurality of dimensions refer to the dimensions for calculating the matching point cloud pair, which include but are not limited to rotation dimensions and translation dimensions, that is, the matching point cloud pair is calculated by rotation and translation, and then the rigid transformation relationship between the source point cloud and the target point cloud is constructed according to the rotation matrix and the translation vector, and then each point in the source point cloud is rotated and / or translated according to the rigid transformation relationship, to obtain the registration result of the source point cloud to the target point cloud.

[0099] In this embodiment, the source point cloud hybrid probability distribution map is generated according to the source point cloud, and the target point cloud hybrid probability distribution map is generated according to the target point cloud; the source point cloud feature and the target point cloud feature are obtained by performing nonlinear calculation on the source point cloud hybrid probability distribution map and the target point cloud hybrid probability distribution map respectively; the point cloud correspondence matrix is generated according to the source point cloud feature and the target point cloud feature; the registration result of the source point cloud to the target point cloud is determined according to the point cloud correspondence matrix; in the above manner, the point cloud correspondence matrix is generated according to the source point cloud feature and the target point cloud feature, and then the registration result of the source point cloud to the target point cloud is determined according to the point cloud correspondence matrix, so as to effectively improve the accuracy, stability and generalization ability of the registered point cloud.

[0100] In an embodiment, as Figure 3The second embodiment of the registration method based on the probability distribution distance feature description is proposed based on the first embodiment, and the step S10 comprises:

[0101] In step S101, an initial source point cloud and an initial point cloud are obtained, and the initial source point cloud and the initial point cloud are calibrated.

[0102] It should be understood that after obtaining the initial source point cloud and the initial point cloud, since the depth sensing device is easily affected by environmental interference during measurement, the obtained acquisition data is often accompanied by interference such as noise, holes and abnormal points, and therefore, the three-dimensional point cloud (the initial source point cloud and the initial point cloud) needs to be preprocessed before formally performing a calculation task, for example, the preprocessing includes but is not limited to calibration and quality enhancement, that is, the initial source point cloud and the initial point cloud are first calibrated, the calibration can be realized by semi-automatic and automatic calibration by means of external object reference and checkerboard method, and the calibration can be camera calibration.

[0103] In step S102, the calibrated initial source point cloud and the initial point cloud are respectively subjected to quality enhancement to obtain a source point cloud and a target point cloud.

[0104] It can be understood that the source point cloud refers to the point cloud after the initial source point cloud is calibrated and quality enhanced, and similarly, the target point cloud refers to the point cloud after the initial point cloud is calibrated and quality enhanced, and the quality enhancement can be realized by sampling optimization, filtering denoising and hole filling.

[0105] In step S103, a source point cloud model is generated according to the source point cloud, and a target point cloud model is generated according to the target point cloud.

[0106] In step S104, a target proximity algorithm is used to respectively perform semantic segmentation on the source point cloud model and the target point cloud model to obtain a source point cloud segmentation result and a target point cloud segmentation result.

[0107] It can be understood that after obtaining the source point cloud model and the target point cloud model, the target proximity algorithm is used to respectively perform semantic segmentation on the source point cloud model and the target point cloud model to realize grouping of the point cloud data, which can help the point cloud features to embed certain global characteristics, and in some point clouds with obvious feature distribution, after the semantic segmentation is completed, the source point cloud segmentation result and the target point cloud segmentation result are obtained.

[0108] In step S105, a source point cloud mixed probability distribution map is generated according to the source point cloud segmentation result, and a target point cloud mixed probability distribution map is generated according to the target point cloud segmentation result.

[0109] It should be understood that after obtaining the source point cloud segmentation result and the target point cloud segmentation result, the source point cloud segmentation result and the target point cloud segmentation result are respectively labeled, and the labels are matched. The specific process of matching is that the features of each segmentation part are extracted to solve the similarity. When the solved similarity is greater than a similarity threshold, it is considered that the segmentation parts with the highest similarity in the source point cloud and the target point cloud are matched with each other and have the same label. Then, after the matching is completed, a source point cloud mixed probability distribution map is generated according to the source point cloud segmentation result, and a target point cloud mixed probability distribution map is generated according to the target point cloud segmentation result.

[0110] Further, step S105 comprises: calculating the source point cloud segmentation result to obtain a source point cloud segmentation part Gaussian distribution and a neighborhood Gaussian distribution of each point of the source point cloud; splicing the neighborhood Gaussian distribution of each point of the source point cloud and the source point cloud segmentation part Gaussian distribution to which each point of the source point cloud belongs to obtain a source point cloud mixed Gaussian distribution; generating a source point cloud mixed probability distribution map according to the source point cloud mixed Gaussian distribution and a preset graph structure; calculating the target point cloud segmentation result to obtain a target point cloud segmentation part Gaussian distribution and a neighborhood Gaussian distribution of each point of the target point cloud; splicing the neighborhood Gaussian distribution of each point of the target point cloud and the target point cloud segmentation part Gaussian distribution to which each point of the target point cloud belongs to obtain a target point cloud mixed Gaussian distribution; and generating a target point cloud mixed probability distribution map according to the target point cloud mixed Gaussian distribution and the preset graph structure.

[0111] It can be understood that after obtaining the source point cloud segmentation result, the source point cloud is divided into m parts according to the source point cloud segmentation result, and the corresponding Gaussian distribution, i.e., the source point cloud segmentation part Gaussian distribution guass(s_m i is calculated. Then, by setting a neighborhood radius r value, each point in the source point cloud searches for neighboring points and calculates a neighborhood Gaussian distribution as the distribution of the point, i.e., the neighborhood Gaussian distribution guass(s_p i of each point of the source point cloud. Similarly, according to the target point cloud segmentation result, the target point cloud is divided into m parts and the corresponding Gaussian distribution, i.e., the target point cloud segmentation part Gaussian distribution guass(t_m i is calculated. Then, by setting a neighborhood radius r value, each point in the target point cloud searches for neighboring points and calculates a neighborhood Gaussian distribution as the distribution of the point, i.e., the neighborhood Gaussian distribution guass(t_p i of each point of the target point cloud. Then, the source point cloud mixed Gaussian distribution <guass(s_p i , guass(s_m i )> and the target point cloud mixed Gaussian distribution <guass(t_p i , guass(t_m iThen, a source point cloud mixture probability distribution map is generated according to the source point cloud mixture Gaussian distribution and a preset graph structure, and a target point cloud mixture probability distribution map is generated according to the target point cloud mixture Gaussian distribution and the preset graph structure, specifically, the source point cloud mixture Gaussian distribution and the target point cloud mixture Gaussian distribution are modeled using the preset graph structure, each mixture Gaussian distribution is taken as a vertex in the graph, and an edge is established between any two vertices in a fully connected manner to obtain the source point cloud mixture probability distribution map and the target point cloud mixture probability distribution map.

[0112] In the embodiment, the initial source point cloud and the initial point cloud are obtained, and the initial source point cloud and the initial point cloud are calibrated. The calibrated initial source point cloud and the initial point cloud are subjected to quality enhancement respectively to obtain a source point cloud and a target point cloud. A source point cloud model is generated according to the source point cloud, and a target point cloud model is generated according to the target point cloud. The source point cloud model and the target point cloud model are subjected to semantic segmentation respectively by a target proximity algorithm to obtain a source point cloud segmentation result and a target point cloud segmentation result. A source point cloud mixture probability distribution map is generated according to the source point cloud segmentation result, and a target point cloud mixture probability distribution map is generated according to the target point cloud segmentation result. In the above manner, the source point cloud model and the target point cloud model are generated after the initial source point cloud and the initial point cloud are calibrated and subjected to quality enhancement, then the source point cloud model and the target point cloud model are subjected to semantic segmentation respectively by the target proximity algorithm, and finally the source point cloud mixture probability distribution map and the target point cloud mixture probability distribution map are generated according to the source point cloud segmentation result and the target point cloud segmentation result, so that the accuracy of generating the point cloud mixture probability distribution map can be effectively improved.

[0113] In addition, the embodiment of the present application further provides a storage medium, and the storage medium stores a registration program based on probability distribution distance feature description. When the registration program based on probability distribution distance feature description is executed by a processor, the steps of the registration method based on probability distribution distance feature description are implemented.

[0114] Since the storage medium adopts all the technical solutions of the above-mentioned embodiments, it at least has all the beneficial effects brought by the technical solutions of the above-mentioned embodiments, which will not be described here one by one.

[0115] In addition, with reference to Figure 4 , the embodiment of the present application further provides a registration device based on probability distribution distance feature description, which comprises:

[0116] The generation module 10 is configured to generate a source point cloud mixture probability distribution map according to a source point cloud, and generate a target point cloud mixture probability distribution map according to a target point cloud.

[0117] The computing module 20 is configured to perform nonlinear calculation on the source point cloud mixture probability distribution map and the target point cloud mixture probability distribution map respectively, and take the calculation results as source point cloud features and target point cloud features.

[0118] The generating module 10 is further configured to generate a point cloud correspondence relationship matrix according to the source point cloud features and the target point cloud features.

[0119] The determining module 30 is configured to determine a registration result of the source point cloud to the target point cloud according to the point cloud correspondence relationship matrix.

[0120] In the embodiment, a source point cloud mixture probability distribution map is generated according to a source point cloud, and a target point cloud mixture probability distribution map is generated according to a target point cloud; nonlinear calculation is performed on the source point cloud mixture probability distribution map and the target point cloud mixture probability distribution map respectively, and the calculation results are taken as source point cloud features and target point cloud features; a point cloud correspondence relationship matrix is generated according to the source point cloud features and the target point cloud features; and a registration result of the source point cloud to the target point cloud is determined according to the point cloud correspondence relationship matrix. In this way, the point cloud correspondence relationship matrix is generated according to the source point cloud features and the target point cloud features, and then the registration result of the source point cloud to the target point cloud is determined according to the point cloud correspondence relationship matrix, so that the accuracy, stability and generalization ability of the registered point cloud can be effectively improved.

[0121] It should be noted that the above-described workflow is only illustrative and does not limit the protection scope of the present application. In actual applications, a person skilled in the art can select part or all of the above-described workflow to achieve the purpose of the embodiment according to actual needs, which is not limited herein.

[0122] In addition, technical details not described in detail in the embodiment can be referred to the registration method based on probability distribution distance feature description provided by any embodiment of the present application, which will not be described herein.

[0123] In addition, it should be noted that in this document, the terms "comprise", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such a process, method, article or system. Without more limitations, the element defined by the statement "comprises a" does not exclude the presence of another identical element in the process, method, article or system that includes the element.

[0124] The above-mentioned embodiment numbers of the present application are only for description, and do not represent the advantages and disadvantages of the embodiments.

[0125] Those skilled in the art can clearly understand the above-mentioned embodiment method can be realized by means of software and the necessary general hardware platform, of course, can also be realized by hardware, but in many cases, the former is a better embodiment. Based on such understanding, the technical solutions of the present application essentially or say the part of the prior art contribution can be embodied in the form of software products, the computer software product is stored in a storage medium (such as read only memory (Read Only Memory, ROM) / RAM, disk, optical disk), including a number of instructions to make a terminal device (may be a mobile phone, computer, integrated platform workstation, or network equipment, etc.) executes the method described in various embodiments of the present application.

[0126] The above is only the preferred embodiment of the present application, not the patent scope of the present application, any equivalent structure or equivalent flow transformation using the content of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A registration method based on probability distribution distance feature description, characterized in that: The registration method based on probability distribution distance feature description comprises the following steps: generating a source point cloud mixed probability distribution map according to the source point cloud, and generating a target point cloud mixed probability distribution map according to the target point cloud; Performing nonlinear calculations on the source point cloud mixed probability distribution map and the target point cloud mixed probability distribution map, respectively, and using the calculation results as source point cloud features and target point cloud features, respectively; Generate a point cloud correspondence matrix according to the source point cloud features and the target point cloud features; Determining a registration result from the source point cloud to the target point cloud according to the point cloud correspondence matrix; Generating a source point cloud mixed probability distribution map according to the source point cloud, and generating a target point cloud mixed probability distribution map according to the target point cloud, includes: Acquire an initial source point cloud and an initial point cloud, and calibrate the initial source point cloud and the initial point cloud; The quality of the calibrated initial source point cloud and the initial point cloud is enhanced to obtain the source point cloud and the target point cloud; generating a source point cloud model according to the source point cloud, and generating a target point cloud model according to the target point cloud; Performing semantic segmentation on the source point cloud model and the target point cloud model respectively by using a target proximity algorithm to obtain a source point cloud segmentation result and a target point cloud segmentation result; Generating a source point cloud mixed probability distribution map according to the source point cloud segmentation result, and generating a target point cloud mixed probability distribution map according to the target point cloud segmentation result; Generating a source point cloud mixed probability distribution map according to the source point cloud segmentation result, and generating a target point cloud mixed probability distribution map according to the target point cloud segmentation result, comprises: Calculating the source point cloud segmentation result to obtain the Gaussian distribution of the source point cloud segmentation part and the Gaussian distribution of the neighborhood of each point in the source point cloud; splicing the neighborhood Gaussian distribution of each point in the source point cloud and the Gaussian distribution of the source point cloud segment to which each point in the source point cloud belongs, to obtain a mixed Gaussian distribution of the source point cloud; Generate a source point cloud mixed probability distribution map according to the source point cloud mixed Gaussian distribution and a preset map structure; Calculating the target point cloud segmentation result to obtain the Gaussian distribution of the target point cloud segmentation part and the Gaussian distribution of the neighborhood of each point in the target point cloud; The neighborhood Gaussian distribution of each point of the target point cloud is concatenated with the Gaussian distribution of the target point cloud segmentation part to which each point of the target point cloud belongs, so as to obtain a mixed Gaussian distribution of the target point cloud; A target point cloud mixed probability distribution map is generated according to the target point cloud mixed Gaussian distribution and a preset map structure.

2. The registration method based on probability distribution distance feature description according to claim 1, characterized in that: The performing nonlinear calculation on the source point cloud mixed probability distribution map and the target point cloud mixed probability distribution map respectively, and using the calculation results as the source point cloud features and the target point cloud features, includes: Calculate the first probability distribution distance between the target vertex in the source point cloud mixed probability distribution map and each vertex in the target point cloud mixed probability distribution map; Generate a source vertex vector according to the first probability distribution distance and the number of source vertices; Calculate a second probability distribution distance between the target vertex and the segmented portion of the source point cloud in the source point cloud mixed probability distribution graph; Generate a first vertex matrix according to the second probability distribution distance and the number of source vertices; Calculate a third probability distribution distance between the target vertex and the target point cloud segmentation part in the source point cloud mixed probability distribution graph; Generate a second vertex matrix according to the third probability distribution distance and the number of source vertices; Extending the target vertex vector to the first vertex matrix and the second vertex matrix respectively; Generate a source point cloud feature matrix according to the expanded first vertex matrix, the second vertex matrix and the target vertex vector; Calculate a fourth probability distribution distance between a target vertex in the target point cloud mixed probability distribution graph and each vertex in the source point cloud mixed probability distribution graph; generating a target vertex vector according to a fourth probability distribution distance and the number of target vertices; Calculate the fifth probability distribution distance between the target vertex and the target point cloud segmentation part in the target point cloud mixed probability distribution graph; generating a third vertex matrix according to the fifth probability distribution distance and the number of target vertices; Calculate the sixth probability distribution distance between the target vertex and the source point cloud segmentation part in the target point cloud mixed probability distribution graph; generating a fourth vertex matrix according to the sixth probability distribution distance and the number of target vertices; Extending the target vertex vector to the third vertex matrix and the fourth vertex matrix respectively; Generate a target point cloud feature matrix according to the expanded third vertex matrix, the fourth vertex matrix and the target vertex vector; The source point cloud features and the target point cloud features are obtained according to the source point cloud feature matrix and the target point cloud feature matrix.

3. The registration method based on probability distribution distance feature description according to claim 2, characterized in that: The obtaining of source point cloud features and target point cloud features according to the source point cloud feature matrix and the target point cloud feature matrix includes: Performing multi-layer perceptron calculation on the source point cloud feature matrix and the target point cloud feature matrix respectively to obtain a source point cloud feature description matrix and a target point cloud feature description matrix; Preprocessing the source point cloud feature description matrix and the target point cloud feature description matrix respectively to obtain a source point cloud feature description and a target point cloud feature description; The source point cloud features and the target point cloud features are obtained according to the source point cloud feature description and the target point cloud feature description.

4. The registration method based on probability distribution distance feature description according to claim 1, characterized in that: Generating a point cloud correspondence matrix according to the source point cloud features and the target point cloud features includes: Performing dot product calculation on the source point cloud features and the target point cloud features to obtain a number of feature similarities; Constructing a feature similarity matrix based on the plurality of feature similarities; The feature similarity matrix is ​​calculated by row and column calculation using the Sinkhorn algorithm to obtain a point cloud correspondence matrix.

5. The registration method based on probability distribution distance feature description according to claim 4, characterized in that: Determining the registration result of the source point cloud to the target point cloud according to the point cloud correspondence matrix includes: Obtaining matching point cloud pairs according to the point cloud correspondence matrix; Calculating the matching point cloud pair using an SVD algorithm according to a number of dimensions to obtain a rotation matrix and a translation vector; Constructing a rigid transformation relationship between the source point cloud and the target point cloud according to the rotation matrix and translation vector; Each point of the source point cloud is rotated and / or translated according to the rigid transformation relationship to obtain a registration result of the source point cloud to the target point cloud.

6. A registration device based on probability distribution distance feature description, capable of implementing the registration method based on probability distribution distance feature description according to any one of claims 1 to 5, characterized in that: The registration device based on probability distribution distance feature description includes: A generation module, configured to generate a source point cloud mixed probability distribution map based on the source point cloud, and to generate a target point cloud mixed probability distribution map based on the target point cloud; a calculation module, configured to perform nonlinear calculations on the source point cloud mixed probability distribution map and the target point cloud mixed probability distribution map, respectively, and use the calculation results as source point cloud features and target point cloud features; The generating module is further configured to generate a point cloud correspondence matrix according to the source point cloud features and the target point cloud features; A determination module is used to determine the registration result of the source point cloud to the target point cloud according to the point cloud correspondence matrix.

7. A registration device based on probability distribution distance feature description, characterized in that: The registration device based on probability distribution distance feature description includes: a memory, a processor, and a registration program based on probability distribution distance feature description stored in the memory and executable on the processor, wherein the registration program based on probability distribution distance feature description is configured to implement the registration method based on probability distribution distance feature description as described in any one of claims 1 to 5.

8. A storage medium, characterized in that: The storage medium stores a registration program based on probability distribution distance feature description, and when the registration program based on probability distribution distance feature description is executed by the processor, the registration method based on probability distribution distance feature description according to any one of claims 1 to 5 is implemented.

Images