Point cloud processing method and device based on geometric decomposition and complementation, equipment and medium
Through geometric decomposition and complementary point cloud processing methods, the shape features of point cloud data are extracted and feature vectors are generated, which solves the problem of low semantic analytical accuracy in point cloud processing in the prior art, improves processing efficiency and adaptability, and Effective compression of data is realized.
Patent Information
- Application Number
- CN202411015165.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-25
- Publication Date
- 2024-12-06
AI Technical Summary
The prior art ignores shape features in point cloud processing, resulting in low semantic parsing accuracy and processing network depth makes it impossible to apply to real-time scenarios.
Using a point cloud processing method based on geometric decomposition and complementarity, the original point cloud data is geometrically decomposed through feature filters to obtain feature point cloud data, and the feature learning algorithm is used for complementary learning, and feature vectors are generated, and then the feature splicing algorithm is used to translate the feature vectors. It is spliced into the target point cloud feature.
It improves the comprehensiveness and accuracy of semantic analytics of point cloud data, making point cloud data easier to process and analyze, adapt to different application scenarios, and realizes data compression, reduces storage and transmission requirements, while retaining key information.
Smart Images

Figure CN119091442A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of data processing, and in particular to a point cloud processing method, device, equipment and medium based on geometric decomposition and complementarity. Background Art
[0002] With the development of 3D perception technology, point cloud, as an important form of 3D data representation, has rich spatial structure information. Therefore, it has gradually become a hot research topic in industry and academia in recent years and is widely used in autonomous driving, robotics, virtual reality, augmented reality and other fields. In recent years, the application of deep learning has also made point cloud processing more accurate.
[0003] However, existing technologies treat all points in the point cloud in the same way, ignoring some of the shape features, which makes it impossible to better obtain the features in the point cloud. At the same time, the point cloud processing network is designed to be deeper and deeper, making it impossible to apply it to real-time scenarios.
[0004] Therefore, how to solve the low accuracy of semantic parsing of point cloud data has become a technical problem that needs to be solved urgently. Summary of the invention
[0005] The present application provides a point cloud processing method, device, equipment and medium based on geometric decomposition and complementarity, aiming to improve the accuracy of semantic analysis of point cloud data.
[0006] In a first aspect, the present application provides a point cloud processing method based on geometric decomposition and complementation, the point cloud processing method based on geometric decomposition and complementation comprising the following steps:
[0007] Based on the feature filter, the original point cloud data is geometrically decomposed to obtain at least one feature point cloud data;
[0008] Based on the feature learning algorithm, the original point cloud data and each of the feature point cloud data are respectively subjected to complementary learning to obtain a feature vector corresponding to each of the feature point cloud data;
[0009] Based on the feature stitching algorithm, the feature vectors corresponding to the feature point cloud data are stitched together to obtain the target point cloud features.
[0010] In a second aspect, the present application further provides a point cloud processing device based on geometric decomposition and complementation, the point cloud processing device based on geometric decomposition and complementation comprising:
[0011] A point cloud data decomposition module, used for geometrically decomposing the original point cloud data based on a feature filter to obtain at least one feature point cloud data;
[0012] A feature vector acquisition module, used for performing complementary learning on the original point cloud data and each of the feature point cloud data based on a feature learning algorithm to obtain a feature vector corresponding to each of the feature point cloud data;
[0013] The feature stitching module is used to stitch the feature vectors corresponding to the feature point cloud data based on the feature stitching algorithm to obtain the target point cloud features.
[0014] In a third aspect, the present application also provides a computer device, comprising a processor, a memory, and a computer program stored in the memory and executable by the processor, wherein when the computer program is executed by the processor, the steps of the point cloud processing method based on geometric decomposition and complementarity as described above are implemented.
[0015] In a fourth aspect, the present application also provides a computer-readable storage medium, on which a computer program is stored, wherein when the computer program is executed by a processor, the steps of the point cloud processing method based on geometric decomposition and complementation as described above are implemented.
[0016] The present application provides a point cloud processing method, device, computer equipment and storage medium based on geometric decomposition and complementation. The method of the present application includes geometric decomposition of the original point cloud data based on a feature filter to obtain at least one feature point cloud data; based on a feature learning algorithm, the original point cloud data is respectively complementary to each of the feature point cloud data to obtain the feature vector corresponding to each of the feature point cloud data; based on a feature splicing algorithm, the feature vectors corresponding to each of the feature point cloud data are spliced to obtain the target point cloud feature. In the above manner, the present application performs geometric decomposition through a feature filter to decompose the original point cloud data into a plurality of different feature point cloud data, and performs feature learning through feature complementation, so that richer information can be extracted from the original point cloud data, thereby obtaining global and complementary point cloud geometric semantics, and improving the comprehensiveness and accuracy of the semantic analysis of the original point cloud data. The generation of feature vectors makes point cloud data easier to process and analyze, can better adapt to different application scenarios, and improve the adaptability of point cloud data applications. At the same time, the splicing of feature vectors may achieve data compression, reduce the demand for storage and transmission, while retaining key information, reducing the amount of data while maintaining data integrity. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.
[0018] Figure 1 A schematic flow chart of a first embodiment of a point cloud processing method based on geometric decomposition and complementarity provided by the present application;
[0019] Figure 2 A schematic diagram of a feature vector calculation process provided in an embodiment of the present application;
[0020] Figure 3 A schematic flow chart of a second embodiment of a point cloud processing method based on geometric decomposition and complementation provided by the present application;
[0021] Figure 4 It is a structural schematic diagram of a first embodiment of a point cloud processing device based on geometric decomposition and complementation provided by the present application;
[0022] Figure 5 It is a schematic block diagram of the structure of a computer device provided in an embodiment of the present application.
[0023] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION
[0024] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.
[0025] The flowcharts shown in the accompanying drawings are only examples and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may also be decomposed, combined or partially merged, so the actual execution order may change according to actual conditions.
[0026] In conjunction with the accompanying drawings, some embodiments of the present application are described in detail below. In the absence of conflict, the following embodiments and features in the embodiments can be combined with each other.
[0027] Please refer to Figure 1 , Figure 1 A schematic flow chart of a first embodiment of a point cloud processing method based on geometric decomposition and complementarity provided in the present application.
[0028] like Figure 1 As shown, the point cloud processing method based on geometric decomposition and complementation includes steps S101 to S103.
[0029] S101, performing geometric decomposition on original point cloud data based on a feature filter to obtain at least one feature point cloud data;
[0030] In one embodiment, the feature point cloud data includes sharp point cloud data and flat point cloud data, etc. Alternatively, the feature point cloud data may include a border area and a non-border area (such as a plane), etc.
[0031] In one embodiment, the constructed feature filter is used to calculate the feature value of each node in the original point cloud data.
[0032] Furthermore, based on the feature filter, the feature value of each node in the original point cloud data is calculated; based on the feature value of each node and a preset data division rule, the feature value of each node is divided into data to obtain at least one of the feature point cloud data.
[0033] In one embodiment, a feature filter is applied to the feature vector of each point, which may be obtained by weighted averaging or other mathematical operations, and then the feature value of each point in the original point cloud data is calculated and updated to reflect the features of the points in its neighborhood.
[0034] In one embodiment, the feature filter may be a filter based on local curvature, or may be other types of filters such as a filter based on an edge detection algorithm.
[0035] In one embodiment, the method of calculating the eigenvalue by the feature filter includes but is not limited to normal and curvature calculation, eigenvalue analysis, PFH (Point Feature Histogram), FPFH (Fast Point Feature Histogram), 3D Shape Context and Spin Image.
[0036] Among them, normal and curvature calculation: point cloud normal and curvature are the basis of point cloud data feature extraction, which can be calculated by least squares fitting or covariance matrix-based method. For each point, the normal vector and curvature of the point can be calculated based on the position of the surrounding points and the normal vector.
[0037] Eigenvalue Analysis: Eigenvalue analysis is a method used to calculate the principal curvature and principal direction in point cloud data. It determines the principal direction and principal curvature of point cloud data by solving the eigenvectors and eigenvalues of the covariance matrix.
[0038] PFH (Point Feature Histogram): The PFH algorithm is a histogram-based point cloud feature extraction algorithm that can describe the relationship between a point and its surrounding points. It represents the local shape features in point cloud data by calculating the normal difference, distance, and angle between point pairs.
[0039] FPFH (Fast Point Feature Histogram): The FPFH algorithm is an improvement of the PFH algorithm, which can calculate the features of point cloud data in a shorter time. It describes the local shape features in point cloud data by calculating the normal difference, distance and angle between point pairs and adding the normal vector information of the points.
[0040] 3D Shape Context: The 3D Shape Context algorithm is a global feature extraction algorithm based on point cloud data that can describe the shape and structure information of point cloud data. It constructs the feature vector of point cloud data by calculating the distance, angle and direction relationship between a point and surrounding points.
[0041] Spin Image: The Spin Image algorithm is a point cloud feature extraction algorithm based on image descriptors, which can describe the local shape features in point cloud data. It represents the features of point cloud data by projecting the point cloud onto a two-dimensional image plane and calculating the grayscale histogram of each pixel.
[0042] In one embodiment, a feature threshold is determined based on the data division rule; and based on a comparison result between the feature value of each node and the feature threshold, the feature value of each node is divided into at least one of the feature point cloud data.
[0043] In one embodiment, the feature threshold may be set according to the statistical characteristics of the point cloud data or an optimal threshold determined by experiments. The feature threshold may be fixed or dynamically adjusted, depending on the characteristics of the data and application requirements.
[0044] In one embodiment, the feature value of each node is compared with the corresponding feature threshold. According to the comparison result, the feature value of the node is divided into different categories. For example, if the sharpness feature value of the node is higher than the sharpness threshold, it is divided into sharp point cloud data; if the flatness feature value of the node is lower than the flatness threshold, it is divided into flat point cloud data. Or, if the edge feature value of the node is higher than the edge threshold, it is divided into border area point cloud data.
[0045] In one embodiment, multiple feature thresholds may be set according to data distribution of different feature point cloud data.
[0046] For example, a sharpness threshold can be defined, which is based on the calculation of local curvature. If the local curvature of a point is greater than this threshold, then the point is considered sharp and is classified into the sharp point cloud data. Similarly, a flatness threshold can be defined based on the consistency of local normal vectors. If the normal vector of a point does not vary much within a small neighborhood, then the point is considered flat and is classified into the flat point cloud data.
[0047] In another embodiment, based on the data division rule, the feature allocation ratio corresponding to the feature point cloud data is determined; based on the numerical sorting result of the feature value of each node and the feature allocation ratio, the feature value of each node is divided into at least one of the feature point cloud data.
[0048] In one embodiment, the proportion of each type of feature point cloud data in the total data may be predetermined, and this proportion may be based on prior knowledge, statistical characteristics of the data, or requirements of a specific application.
[0049] In one embodiment, the feature values of each node in the point cloud are calculated, and these feature values will be used for subsequent sorting and classification. The feature values of all nodes are sorted from small to large or from large to small, depending on the type of feature (for example, sharpness can be sorted from large to small).
[0050] Exemplarily, the division point is determined according to the feature allocation ratio and the sorting result. For example, if the sharp point cloud data should account for 20% of the total data, the node at the 20% position is found in the sorted feature value list. According to the division point, the feature value of the node is divided into different feature point cloud data categories. The feature value of the node before the division point is classified as one feature point cloud data, and the feature value after the division point is classified as another feature point cloud data.
[0051] In a specific embodiment, suppose there are 10,000 nodes and it is decided that sharp point cloud data should account for 20%. First, the sharpness eigenvalues of all nodes are calculated and sorted. Then, the eigenvalue of the 2000th node in the sorted list is found and used as the dividing point. All nodes with sharpness eigenvalues higher than this dividing point will be classified as sharp point cloud data.
[0052] Optionally, the feature allocation ratio or partitioning strategy can be adjusted according to feedback from actual applications to optimize the quality of feature point cloud data.
[0053] S102, based on a feature learning algorithm, performing complementary learning on the original point cloud data and each of the feature point cloud data, to obtain a feature vector corresponding to each of the feature point cloud data;
[0054] In one embodiment, after distinguishing the border and the non-border, the original point cloud and the two subsets should be used to perform complementary learning of geometric information. Figure 2 As shown in the figure, the point features of flat areas and sharp areas are learned respectively by using dot product attention. The complementary information between the original point cloud data and the feature point cloud data is used for learning, that is, the feature vectors of the original data and the subset data are fused or interacted. The feature vectors learned from different point cloud data are fused to form a comprehensive feature representation, which helps to improve the model's understanding of the original point cloud data.
[0055] In one embodiment, a dot product attention mechanism can be applied in feature learning of flat areas and sharp areas. Dot product attention allows the model to focus on the most important parts of the point cloud, thereby more effectively capturing geometric features. For example, for border areas, it may be necessary to pay attention to the sharpness and variability of the boundaries, so dot product attention can be used to enhance the learning of these features. For non-border areas, it may be necessary to pay attention to the flatness and consistency of the surface, and dot product attention is also applied to extract features of these areas.
[0056] Furthermore, based on a multi-layer perceptron, a nonlinear transformation is performed on the original point cloud data and each of the feature point cloud data to obtain a first vector corresponding to the original point cloud data and a second vector corresponding to each of the feature point cloud data; based on the first vector and the second vector corresponding to each of the feature point cloud data, an attention score corresponding to each of the feature point cloud data is calculated; based on each of the feature point cloud data and the attention score corresponding to each of the feature point cloud data, the feature vector corresponding to each of the feature point cloud data is calculated.
[0057] In one embodiment, a multilayer perceptron (MLP) is a common and powerful neural network model that is widely used in classification and regression problems and has achieved excellent performance in many practical applications.
[0058] In one embodiment, a multi-layer perception network can be designed, which can accept point cloud data as input and has multiple hidden layers to perform nonlinear transformation. The original point cloud data and each feature point cloud data are input into the MLP, and the first vector and the second vector after nonlinear transformation are obtained through forward propagation. The first vector and the second vector are used in combination with the attention mechanism to calculate the attention score of each feature point cloud data. According to the attention score of each feature point cloud data, the weight in the second vector is adjusted to highlight the important feature points, and finally the feature vector corresponding to each feature point cloud data is calculated.
[0059] For example, for the feature point cloud data in the border area, MLP can learn the geometric features related to the boundary, and the attention mechanism can further emphasize the sharpness and variability of the boundary.
[0060] In one embodiment, the feature vector calculation process of each feature point cloud data is performed separately, that is, each type of feature point cloud data is interactively learned with the original point cloud data.
[0061] In one embodiment, in the embodiment of the present application, the feature vector extraction process of each feature point cloud data is the same, and the embodiment of the present application will not go into details about this. Only the feature point cloud data of the flat area is used as an example for explanation, and the other feature point cloud data are the same.
[0062] In one embodiment, the original point cloud data and the feature power data corresponding to the flat area are first passed to the multi-layer perceptron for a nonlinear transformation to obtain the first vector corresponding to the original point cloud data and the second vector corresponding to the flat area. Then the dot product of the two vectors is calculated as the attention score, which can be intuitively understood as the importance of different points in the area to the overall point cloud. It can be expressed as:
[0063] Score attention =φ(X origin )φ(X Gentle ) T
[0064] After that, the attention score is multiplied by the feature vector of the flat area to obtain a new feature vector with the same shape as the original point cloud. It should be noted that only the features of the flat area are obtained at this time, so the embodiment of the present application also refers to the design of ResNet and adds a residual connection, which can be expressed as:
[0065] Y Gentle =X Origin +Score attention φ(X Gentle )
[0066] Through the above method, we can get Y Gentle (Flat region feature vector) and Y Sharp (Sharp Area Eigenvectors) Two eigenvectors.
[0067] S103 . Based on a feature stitching algorithm, stitching operations are performed on the feature vectors corresponding to the feature point cloud data to obtain target point cloud features.
[0068] In one embodiment, the feature vectors corresponding to the feature point cloud data are spliced through a splicing operation, thereby effectively obtaining sufficient point cloud features through the two steps of geometric decomposition and complementation.
[0069] In one embodiment, each feature vector is normalized before concatenation to ensure that they have a consistent scale and to prevent certain features from having a disproportionate impact on the concatenation result due to a large value range.
[0070] In one embodiment, all feature vectors are concatenated in a certain order, and the concatenation order can be determined according to the importance of the features or application requirements.
[0071] In one embodiment, based on the splicing, feature information can be further integrated through feature fusion technology. For example, weighted average, principal component analysis (PCA) or other fusion algorithms can be used to optimize the feature vector.
[0072] In one embodiment, the spliced feature vector is the target point cloud feature, which integrates the geometric and structural information of the original point cloud data and its different feature point cloud data.
[0073] This embodiment provides a point cloud processing method based on geometric decomposition and complementation. The method of this application performs geometric decomposition through feature filters, decomposes the original point cloud data into multiple different feature point cloud data, and performs feature learning through feature complementation. It can extract richer information from the original point cloud data, thereby obtaining global and complementary point cloud geometric semantics, and improving the comprehensiveness and accuracy of the semantic analysis of the original point cloud data. The generation of feature vectors makes point cloud data easier to process and analyze, and can better adapt to different application scenarios, improving the adaptability of point cloud data applications. At the same time, the splicing of feature vectors may achieve data compression, reduce storage and transmission requirements, while retaining key information, reducing the amount of data while maintaining data integrity.
[0074] Please refer to Figure 3 , Figure 3 A schematic flow chart of a second embodiment of a point cloud processing method based on geometric decomposition and complementation provided in the present application.
[0075] In this embodiment, Figure 3 As shown, based on the above Figure 1 In the illustrated embodiment, before step S101, the following steps are further included:
[0076] S201, constructing an adjacency matrix based on the distance between each node and at least one neighboring point in the original point cloud data;
[0077] In one embodiment, the original point cloud data is a data set. For the data set P of the original point cloud data, 1 , p 2 , p 3 ,…,p n}, each point p i has a d-dimensional feature vector.
[0078] First, we can calculate the Euclidean distance between all pairs of points. i and p j , the Euclidean distance between two points is calculated as:
[0079]
[0080] Among them, x ik and x jk They are point p i and p j The eigenvalue in the kth dimension.
[0081] In one embodiment, for each point p i , we can find the K nearest neighboring points by sorting the distance list of each point.
[0082] In one embodiment, the adjacency matrix A is an n×n matrix used to represent the connection relationship between vertices in the graph. ij Represents point p i and p j If p i and p j is a neighboring point, then A ij Nonzero; otherwise zero.
[0083] In one embodiment, when constructing the adjacency matrix, an n×n matrix A is first initialized, and all elements are set to 0. For each point p i , fill the distances of its K nearest neighbor points into the corresponding positions of matrix A, and thus construct the adjacency matrix.
[0084] S202, based on the distance of each node in the adjacency matrix, calculating the weight corresponding to each node to obtain a weight matrix;
[0085] In one embodiment, the adjacency matrix provides connection information between vertices, but generally does not include the strength or weight of the connection. The weight matrix further refines the weight of the connection based on the adjacency matrix. For example, the weight matrix can be generated by calculating the distance between vertex pairs, and the weight can be the inverse of the distance or other measurement.
[0086] In one embodiment, the weight matrix W is used to represent the similarity or connection strength between points. The weight can be calculated by the inverse of the distance, that is:
[0087]
[0088] Among them, d(p i , p j ) is point p i and p jThe distance between.
[0089] In one embodiment, when constructing the weight matrix, an n×n matrix W is first initialized, and all elements are set to 0. For each non-zero element A in the adjacency matrix A ij , calculate its corresponding weight W ij .
[0090] S203: construct the feature filter based on the weight matrix.
[0091] In one embodiment, a feature filter is a mechanism for processing data features, particularly in image processing and signal processing. It can enhance or smooth features based on an adjacency matrix and a weight matrix. The feature filter processes data by applying a graph filtering algorithm (such as Laplace filtering, Gaussian filtering, etc.) to extract or enhance useful features in the data.
[0092] In one embodiment, the weight matrix provides weight information between vertices, and the feature filter uses these weights to process data. For example, in image processing, the weight matrix can be used to define the interaction between pixels to achieve image smoothing or feature enhancement.
[0093] In another embodiment, the adjacency matrix can also be used directly for feature filtering. For example, in a graph convolutional network (GCN), the adjacency matrix defines the connection relationship between nodes, and features are extracted through a graph convolution operation.
[0094] This embodiment can effectively capture the local structure and pattern in the point cloud data by constructing the adjacency matrix and the weight matrix. The application of feature filters can improve the quality of data and the expressiveness of features, which is helpful for subsequent data analysis and machine learning tasks.
[0095] See also Figure 4 , Figure 4 It is a structural schematic diagram of a first embodiment of a point cloud processing device based on geometric decomposition and complementation provided in the present application. The point cloud processing device based on geometric decomposition and complementation is used to execute the aforementioned point cloud processing method based on geometric decomposition and complementation.
[0096] like Figure 4 As shown, the point cloud processing device 300 based on geometric decomposition and complementation includes: a data acquisition module 301, an abnormal data screening module 302, a fault coefficient calculation module 303 and a fault type determination module 304.
[0097] The point cloud data decomposition module 301 is used to geometrically decompose the original point cloud data based on the feature filter to obtain at least one feature point cloud data;
[0098] A feature vector acquisition module 302 is used to perform complementary learning on the original point cloud data and each of the feature point cloud data based on a feature learning algorithm to obtain a feature vector corresponding to each of the feature point cloud data;
[0099] The feature stitching module 303 is used to stitch the feature vectors corresponding to the feature point cloud data based on a feature stitching algorithm to obtain target point cloud features.
[0100] In one embodiment, the point cloud processing device 300 based on geometric decomposition and complementation further includes a filter construction module, including:
[0101] An adjacency matrix construction unit, configured to construct an adjacency matrix based on the distance between each node and at least one neighboring point in the original point cloud data;
[0102] A weight matrix calculation unit, used to calculate the weight corresponding to each node based on the distance of each node in the adjacency matrix to obtain a weight matrix;
[0103] A filter construction unit is used to construct the feature filter based on the weight matrix.
[0104] In one embodiment, the point cloud data decomposition module 301 includes:
[0105] An eigenvalue calculation unit, used to calculate the eigenvalue of each node in the original point cloud data based on the eigenfilter;
[0106] The data partitioning unit is used to perform data partitioning on the characteristic values of each node based on the characteristic values of each node and a preset data partitioning rule to obtain at least one of the characteristic point cloud data.
[0107] In one embodiment, the data partitioning unit includes:
[0108] A feature threshold determination subunit, used to determine a feature threshold based on the data partitioning rule;
[0109] The data division first subunit is used to divide the feature value of each node into at least one of the feature point cloud data based on the comparison result between the feature value of each node and the feature threshold.
[0110] In one embodiment, the data partitioning unit further includes:
[0111] A feature allocation ratio determination subunit, used to determine the feature allocation ratio corresponding to the feature point cloud data based on the data division rule;
[0112] The second data division subunit is used to divide the feature value of each node into at least one of the feature point cloud data based on the numerical sorting result of the feature value of each node and the feature allocation ratio.
[0113] In one embodiment, the feature vector obtaining module 302 includes:
[0114] A nonlinear transformation unit, configured to perform nonlinear transformation on the original point cloud data and each of the feature point cloud data based on a multi-layer perceptron, to obtain a first vector corresponding to the original point cloud data and a second vector corresponding to each of the feature point cloud data;
[0115] An attention score calculation unit, used for calculating the attention score corresponding to each feature point cloud data based on the first vector and the second vector corresponding to each feature point cloud data;
[0116] A feature vector calculation unit is used to calculate the feature vector corresponding to each feature point cloud data based on each feature point cloud data and the attention score corresponding to each feature point cloud data.
[0117] In one embodiment, the feature point cloud data includes sharp point cloud data and flat point cloud data.
[0118] It should be noted that technical personnel in the relevant field can clearly understand that, for the convenience and conciseness of description, the specific working process of the above-described device and each module can refer to the corresponding process in the aforementioned embodiment of the point cloud processing method based on geometric decomposition and complementarity, and will not be repeated here.
[0119] The apparatus provided in the above embodiment may be implemented in the form of a computer program. The computer program may be Figure 5 Runs on the computer device shown.
[0120] See also Figure 5 , Figure 5 1 is a schematic block diagram of a computer device provided in an embodiment of the present application. The computer device may be a server.
[0121] See also Figure 5 The computer device includes a processor, a memory and a network interface connected through a system bus, wherein the memory may include a non-volatile storage medium and an internal memory.
[0122] The non-volatile storage medium can store an operating system and a computer program. The computer program includes program instructions, and when the program instructions are executed, the processor can execute any point cloud processing method based on geometric decomposition and complementation.
[0123] The processor is used to provide computing and control capabilities and support the operation of the entire computer equipment.
[0124] The internal memory provides an environment for the operation of the computer program in the non-volatile storage medium. When the computer program is executed by the processor, the processor can execute any point cloud processing method based on geometric decomposition and complementation.
[0125] The network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art will appreciate that Figure 5 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.
[0126] It should be understood that the processor may be a central processing unit (CPU), and the processor may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.
[0127] In one embodiment, the processor is used to run a computer program stored in the memory to implement the following steps:
[0128] Based on the feature filter, the original point cloud data is geometrically decomposed to obtain at least one feature point cloud data;
[0129] Based on the feature learning algorithm, the original point cloud data and each of the feature point cloud data are respectively subjected to complementary learning to obtain a feature vector corresponding to each of the feature point cloud data;
[0130] Based on the feature stitching algorithm, the feature vectors corresponding to the feature point cloud data are stitched together to obtain the target point cloud features.
[0131] In one embodiment, before implementing the geometric decomposition of the original point cloud data based on the feature filter to obtain at least one feature point cloud data, the processor is further used to implement:
[0132] Constructing an adjacency matrix based on the distance between each node and at least one neighboring point in the original point cloud data;
[0133] Based on the distance between each node in the adjacency matrix, the weight corresponding to each node is calculated to obtain a weight matrix;
[0134] Based on the weight matrix, the feature filter is constructed.
[0135] In one embodiment, when the processor performs geometric decomposition on the original point cloud data based on the feature filter to obtain at least one feature point cloud data, it is used to implement:
[0136] Based on the feature filter, calculating the feature value of each node in the original point cloud data;
[0137] Based on the characteristic value of each node and a preset data division rule, data division is performed on the characteristic value of each node to obtain at least one of the characteristic point cloud data.
[0138] In one embodiment, when the processor performs data division on the feature values of each node based on the feature values of each node and a preset data division rule to obtain at least one feature point cloud data, it is used to implement:
[0139] Based on the data partitioning rule, determining a feature threshold;
[0140] Based on the comparison result between the characteristic value of each node and the characteristic threshold, the characteristic value of each node is divided into at least one of the characteristic point cloud data.
[0141] In one embodiment, when the processor performs data division on the feature values of each node based on the feature values of each node and a preset data division rule to obtain at least one feature point cloud data, it is also used to implement:
[0142] Based on the data division rule, determining a feature allocation ratio corresponding to the feature point cloud data;
[0143] Based on the numerical sorting results of the feature values of the nodes and the feature allocation ratio, the feature values of the nodes are divided into at least one of the feature point cloud data.
[0144] In one embodiment, when the processor implements the feature-based learning algorithm to perform complementary learning on the original point cloud data and each of the feature point cloud data to obtain a feature vector corresponding to each of the feature point cloud data, it is used to implement:
[0145] Based on a multi-layer perceptron, nonlinear transformation is performed on the original point cloud data and each of the feature point cloud data to obtain a first vector corresponding to the original point cloud data and a second vector corresponding to each of the feature point cloud data;
[0146] Calculate the attention score corresponding to each feature point cloud data based on the first vector and the second vector corresponding to each feature point cloud data;
[0147] Based on each of the feature point cloud data and the attention score corresponding to each of the feature point cloud data, the feature vector corresponding to each of the feature point cloud data is calculated.
[0148] In one embodiment, the feature point cloud data includes sharp point cloud data and flat point cloud data.
[0149] A computer-readable storage medium is also provided in an embodiment of the present application, wherein the computer-readable storage medium stores a computer program, wherein the computer program includes program instructions, and the processor executes the program instructions to implement any one of the point cloud processing methods based on geometric decomposition and complementarity provided in the embodiments of the present application.
[0150] The computer-readable storage medium may be an internal storage unit of the computer device described in the foregoing embodiment, such as a hard disk or memory of the computer device. The computer-readable storage medium may also be an external storage device of the computer device, such as a plug-in hard disk, a smart memory card (SmartMedia Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc., equipped on the computer device.
[0151] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed in the present application, and these modifications or replacements should be included in the protection scope of the present application. Therefore, the protection scope of the present application shall be based on the protection scope of the claims.
Claims
1. A point cloud processing method based on geometric decomposition and complementarity, characterized in that: The method comprises: Based on the feature filter, the original point cloud data is geometrically decomposed to obtain at least one feature point cloud data; Based on the feature learning algorithm, the original point cloud data and each of the feature point cloud data are respectively subjected to complementary learning to obtain a feature vector corresponding to each of the feature point cloud data; Based on the feature stitching algorithm, the feature vectors corresponding to the feature point cloud data are stitched together to obtain the target point cloud features.
2. The point cloud processing method based on geometric decomposition and complementarity according to claim 1, characterized in that: Before geometrically decomposing the original point cloud data based on the feature filter to obtain at least one feature point cloud data, the method further includes: Constructing an adjacency matrix based on the distance between each node and at least one neighboring point in the original point cloud data; Based on the distance between each node in the adjacency matrix, the weight corresponding to each node is calculated to obtain a weight matrix; Based on the weight matrix, the feature filter is constructed.
3. The point cloud processing method based on geometric decomposition and complementarity according to claim 1, characterized in that: The method of geometrically decomposing the original point cloud data based on the feature filter to obtain at least one feature point cloud data includes: Based on the feature filter, calculating the feature value of each node in the original point cloud data; Based on the characteristic value of each node and a preset data division rule, data division is performed on the characteristic value of each node to obtain at least one of the characteristic point cloud data.
4. The point cloud processing method based on geometric decomposition and complementarity according to claim 3, characterized in that: The step of performing data division on the characteristic values of each node based on the characteristic values of each node and a preset data division rule to obtain at least one characteristic point cloud data includes: Based on the data partitioning rule, determining a feature threshold; Based on the comparison result between the characteristic value of each node and the characteristic threshold, the characteristic value of each node is divided into at least one of the characteristic point cloud data.
5. The point cloud processing method based on geometric decomposition and complementation according to claim 3, characterized in that: The step of performing data division on the characteristic values of each node based on the characteristic values of each node and a preset data division rule to obtain at least one characteristic point cloud data further includes: Based on the data division rule, determining a feature allocation ratio corresponding to the feature point cloud data; Based on the numerical sorting results of the feature values of the nodes and the feature allocation ratio, the feature values of the nodes are divided into at least one of the feature point cloud data.
6. The point cloud processing method based on geometric decomposition and complementation according to claim 1, characterized in that: Based on the feature learning algorithm, the original point cloud data and each of the feature point cloud data are respectively subjected to complementary learning to obtain a feature vector corresponding to each of the feature point cloud data, including: Based on a multi-layer perceptron, nonlinear transformation is performed on the original point cloud data and each of the feature point cloud data to obtain a first vector corresponding to the original point cloud data and a second vector corresponding to each of the feature point cloud data; Calculate the attention score corresponding to each feature point cloud data based on the first vector and the second vector corresponding to each feature point cloud data; Based on each of the feature point cloud data and the attention score corresponding to each of the feature point cloud data, the feature vector corresponding to each of the feature point cloud data is calculated.
7. The point cloud processing method based on geometric decomposition and complementation according to claim 1, characterized in that: The feature point cloud data includes sharp point cloud data and flat point cloud data.
8. A point cloud processing device based on geometric decomposition and complementation, characterized in that: The point cloud processing device based on geometric decomposition and complementation includes: A point cloud data decomposition module, used for geometrically decomposing the original point cloud data based on a feature filter to obtain at least one feature point cloud data; A feature vector acquisition module, used for performing complementary learning on the original point cloud data and each of the feature point cloud data based on a feature learning algorithm to obtain a feature vector corresponding to each of the feature point cloud data; The feature stitching module is used to stitch the feature vectors corresponding to the feature point cloud data based on the feature stitching algorithm to obtain the target point cloud features.
9. A computer device, characterized in that: The computer device includes a processor, a memory, and a computer program stored in the memory and executable by the processor, wherein when the computer program is executed by the processor, the steps of the point cloud processing method based on geometric decomposition and complementation as described in any one of claims 1 to 7 are implemented.
10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, wherein when the computer program is executed by a processor, the steps of the point cloud processing method based on geometric decomposition and complementation as claimed in any one of claims 1 to 7 are implemented.
Citation Information
Patent Citations
Construction method and device of 3D target detection model based on LiDAR point cloud and RGB image
CN118212405A
