A point cloud registration method, device, equipment and medium

Abstract

The application discloses a point cloud registration method, device and equipment and a medium, and relates to the technical field of data processing. The method comprises the following steps: obtaining point cloud data to be processed and a point cloud registration request of a to-be-processed task; determining a point cloud registration algorithm of the point cloud data to be processed by using a trained registration matching model, and performing registration on the point cloud data to be processed by using the point cloud registration algorithm to obtain a transformation matrix; dividing the point cloud data to be processed into a plurality of sub-point cloud data according to the point cloud registration request; converting each sub-point cloud data into a homogeneous coordinate matrix of a preset shape, performing coordinate transformation on the homogeneous coordinate matrix by using the transformation matrix, and obtaining a restored point cloud matrix of each sub-point cloud data; determining the number of the sub-point cloud data, splicing the restored coordinate matrix according to the number, and obtaining point cloud data after registration. The application can greatly shorten the entire registration cycle while maintaining high-quality output by comprehensively using multiple strategies.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a point cloud registration method, device, equipment and medium. BACKGROUND

[0002] Point cloud registration is one of the key technologies in the field of computer vision and robot navigation, and its main purpose is to align point cloud data from different perspectives or time points to achieve environmental modeling, object recognition and other functions. With the development of applications such as autonomous driving and unmanned aerial vehicle navigation, higher requirements are put forward for the speed and robustness of point cloud registration algorithms.

[0003] However, the current widely used point cloud registration algorithms, such as Iterative Closest Point (ICP) and its variants, although perform well in accuracy, are inefficient in processing large-scale point cloud data, while some lightweight algorithms are fast but sacrifice part of the accuracy.

[0004] Therefore, how to shorten the response time while ensuring high accuracy of point cloud registration has become a problem to be solved in the industry. SUMMARY

[0005] Therefore, the embodiments of the present application provide a point cloud registration method, device, equipment and medium to solve the problem of poor effect when processing large-scale point cloud data.

[0006] According to a first aspect, the embodiments of the present application provide a point cloud registration method, which comprises:

[0007] Obtaining point cloud data to be processed and a point cloud registration request of a to-be-processed task;

[0008] Determining a point cloud registration algorithm of the point cloud data to be processed by using a trained registration matching model, and performing registration on the point cloud data to be processed by using the point cloud registration algorithm to obtain a transformation matrix; the registration matching model is trained by using a decision tree algorithm;

[0009] According to the point cloud registration request, the point cloud data to be processed is divided into a plurality of sub-point cloud data;

[0010] In a parallel processing manner, each sub-point cloud data is converted into a homogeneous coordinate matrix of a preset shape, and the homogeneous coordinate matrix is subjected to coordinate transformation by using the transformation matrix to obtain a restored point cloud matrix of each sub-point cloud data;

[0011] Determining the number of the sub-point cloud data, and splicing the restored coordinate matrix according to the number to obtain the registered point cloud data.

[0012] In conjunction with the first aspect, in the first embodiment of the first aspect, the registration and matching model is trained through the following steps:

[0013] Acquire historical point cloud data that has been registered in different scenarios, and determine the point cloud registration algorithm used in the historical point cloud data;

[0014] Feature extraction is performed on historical point cloud data to obtain preset historical point cloud features;

[0015] Historical point cloud features are preprocessed; the preprocessing methods include standardization and normalization.

[0016] By using historical point cloud features as influencing factors, a mapping relationship between the influencing factors of historical point cloud data and point cloud registration algorithms is established to obtain historical training data of historical point cloud data.

[0017] A decision tree is constructed using historical training data, and a registration and matching model is obtained based on the constructed decision tree; the registration and matching model is used to obtain the point cloud registration algorithm for the point cloud data to be processed.

[0018] In conjunction with the first embodiment of the first aspect, in the second embodiment of the first aspect, the step of constructing a decision tree using historical training data and obtaining a registration and matching model based on the constructed decision tree specifically includes:

[0019] Collect all historical training data to obtain the historical training set, place the historical training set at the root node of the decision tree, and calculate the information entropy of the root node.

[0020] Traverse each influencing factor, determine the conditional entropy of each influencing factor, and determine the information gain of each influencing factor based on the information entropy of the root node and the conditional entropy of the influencing factors.

[0021] Determine the splitting information for each impact factor, and determine the information gain ratio for each impact factor based on the splitting information and information gain;

[0022] The influence factor with the largest information gain ratio is used as the splitting feature of the root node, and the root node is split according to the splitting feature to obtain the child nodes of the root node.

[0023] The child nodes are recursively split until a preset stopping condition is met, resulting in a decision tree.

[0024] The classification performance of child nodes and subtrees composed of child nodes is determined, and the decision tree is pruned from bottom to top based on the classification performance to obtain the registration and matching model.

[0025] In a third implementation form of the first aspect, in the first implementation form of the first aspect, the point cloud registration algorithm of the to-be-processed point cloud data is determined by using the trained registration matching model, and the to-be-processed point cloud data is registered by using the point cloud registration algorithm to obtain the transformation matrix, and the method specifically comprises the following steps.

[0026] The to-be-processed point cloud data is subjected to feature extraction to obtain preset to-be-processed point cloud features;

[0027] The to-be-processed point cloud features are preprocessed, and the to-be-processed point cloud features are input into the registration matching model to obtain the point cloud registration algorithm of the to-be-processed point cloud data output by the registration matching model;

[0028] The to-be-processed point cloud data is registered by using the point cloud registration algorithm to obtain the transformation matrix of each frame of point cloud data;

[0029] In a case where the to-be-processed point cloud data fails to be registered by using the point cloud registration algorithm, the to-be-processed point cloud data is registered by using a preset algorithm to obtain the transformation matrix of each frame of point cloud data.

[0030] In a fourth implementation form of the first aspect, in the first implementation form of the first aspect, each sub-point cloud data is converted into a homogeneous coordinate matrix of a preset shape by using a parallel processing mode, and the homogeneous coordinate matrix is subjected to coordinate transformation by using the transformation matrix to obtain a restored point cloud matrix of each sub-point cloud data, and the method specifically comprises the following steps.

[0031] Each sub-point cloud data is converted into a homogeneous coordinate matrix of a preset shape by using a parallel processing mode;

[0032] The homogeneous coordinate matrix is subjected to matrix transposition to obtain a transposed coordinate matrix;

[0033] A corresponding sub-transformation matrix is extracted from the transformation matrix according to the point cloud data of the first frame and the last frame of the sub-point cloud data, the transposed coordinate matrix of the sub-point cloud data is subjected to matrix transformation by using the sub-transformation matrix to obtain a transformed coordinate matrix;

[0034] The transformed coordinate matrix is restored to the preset shape to obtain the restored point cloud matrix.

[0035] In a fifth implementation form of the first aspect, in the first implementation form of the first aspect, the number of the sub-point cloud data is determined, and the restored coordinate matrix is spliced according to the number to obtain the registered point cloud data, and the method specifically comprises the following steps.

[0036] All the restored point cloud matrices are collected;

[0037] The number of the sub-point cloud data is determined, and all the restored coordinate matrices are spliced into the restored point cloud matrix according to the number to obtain a spliced point cloud matrix;

[0038] Convert the spliced point cloud matrix into point cloud data to obtain the registered point cloud data.

[0039] In combination with the first aspect, in a sixth implementation form of the first aspect, the method further comprises the following steps:

[0040] The network address of the first frame of point cloud data of the to-be-processed point cloud data is determined, a hash value is generated according to the network address, the hash value is matched with the cache data stored in the preset database, and in the case of determining cache hit, the matched cache data is called from the preset database as the transformation matrix of the to-be-processed point cloud data; the hash value is used as the unique identifier of the to-be-processed point cloud data, and the preset database stores historical transformation matrices of several frames.

[0041] According to the second aspect, an embodiment of the present application further provides a point cloud registration device, the device comprising:

[0042] The registration request module is configured to obtain the point cloud registration request of the to-be-processed point cloud data and the to-be-processed task.

[0043] The point cloud registration module is configured to determine the point cloud registration algorithm of the to-be-processed point cloud data by using the trained registration matching model, and to perform registration on the to-be-processed point cloud data by using the point cloud registration algorithm to obtain the transformation matrix; the registration matching model is trained by using the decision tree algorithm.

[0044] The data segmentation module is configured to segment the to-be-processed point cloud data into a plurality of sub-point cloud data according to the point cloud registration request.

[0045] The coordinate transformation module is configured to convert each sub-point cloud data into a homogeneous coordinate matrix of a preset shape in a parallel processing manner, and to perform coordinate transformation on the homogeneous coordinate matrix by using the transformation matrix to obtain the restored point cloud matrix of each sub-point cloud data.

[0046] The matrix splicing module is configured to determine the number of the sub-point cloud data, splice the restored coordinate matrix according to the number, and obtain the registered point cloud data.

[0047] According to the third aspect, an embodiment of the present application further provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the point cloud registration method according to any one of the above when executing the program.

[0048] According to the fourth aspect, an embodiment of the present application further provides a non-transitory computer readable storage medium having a computer program stored thereon, wherein the computer program is executable by a processor to implement the steps of the point cloud registration method according to any one of the above.

[0049] The point cloud registration method, device, equipment and medium provided by the application determine the point cloud registration algorithm of the to-be-processed point cloud data by using the trained registration matching model, and then register the to-be-processed point cloud data by using the point cloud registration algorithm to obtain a transformation matrix. The registration matching model is trained by using a decision tree algorithm. The flexible switching of the point cloud registration algorithm of the to-be-processed point cloud data under the intelligent decision support improves the response speed and reliability of the overall system. The to-be-processed point cloud data of the main process is divided into a plurality of sub-point cloud data, and each sub-point cloud data is allocated to a sub-process for processing. When the sub-point cloud data is processed, a multi-process parallel processing mode can be adopted to fully utilize the advantages of the modern processor architecture, greatly improve the data processing efficiency, and further enhance the throughput. In this way, the parallel processing mode can be used to convert each sub-point cloud data into a homogeneous coordinate matrix of a preset shape and perform coordinate transformation on the homogeneous coordinate matrix by using the transformation matrix to obtain a restored point cloud matrix of each sub-point cloud data. One matrix multiplication by using the transformation matrix can perform correct transformation under the homogeneous coordinate. For large-scale point cloud processing, the processing efficiency can be significantly improved, and time-balanced point cloud registration is realized, and the processing efficiency is further improved. The application comprehensively uses a plurality of strategies to greatly shorten the entire registration period while maintaining high-quality output. BRIEF DESCRIPTION OF DRAWINGS

[0050] The features and advantages of the application will be more clearly understood through reference to the following drawings, which are presented for illustrative purposes only and are not intended to limit the application in any way. In the drawings:

[0051] Figure 1 Fig. 1 shows a flowchart of a point cloud registration method provided by the application;

[0052] Figure 2 Fig. 3 shows a schematic diagram of a multi-process parallel processing adopted in the point cloud registration method provided by the application;

[0053] Figure 3 Fig. 4 shows another flowchart of a point cloud registration method provided by the application;

[0054] Figure 4 Fig. 5 shows a structural schematic diagram of a point cloud registration device provided by the application;

[0055] Figure 5 Fig. 6 is a hardware structure schematic diagram of an electronic device provided by an embodiment of the application. DETAILED DESCRIPTION

[0056] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0057] Point cloud registration is one of the key technologies in the field of computer vision and robot navigation, and its main purpose is to align point cloud data from different perspectives or time points to achieve environment modeling, object recognition and other functions. With the development of applications such as autonomous driving and unmanned aerial vehicle navigation, higher requirements are put forward for the speed and robustness of point cloud registration algorithms.

[0058] However, the currently widely used point cloud registration algorithms such as ICP and its variants, although they perform well in accuracy, are inefficient in processing large-scale point cloud data, while some lightweight algorithms are fast but sacrifice part of the accuracy.

[0059] Specifically, the currently widely used point cloud registration methods include:

[0060] ICP algorithm: As one of the most basic point cloud registration algorithms, ICP algorithm is simple and easy to implement, but its convergence speed is slow and it is easy to fall into local optimal solution, especially when it faces point cloud data with large noise or serious partial occlusion.

[0061] Voxelized Generalized Iterative Closest Point (VGICP) algorithm: This is an improved ICP algorithm based on voxel grid, which can improve computational efficiency by reducing the dimension of search space. However, when dealing with large-scale point cloud data or high-density point cloud, the computational complexity of VGICP is still high, resulting in long running time.

[0062] Generalized ICP (GICP) algorithm: Compared with traditional ICP algorithm, GICP considers the local feature information of point cloud, which improves the registration accuracy. However, the enhanced accuracy of GICP is at the expense of computational efficiency.

[0063] It can be seen that the above point cloud registration algorithm often needs a long time to complete a complete registration process, the response time is long, especially when facing complex scenes or multiple consecutive images, which limits their applicability in real-time applications; secondly, in order to pursue fast processing speed, some optimized point cloud registration algorithms may reduce the tolerance to noise and outliers, resulting in lower robustness, thereby affecting the reliability of the final result, and the above point cloud registration algorithm lacks the ability to flexibly adjust its strategy according to specific application scenarios, and is difficult to cope with diversified working conditions.

[0064] In summary, how to shorten the response time while ensuring high accuracy of point cloud registration has become a problem to be solved in the industry.

[0065] Due to the above technical problems, a point cloud registration method is provided in the embodiments of the present application, which aims to greatly shorten the entire registration cycle while maintaining high-quality output. The point cloud registration method of the embodiments of the present application can be used in electronic devices, including but not limited to computers, mobile terminals, etc. Figure 1 is a flowchart of the point cloud registration method according to the embodiments of the present application, as shown in Figure 1 The method can include the following steps:

[0066] S101, obtaining point cloud data to be processed and a point cloud registration request of a to-be-processed task.

[0067] S102, determining a point cloud registration algorithm of the to-be-processed point cloud data by using a trained registration matching model, and performing registration on the to-be-processed point cloud data by using the point cloud registration algorithm to obtain a transformation matrix.

[0068] The to-be-processed point cloud data is new data that needs to be registered. In order to more effectively select a suitable point cloud registration algorithm for the to-be-processed point cloud data, in the embodiments of the present application, the registration matching model is trained by using a decision tree algorithm, that is, the to-be-processed point cloud features after preprocessing of the to-be-processed point cloud data are used as input data, the registration matching model processes the input data to obtain a predicted classification result, and the predicted classification result is used as output data and used to indicate the point cloud registration algorithm of the to-be-processed point cloud data.

[0069] The decision tree of the registration matching model starts from the root node of the decision tree when applied, and traverses layer by layer downward until reaching the leaf node. The leaf node represents a final predicted classification result, that is, the point cloud registration algorithm in the scene corresponding to the to-be-processed point cloud data, and further determines the final point cloud registration algorithm.

[0070] Specifically, step S102 includes the following steps:

[0071] S1021, performing feature extraction on the to-be-processed point cloud data to extract predetermined to-be-processed point cloud features.

[0072] The preset to-be-processed point cloud features specifically include but are not limited to:

[0073] Point cloud quantity : used for representing the point number and of the to-be-processed point cloud data, which can be determined by traversing the file of the to-be-processed point cloud data; frame number : used for representing the total frame number of the to-be-processed point cloud data; point cloud average density : Scene type: used for representing the collection scene of the to-be-processed point cloud data, such as driving, underground parking, viaduct, etc.

[0074] S1022, pre-processing the to-be-processed point cloud features, and inputting the to-be-processed point cloud features into the registration matching model to obtain a point cloud registration algorithm of the to-be-processed point cloud data output by the registration matching model. The point cloud registration algorithm stored in the registration matching model includes an SM-VGICP algorithm, an SM-GICP algorithm, an F-ICP algorithm, an ICP algorithm, a VGICP algorithm, a GICP algorithm, etc.

[0075] The pre-processing mode includes standardization and normalization. The standardization is a mean and variance standardization processing of the feature values of the to-be-processed point cloud features, and the normalization can convert each feature value into the range of [0, 1].

[0076] S1023, registering the to-be-processed point cloud data by using the point cloud registration algorithm to obtain a transformation matrix of each frame of point cloud data, the transformation matrix being an RT matrix, and each frame of point cloud data of the to-be-processed point cloud data having a corresponding RT matrix.

[0077] S1024, in the case that the point cloud registration algorithm fails to register the to-be-processed point cloud data, registering the to-be-processed point cloud data by using a preset algorithm to obtain a transformation matrix of each frame of point cloud data, the transformation matrix still being an RT matrix in this case, and each frame of point cloud data of the to-be-processed point cloud data having a corresponding RT matrix.

[0078] In some cases, the point cloud registration algorithm output by the registration matching model cannot register the to-be-processed point cloud data, and the to-be-processed point cloud data registration failure may occur. At this time, the preset algorithm responsible for registration backup is switched to, the to-be-processed point cloud data is registered by using the preset algorithm, and the transformation matrix can be ensured to be obtained, thereby avoiding the subsequent processing.

[0079] In the embodiment of the application, the preset algorithm is an ICP algorithm.

[0080] S103, according to the point cloud registration request, the to-be-processed point cloud data is divided into a plurality of sub-point cloud data, and the set of the sub-point cloud data is wherein, represents the th sub-point cloud data, and the to-be-processed point cloud data is divided into sub-point cloud data.

[0081] By dividing the to-be-processed point cloud data of the main process into several sub-point cloud data, each sub-point cloud data is allocated to a sub-process for processing. In the process of processing the sub-point cloud data, a multi-process parallel processing mode can be adopted. The corresponding processing of each sub-point cloud data is equivalent to a sub-task. In this way, the advantages of modern processor architecture are fully utilized, the data processing efficiency is greatly improved, and the throughput is further enhanced.

[0082] S104, each sub-point cloud data is converted into a homogeneous coordinate matrix of a preset shape and the homogeneous coordinate matrix is subjected to coordinate transformation by using a transformation matrix, to obtain a restored point cloud matrix of each sub-point cloud data. The preset shape is N*4. In this way, the three-dimensional point cloud data can be upgraded to homogeneous coordinate data. The homogeneous coordinate matrix is obtained, and the homogeneous coordinate matrix is processed and restored by using the RT matrix form transformation matrix, to obtain a restored point cloud matrix. The shape of the restored point cloud matrix is also the preset shape, i.e. N*4.

[0083] In this way, one matrix multiplication by using the transformation matrix can perform correct transformation under the homogeneous coordinate. For large-scale point cloud processing, the processing efficiency can be significantly improved. Please refer to Figure 2 When performing coordinate transformation, the to-be-processed point cloud data is divided into several sub-point cloud data. Therefore, a parallel processing mode can be adopted, and time-balanced point cloud registration is achieved, further improving the processing efficiency.

[0084] S105, the number of the sub-point cloud data is determined, and the restored coordinate matrix is spliced according to the number, to obtain the registered point cloud data.

[0085] When the to-be-processed point cloud data is divided into several sub-point cloud data, a corresponding number can be set for each sub-point cloud data, to determine the order of the several sub-point cloud data in matrix splicing.

[0086] In the process of obtaining the final required registered point cloud data, an adaptive intelligent decision mechanism based on a decision tree is introduced to select a suitable point cloud registration algorithm in combination with a multi-process parallel processing mode. The comprehensive use of multiple strategies can greatly shorten the entire registration period while maintaining high-quality output.

[0087] The point cloud registration method of this invention utilizes a trained registration and matching model to determine the point cloud registration algorithm for the point cloud data to be processed. This algorithm is then used to register the point cloud data, resulting in a transformation matrix. The registration and matching model is trained using a decision tree algorithm. By flexibly switching the point cloud registration algorithm for each piece of point cloud data under intelligent decision support, the overall system's response speed and reliability are improved. The method divides the point cloud data to be processed in the main process into several sub-point cloud data, each of which is assigned to a sub-process for processing. Multi-process parallel processing can be adopted when processing sub-point cloud data, fully utilizing the advantages of modern processor architecture, significantly improving data processing efficiency, and further enhancing throughput. Subsequently, parallel processing can be used to convert each sub-point cloud data into a homogeneous coordinate matrix of a preset shape and perform coordinate transformation on the homogeneous coordinate matrix using the transformation matrix to obtain the restored point cloud matrix for each sub-point cloud data. A single matrix multiplication using the transformation matrix is ​​sufficient for correct transformation in homogeneous coordinates. For large-scale point cloud processing, this significantly improves processing efficiency and achieves time-balanced point cloud registration, further enhancing processing efficiency. This invention utilizes a combination of strategies to significantly shorten the entire registration cycle while maintaining high-quality output.

[0088] In this embodiment of the application, step S104 specifically includes:

[0089] S1041. Using parallel processing, each sub-point cloud data is converted into a homogeneous coordinate matrix of a preset shape. Specifically:

[0090]

[0091] in, Indicates the first Homogeneous coordinate matrix of individual point cloud data; Indicates the first The x-values ​​of the coordinate system for each point cloud data; Indicates the first The coordinate system y-values ​​of a point cloud data; Indicates the first The z-value of the coordinate system of the point cloud data.

[0092] The last column is a vector of 1s, which allows 3D point cloud data to be upscaled to homogeneous coordinate data.

[0093] S1042. Transpose the aligned coordinate matrix to obtain the transposed coordinate matrix. Specifically:

[0094]

[0095] in, Indicates the first The transpose coordinate matrix of the sub-point cloud data. Data of the same dimension of the transpose coordinate matrix is ​​stored contiguously in memory. When performing matrix multiplication later, the system can efficiently prefetch these contiguous data blocks into the cache, greatly improving computational efficiency.

[0096] S1043. Based on the point cloud data of the first and last frames of the sub-point cloud data, extract the corresponding sub-transformation matrix from the transformation matrix, and use the sub-transformation matrix to perform matrix transformation on the transpose coordinate matrix of the sub-point cloud data to obtain the transformed coordinate matrix. Specifically:

[0097]

[0098] in, Indicates the first Transformation coordinate matrix of individual point cloud data; Show the first Sub-transformation matrix of a point cloud data.

[0099] The transformation matrix contains the RT matrix of each frame of point cloud data to be processed. The sub-transformation matrix of the corresponding frame of the sub-point cloud data can be extracted from the total transformation matrix by using the first and last frames of the sub-point cloud data.

[0100] S1044. Restore the transformed coordinate matrix back to the preset shape to obtain the restored point cloud matrix. Specifically:

[0101]

[0102] in, Indicates the first The point cloud matrix is ​​reconstructed from the individual point cloud data; correspondingly... Indicates the restored first... The x-value of the coordinate system of a point cloud data. Indicates the restored first... The y-values ​​of the coordinate system of a point cloud data. Indicates the restored first... The coordinate system z-values ​​of the point cloud data. The shape of the transformed coordinate matrix is ​​4×N. The transformed coordinate matrix is ​​transposed to the original shape through restoration processing to obtain the restored point cloud matrix.

[0103] In this embodiment of the application, step S105 includes the following steps:

[0104] S1051. Gather all the restored point cloud matrices. By gathering the data results returned by all sub-processes, the restored point cloud matrix after a series of processing of all sub-point cloud data is obtained.

[0105] S1052, determine the number of sub-point cloud data, and perform splicing on all the restored coordinate matrices according to the number to obtain a spliced point cloud matrix.

[0106]

[0107] wherein, represents the spliced point cloud matrix obtained by splicing all the restored coordinate matrices.

[0108] S1053, convert the spliced point cloud matrix into point cloud data to obtain the registered point cloud data.

[0109] Since the last column of each of the restored coordinate matrices is a 1 vector, the spliced point cloud matrix needs to be converted into a three-dimensional point cloud data, so as to obtain the required registered point cloud data.

[0110] In the embodiments of the present application, the registration matching model is trained by the following steps:

[0111] S201, obtain historical point cloud data registered under different scenes, and determine a point cloud registration algorithm used by the historical point cloud data.

[0112] In the embodiments of the present application, the point cloud registration algorithm used by the historical point cloud data includes: SM-VGICP algorithm, SM-GICP algorithm, F-ICP algorithm, ICP algorithm, VGICP algorithm and GICP algorithm. It can be understood that the point cloud registration algorithm used by the historical point cloud data can not be limited to the above-mentioned several algorithms.

[0113] wherein, the prefixes SM (small) and F (fast) represent light-weight implementation algorithms of the corresponding algorithms, and the point cloud registration algorithm used by the historical point cloud data includes light-weight algorithms and standard algorithms.

[0114] S202, perform feature extraction on the historical point cloud data to extract preset historical point cloud features.

[0115] Similarly, the preset historical point cloud features specifically include but are not limited to:

[0116] point cloud number : used to represent the point number of the historical point cloud data, which can be determined by traversing the file of the historical point cloud data; frame number : used to represent the total frame number of the historical point cloud data; point cloud average density : ; scene type: used to represent the collection scene of the historical point cloud data, such as driving, underground parking, viaduct, etc.

[0117] S203, pre-process the historical point cloud features, wherein the pre-processing includes standardization and normalization; the standardization is a mean and variance standardization processing of the feature values of the historical point cloud features; and the normalization can convert each feature value to a range of [0, 1].

[0118] These historical point cloud features are pre-processed to ensure that they have the same weight in subsequent decision tree construction.

[0119] S204, taking the historical point cloud features as impact factors, establishing a mapping relationship between the impact factors of the historical point cloud data and the point cloud registration algorithm, and obtaining historical training data of the historical point cloud data. It can be understood that each historical point cloud data can obtain corresponding historical training data.

[0120] Among them, the point cloud registration algorithm can be used as the target value of the impact factor.

[0121] S205, constructing a decision tree using the historical training data, and obtaining a registration matching model based on the constructed decision tree, wherein the registration matching model is used to obtain a point cloud registration algorithm of the to-be-processed point cloud data.

[0122] In the embodiments of the present application, the registration matching model is constructed based on the C4.5 decision tree algorithm. More specifically, step S205 includes the following steps:

[0123] S2051, collecting all the historical training data to obtain a historical training set, placing the historical training set on the root node of the decision tree, and calculating the information entropy of the root node. The calculation formula of the information entropy is:

[0124]

[0125] Among them, represents the information entropy of the historical training set , and the historical training set can be obtained by collecting each historical point cloud data; represents the probability of the th point cloud registration algorithm in the historical training set, and there are point cloud registration algorithms in the historical training set, and each point cloud registration algorithm is a category.

[0126] S2052, traversing each impact factor, determining the conditional entropy of each impact factor, and determining the information gain of each impact factor according to the information entropy of the root node and the conditional entropy of the impact factor. The calculation formula of the information gain is:

[0127]

[0128] Among them, represents the information gain of the impact factor information gain of the influence factor information entropy of the historical training set under the influence factor , i.e. conditional entropy of the influence factor , specifically, The calculation formula of the influence factor

[0129]

[0130] , wherein, represents the i-th sample subset, and the historical training set can be divided into sample subsets according to the sample attributes of the influence factor ; represents the number of samples in the i-th sample subset, represents the number of samples in the i-th sample subset, represents the number of samples in the i-th sample subset, represents the number of samples in the i-th sample subset.

[0131] S2053, determine the split information of each influence factor, and determine the information gain rate of each influence factor according to the split information and the information gain. The calculation formula of the information gain rate is:

[0132]

[0133] , wherein, represents the information gain rate of the influence factor , i.e. information gain ratio; represents the split information (Split Information) of the influence factor , i.e. inherent information amount, specifically, The calculation formula of the influence factor

[0134]

[0135] S2054, the influence factor with the maximum information gain rate is taken as the split feature of the root node, and the root node is split according to the split feature to obtain the child nodes of the root node.

[0136] S2055, recursively split the child nodes until the preset stop condition is met to obtain the decision tree.

[0137] In the C4.5 decision tree algorithm, the impact factor with the maximum information gain ratio is selected as the split feature of the root node when performing the first split, and after the first split is completed, each sub-node obtained by the split is recursively split. The difference between the sub-node and its parent node in the split is that the impact factor with the maximum information gain ratio needs to be selected in addition to the impact factor selected by the parent node, and the impact factor is taken as the split feature of the sub-node. The above process is repeated for each sub-node, and the sub-tree is recursively called to build the sub-tree until the preset stopping condition (such as the number of samples in the node being less than the preset threshold or the information gain ratio of all impact factors being less than the preset gain ratio) is met.

[0138] It should be noted that if the information gain ratios of multiple features are equal, one of the impact factors can be randomly selected as the split feature.

[0139] S2056, determine the classification effect of the sub-node and the sub-tree composed of the sub-node, and prune the decision tree from bottom to top according to the classification effect to obtain a registration matching model.

[0140] In order to avoid overfitting, the C4.5 decision tree algorithm adopts a post-pruning method, that is, after the decision tree is constructed, the decision tree is pruned from bottom to top. In the pruning process, the classification effect of the sub-tree is compared with the classification effect of the single sub-node, and if the sub-tree is not better than the single sub-node, the sub-tree is pruned.

[0141] Please refer to Figure 3 In the embodiment of the present application, the method can further include the following steps:

[0142] S301, obtaining point cloud data to be processed and a point cloud registration request of a task to be processed. The specific content is as shown in Figure 1 as described in step S101, which will not be repeated here.

[0143] S302, determining the network address of the first frame of point cloud data of the point cloud data to be processed, generating a hash value according to the network address, matching the hash value with the cache data stored in the preset database, and in the case of determining cache hit, calling the matching cache data from the preset database as the transformation matrix of the point cloud data to be processed.

[0144] The network address is a uniform resource locator (URL) address. Preferably, the hash value can use MD5 hash value, of course, other types of hash values can also be used. Each time the point cloud registration of the point cloud data to be processed is performed, a string of hash values will be generated for the corresponding batch of tasks, and the hash value is also the unique identifier of the batch of tasks.

[0145] The method also sets a cache mechanism for storing cache data to reduce the time of repeated calculation and improve the overall processing efficiency. The cache data is the historical transformation matrix. A hash value generated based on the URL address of the point cloud data to be processed is used to query whether there is a corresponding cache record in the preset database. If the cache hits, the RT matrix of each frame of point cloud data is obtained from the Hash structure to obtain the transformation matrix.

[0146] The historical cache structure is stored in the Hash type of Redis, and the specific content is as follows:

[0147] Key: Hash value, such as MD5 hash value; Value: Hash type, storing the RT matrix of each frame of point cloud data.

[0148] For example:

[0149] HMSET<md5_hash>frame1 "RT_matrix_1" frame2 "RT_matrix_2"...

[0150] The preset database stores several frames of historical transformation matrices. The preset database is a Redis database. When configuring the preset database, first install and configure the Redis server, and then set the connection pool to ensure stability and performance under high concurrency.

[0151] S303, in the case of determining that the cache does not hit, using the trained registration matching model to determine the point cloud registration algorithm of the point cloud data to be processed, and using the point cloud registration algorithm to register the point cloud data to be processed to obtain the transformation matrix. The specific content of determining the point cloud registration algorithm by the registration matching model and performing registration is as shown in Figure 1 S102, which will not be repeated here.

[0152] After obtaining the transformation matrix, the transformation matrix can be saved as a historical transformation matrix in the preset database for subsequent reuse. The preset database stores the newly stored historical transformation matrix as cache data. When there is a new task to be processed for point cloud registration, step S302 is executed again. If the cache hits, the RT matrix in the cache is loaded. By introducing an efficient cache mechanism, the workload of repeated calculation is greatly reduced, and the processing process is accelerated.

[0153] S304, according to the point cloud registration request, the point cloud data to be processed is divided into several sub-point cloud data. The specific content is as shown in Figure 1 S103, which will not be repeated here.

[0154] S305, each sub-point cloud data is converted into a homogeneous coordinate matrix of a preset shape in a parallel processing manner, and the homogeneous coordinate matrix is subjected to coordinate transformation by using a transformation matrix to obtain a restored point cloud matrix of each sub-point cloud data. Details are as shown in Figure 1 S104, which will not be repeated here.

[0155] S306, the number of the sub-point cloud data is determined, and the restored coordinate matrix is spliced according to the number to obtain the registered point cloud data. Details are as shown in Figure 1 S105, which will not be repeated here.

[0156] The point cloud registration device provided in the embodiments of the present application will be described below. The point cloud registration device described below can be referred to in correspondence with the point cloud registration method described above.

[0157] Due to the above technical problems, in the embodiments of the present application, a point cloud registration device is also provided, which aims to greatly shorten the entire registration period while maintaining high-quality output. Figure 4 is a structural schematic diagram of the point cloud registration method according to the embodiments of the present application, as Figure 4 shown, the device can include:

[0158] The registration request module 10 is configured to obtain point cloud registration requests of to-be-processed point cloud data and to-be-processed tasks.

[0159] The point cloud registration module 20 is configured to determine a point cloud registration algorithm of the to-be-processed point cloud data by using the trained registration matching model, and perform registration on the to-be-processed point cloud data by using the point cloud registration algorithm to obtain a transformation matrix.

[0160] As new data that needs to be subjected to point cloud registration, in order to more effectively select a suitable point cloud registration algorithm for the to-be-processed point cloud data, in the embodiments of the present application, the registration matching model is trained by using a decision tree algorithm, that is, to-be-processed point cloud features after preprocessing of the to-be-processed point cloud data are taken as input data, the registration matching model processes the input data to obtain a predicted classification result, and the predicted classification result is taken as output data and used to indicate the point cloud registration algorithm of the to-be-processed point cloud data.

[0161] The decision tree of the registration matching model starts from a root node when applied, and iteratively traverses downward layer by layer until a leaf node is reached. The leaf node represents a final predicted classification result, that is, a point cloud registration algorithm in a scene corresponding to the to-be-processed point cloud data, and further determines the final point cloud registration algorithm.

[0162] The data segmentation module 30 is configured to segment the to-be-processed point cloud data into a plurality of sub-point cloud data according to the point cloud registration request.

[0163] By dividing the to-be-processed point cloud data of the main process into a plurality of sub-point cloud data, each sub-point cloud data is allocated to a sub-process for processing, and a multi-process parallel processing mode can be adopted when processing the sub-point cloud data. The corresponding processing of each sub-point cloud data is equivalent to a sub-task, so as to fully utilize the advantages of the modern processor architecture, greatly improve the data processing efficiency, and further enhance the throughput.

[0164] The coordinate transformation module 40 is configured to convert each sub-point cloud data into a homogeneous coordinate matrix of a preset shape by using parallel processing, perform coordinate transformation on the homogeneous coordinate matrix by using a transformation matrix, and obtain a restored point cloud matrix of each sub-point cloud data. The preset shape is N x 4. In this way, the three-dimensional point cloud data can be upgraded to homogeneous coordinate data, and the homogeneous coordinate matrix is obtained. Then, the homogeneous coordinate matrix is processed and restored by using the transformation matrix in the form of an RT matrix, and the restored point cloud matrix is obtained. The shape of the restored point cloud matrix is also the preset shape, i.e., N x 4.

[0165] In this way, one matrix multiplication by using the transformation matrix can perform correct transformation under the homogeneous coordinate. For large-scale point cloud processing, the processing efficiency can be significantly improved. Please refer to Figure 2 When performing coordinate transformation, the to-be-processed point cloud data is divided into a plurality of sub-point cloud data, so that the parallel processing mode can be adopted, and the processing efficiency is further improved.

[0166] The matrix splicing module 50 is configured to determine the number of the sub-point cloud data, splice the restored coordinate matrix according to the number, and obtain the registered point cloud data.

[0167] In the process of obtaining the final required registered point cloud data, an adaptive intelligent decision mechanism based on a decision tree is introduced to select a suitable point cloud registration algorithm, and a multi-process parallel processing mode is combined. A plurality of strategies are comprehensively used to greatly shorten the entire registration period while maintaining high-quality output.

[0168] The point cloud registration device of the application determines the point cloud registration algorithm of the to-be-processed point cloud data by using the trained registration matching model, and then registers the to-be-processed point cloud data by using the point cloud registration algorithm to obtain a transformation matrix. The registration matching model is trained by using a decision tree algorithm. The flexible switching of the point cloud registration algorithm of the to-be-processed point cloud data under the intelligent decision support improves the response speed and reliability of the overall system. The to-be-processed point cloud data of the main process is divided into a plurality of sub-point cloud data, and each sub-point cloud data is allocated to a sub-process for processing. The multi-process parallel processing mode can be adopted when the sub-point cloud data is processed, so as to fully utilize the advantages of the modern processor architecture, greatly improve the data processing efficiency, and further enhance the throughput. In this way, the parallel processing mode can be adopted to convert each sub-point cloud data into a homogeneous coordinate matrix of a preset shape, and the homogeneous coordinate matrix is subjected to coordinate transformation by using the transformation matrix to obtain a restored point cloud matrix of each sub-point cloud data. The correct transformation under the homogeneous coordinate can be performed by using the matrix multiplication of the transformation matrix once. For large-scale point cloud processing, the processing efficiency can be significantly improved, and time-balanced point cloud registration is realized, and the processing efficiency is further improved. The application comprehensively uses a plurality of strategies to greatly shorten the entire registration period while maintaining high-quality output.

[0169] Figure 5 An example of a schematic diagram of a physical structure of an electronic device is shown in Figure 5 The electronic device can include a processor 510, a communications interface 520, a memory 530, and a communications bus 540, wherein the processor 510, the communications interface 520, and the memory 530 communicate with each other through the communications bus 540. The processor 510 can invoke a logical command in the memory 530 to execute a point cloud registration method, which includes:

[0170] Obtaining to-be-processed point cloud data and a point cloud registration request of a to-be-processed task;

[0171] Determining the point cloud registration algorithm of the to-be-processed point cloud data by using a trained registration matching model, and registering the to-be-processed point cloud data by using the point cloud registration algorithm to obtain a transformation matrix; the registration matching model is trained by using a decision tree algorithm;

[0172] Dividing the to-be-processed point cloud data into a plurality of sub-point cloud data according to the point cloud registration request;

[0173] Converting each sub-point cloud data into a homogeneous coordinate matrix of a preset shape by using a parallel processing mode, and performing coordinate transformation on the homogeneous coordinate matrix by using the transformation matrix to obtain a restored point cloud matrix of each sub-point cloud data;

[0174] determining the number of the sub-point cloud data, splicing the restored coordinate matrix according to the number, and obtaining the registered point cloud data.

[0175] In addition, the logic instructions in the memory 530 described above can be implemented in the form of a software function unit and sold or used as an independent product, and can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0176] On the other hand, the present application also provides a computer program product, which comprises a computer program stored on a non-transitory computer readable storage medium, and the computer program comprises program instructions, when the program instructions are executed by a computer, the computer can execute the point cloud registration method provided by the above-mentioned method, and the method comprises:

[0177] obtaining point cloud registration request of to-be-processed point cloud data and to-be-processed task;

[0178] determining a point cloud registration algorithm of the to-be-processed point cloud data by using the trained registration matching model, and performing registration on the to-be-processed point cloud data by using the point cloud registration algorithm to obtain a transformation matrix; the registration matching model is trained by using a decision tree algorithm;

[0179] segmenting the to-be-processed point cloud data into a plurality of sub-point cloud data according to the point cloud registration request;

[0180] Converting each sub-point cloud data into a homogeneous coordinate matrix of a preset shape by using a parallel processing mode, and performing coordinate transformation on the homogeneous coordinate matrix by using the transformation matrix to obtain a restored point cloud matrix of each sub-point cloud data;

[0181] determining the number of the sub-point cloud data, splicing the restored coordinate matrix according to the number, and obtaining the registered point cloud data.

[0182] In another aspect, the present application also provides a non-transitory computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the point cloud registration method provided by the above-mentioned method, and the method comprises:

[0183] acquire point cloud registration request of to-be-processed point cloud data and to-be-processed task;

[0184] determine point cloud registration algorithm of to-be-processed point cloud data by using the trained registration matching model, and perform registration on to-be-processed point cloud data by using the point cloud registration algorithm to obtain a transformation matrix; the registration matching model is obtained by training using a decision tree algorithm;

[0185] segment to-be-processed point cloud data into a plurality of sub-point cloud data according to the point cloud registration request;

[0186] convert each sub-point cloud data into a homogeneous coordinate matrix of a preset shape by using a parallel processing mode, and perform coordinate transformation on the homogeneous coordinate matrix by using the transformation matrix to obtain a restored point cloud matrix of each sub-point cloud data;

[0187] determine the number of the sub-point cloud data, and splice the restored coordinate matrix according to the number to obtain the registered point cloud data.

[0188] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing examples, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A point cloud registration method, characterized in that, The method comprises: acquiring point cloud data to be processed and a point cloud registration request of a task to be processed; determining a point cloud registration algorithm of the point cloud data to be processed by using a trained registration matching model, and performing registration on the point cloud data to be processed by using the point cloud registration algorithm to obtain a transformation matrix; the registration matching model is trained by using a decision tree algorithm; segmenting the point cloud data to be processed into a plurality of sub-point cloud data according to the point cloud registration request; in a parallel processing manner, converting each sub-point cloud data into a homogeneous coordinate matrix of a preset shape and performing coordinate transformation on the homogeneous coordinate matrix by using the transformation matrix to obtain a restored point cloud matrix of each sub-point cloud data; determining the number of the sub-point cloud data, and splicing the restored point cloud matrix according to the number to obtain point cloud data after registration; the registration matching model is trained by the following steps: acquiring historical point cloud data after registration in different scenes, and determining a point cloud registration algorithm used by the historical point cloud data; extracting preset historical point cloud features from the historical point cloud data; preprocessing the historical point cloud features; the preprocessing manner comprises standardization and normalization; taking the historical point cloud features as influence factors, establishing a mapping relationship between the influence factors and the point cloud registration algorithm of the historical point cloud data, and obtaining historical training data of the historical point cloud data; constructing a decision tree by using the historical training data, and obtaining the registration matching model based on the constructed decision tree; the registration matching model is used to obtain the point cloud registration algorithm of the point cloud data to be processed; the registration matching model is constructed by using the historical training data, and the registration matching model is obtained based on the constructed decision tree, and specifically comprises: collecting all the historical training data to obtain a historical training set, placing the historical training set on a root node of the decision tree, and calculating information entropy of the root node; determining conditional entropy of each influence factor by traversing each influence factor, and determining information gain of each influence factor according to the information entropy of the root node and the conditional entropy of the influence factor; determining split information of each influence factor, and determining information gain rate of each influence factor according to the split information and the information gain; taking the influence factor with the maximum information gain rate as a split feature of the root node, and splitting the root node according to the split feature to obtain a child node of the root node; recursively splitting the child node until a preset stop condition is met to obtain the decision tree; determining classification effects of the child node and a sub-tree composed of the child node, and pruning the decision tree from bottom to top according to the classification effects to obtain the registration matching model.

2. The point cloud registration method of claim 1, wherein, the registration matching model is constructed by using the historical training data, and the registration matching model is obtained based on the constructed decision tree, and specifically comprises: extracting preset point cloud features of the point cloud data to be processed from the point cloud data to be processed; preprocessing the point cloud features of the point cloud data to be processed, and inputting the point cloud features of the point cloud data to be processed into the registration matching model to obtain a point cloud registration algorithm of the point cloud data to be processed output by the registration matching model; performing registration on the point cloud data to be processed by using the point cloud registration algorithm to obtain a transformation matrix of each frame of point cloud data; In a case where it is determined that the point cloud registration algorithm fails to register the to-be-processed point cloud data, the to-be-processed point cloud data is registered by using a preset algorithm to obtain a transformation matrix of each frame of point cloud data.

3. The point cloud registration method of claim 1, wherein, The parallel processing is adopted to convert each sub-point cloud data into a homogeneous coordinate matrix of a preset shape, and the coordinate transformation is performed on the homogeneous coordinate matrix by using the transformation matrix to obtain a restored point cloud matrix of each sub-point cloud data, and the method specifically comprises the following steps: The parallel processing is adopted to convert each sub-point cloud data into a homogeneous coordinate matrix of a preset shape; The matrix transposition is performed on the homogeneous coordinate matrix to obtain a transposed coordinate matrix; The corresponding sub-transformation matrix is extracted from the transformation matrix according to the point cloud data of the first frame and the last frame of the sub-point cloud data, the matrix transformation is performed on the transposed coordinate matrix of the sub-point cloud data by using the sub-transformation matrix to obtain a transformed coordinate matrix; The transformed coordinate matrix is restored to the preset shape to obtain the restored point cloud matrix.

4. The point cloud registration method of claim 1, wherein, The method further comprises the following steps: All the restored point cloud matrices are collected; The number of the sub-point cloud data is determined, and the restored point cloud matrices are spliced according to the number to obtain registered point cloud data. The method further comprises the following steps:

5. The point cloud registration method of claim 1, wherein, The network address of the first frame of point cloud data of the to-be-processed point cloud data is determined, a hash value is generated according to the network address, the hash value is matched with the cache data stored in the preset database, in a case where cache hit is determined, the matched cache data is called from the preset database as the transformation matrix of the to-be-processed point cloud data; the hash value is used as the unique identifier of the to-be-processed point cloud data, and a plurality of historical transformation matrices are stored in the preset database. The device comprises:

6. A point cloud registration apparatus characterized by comprising: A registration request module is configured to acquire to-be-processed point cloud data and a point cloud registration request of a to-be-processed task; A point cloud registration module is configured to determine a point cloud registration algorithm of the to-be-processed point cloud data by using a trained registration matching model, and to register the to-be-processed point cloud data by using the point cloud registration algorithm to obtain a transformation matrix; the registration matching model is trained by using a decision tree algorithm; A data segmentation module is configured to segment the to-be-processed point cloud data into a plurality of sub-point cloud data according to the point cloud registration request; A coordinate transformation module is configured to convert each sub-point cloud data into a homogeneous coordinate matrix of a preset shape by using parallel processing, and to perform coordinate transformation on the homogeneous coordinate matrix by using a transformation matrix to obtain a restored point cloud matrix of each sub-point cloud data; A matrix splicing module is configured to determine the number of the sub-point cloud data, and to splice the restored point cloud matrices according to the number to obtain registered point cloud data. The registration matching model is trained by using the following steps: Historical point cloud data completed by registration in different scenes is acquired, and a point cloud registration algorithm used by the historical point cloud data is determined; Feature extraction is performed on the historical point cloud data to extract preset historical point cloud features; Preprocessing is performed on the historical point cloud features; the preprocessing mode comprises standardization and normalization. ​ The historical point cloud features are taken as influence factors, a mapping relationship between influence factors of historical point cloud data and point cloud registration algorithms is established, and historical training data of the historical point cloud data is obtained; A decision tree is constructed using the historical training data, and a registration matching model is obtained based on the constructed decision tree; the registration matching model is used to obtain a point cloud registration algorithm of to-be-processed point cloud data; The decision tree is constructed using the historical training data, and the registration matching model is obtained based on the constructed decision tree, specifically including: All historical training data is collected to obtain a historical training set, the historical training set is placed on a root node of a decision tree, and information entropy of the root node is calculated; Each influence factor is traversed to determine conditional entropy of each influence factor, and information gain of each influence factor is determined according to the information entropy of the root node and the conditional entropy of the influence factor; Splitting information of each influence factor is determined, and information gain rate of each influence factor is determined according to the splitting information and the information gain; An influence factor with the maximum information gain rate is taken as a splitting feature of the root node, and the root node is split according to the splitting feature to obtain a child node of the root node; The child node is recursively split until a preset stop condition is met, and a decision tree is obtained; Classification effects of the child node and a sub-tree composed of the child node are determined, and the decision tree is pruned from bottom to top according to the classification effects to obtain a registration matching model.

7. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor implements the steps of the point cloud registration method according to any one of claims 1 to 5 when executing the program.

8. A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program implements the steps of the point cloud registration method according to any one of claims 1 to 5 when executed by the processor.

Images