Splicing method and device for point cloud data of automatic driving vehicle

Abstract

The invention provides an automatic driving vehicle point cloud data splicing method and device, and the method comprises the steps: obtaining source point cloud data collected by a laser radar for a target scene based on the target point cloud data of the target scene in an analog simulation test when a vehicle in an automatic driving mode is subjected to the analog simulation test; based on the source point cloud data and the target point cloud data, determining a corresponding similar transformation matrix between the source point cloud data and the target point cloud data by using a feature model; based on the similar transformation matrix, registering a plurality of corresponding point cloud pairs between the source point cloud data and the target point cloud data, and determining a target change matrix of each point cloud pair; and based on the target change matrix, splicing and fusing the source point cloud data and the target point cloud data to obtain a three-dimensional model corresponding to the target scene. Through the method, the accuracy, efficiency and stability of splicing the point cloud data of the autonomous vehicle are improved.

Description

Technical Field

[0001] The present application relates to the field of autonomous driving point cloud technology, and in particular to a method and device for splicing point cloud data of an autonomous driving vehicle. Background Art

[0002] Point cloud data, as a collection of points in three-dimensional space, can intuitively represent the shape and structure of an object. It is an important data form for three-dimensional reconstruction during autonomous driving. Point cloud stitching technology is the core step in integrating the collected point cloud data into a three-dimensional model. The accuracy and stability of point cloud stitching directly affect the effect and application of three-dimensional reconstruction.

[0003] Currently, when stitching point cloud data collected by autonomous vehicles, the collected source point cloud data contains certain noise and outliers, and the source point cloud data and the point cloud data constructed by the vehicle do not have a unified coordinate system standard, which reduces the stability of the point cloud data stitching; in addition, the traditional point cloud stitching and registration method usually includes coarse registration and fine registration, but this method makes the stitched point cloud data inaccurate and inefficient. Summary of the Invention

[0004] In view of this, the purpose of the present application is to provide a method and device for splicing point cloud data of an autonomous driving vehicle. When conducting a simulation test on a vehicle in an autonomous driving mode, the target point cloud data corresponding to the target scene is used as a reference to obtain the source point cloud data collected by the laser radar for the target scene, and the similarity transformation matrix corresponding to the source point cloud data and the target point cloud data is determined by using a pre-trained feature model. The source point cloud data and the target point cloud data are aligned to determine the target change matrix, and then based on the target change matrix, the source point cloud data and the target point cloud data are fused to obtain a three-dimensional model corresponding to the target scene, thereby improving the accuracy and efficiency of splicing point cloud data of autonomous driving vehicles, and thereby improving the stability of point cloud data splicing.

[0005] The present application provides a method for splicing point cloud data of an autonomous driving vehicle, the method comprising:

[0006] When conducting a simulation test on a vehicle in an autonomous driving mode, obtaining source point cloud data of a target scene in the simulation test, collected by a laser radar disposed in the vehicle, based on preset target point cloud data corresponding to the target scene;

[0007] Based on the source point cloud data and the target point cloud data, using a pre-trained feature model, determining a similarity transformation matrix corresponding to the source point cloud data and the target point cloud data;

[0008] Based on the similarity transformation matrix, registering a plurality of corresponding point cloud pairs between the source point cloud data and the target point cloud data, and determining a target change matrix for each of the point cloud pairs;

[0009] Based on the target change matrix, the source point cloud data and the target point cloud data are spliced ​​and fused to obtain a three-dimensional model corresponding to the target scene.

[0010] Furthermore, the target scene in the simulation test corresponds to preset target point cloud data, and the source point cloud data collected by the laser radar disposed in the vehicle for the target scene is obtained, including:

[0011] Obtaining original source point cloud data collected by a laser radar installed in the vehicle for a target scene in a simulation test;

[0012] Performing denoising and filtering processing on the original source point cloud data to obtain source point cloud data to be processed;

[0013] Based on preset target point cloud data corresponding to the target scene, the source point cloud data to be processed is normalized to obtain source point cloud data corresponding to the target scene.

[0014] Furthermore, the determining of the similarity transformation matrix corresponding to the source point cloud data and the target point cloud data by using a pre-trained feature model based on the source point cloud data and the target point cloud data includes:

[0015] Inputting the source point cloud data and the target point cloud data into a pre-trained local feature extraction model respectively, and obtaining a plurality of local features corresponding to the target scene output by the local feature extraction model;

[0016] Inputting the source point cloud data and the target point cloud data into a pre-trained global feature extraction model respectively, and obtaining a plurality of global features corresponding to the target scene output by the global feature extraction model;

[0017] Based on the local features and the global features, a similarity transformation matrix corresponding to the source point cloud data and the target point cloud data is calculated and determined.

[0018] Furthermore, the local feature extraction model is pre-trained by the following steps:

[0019] Selecting training data from a preset training data set according to a preset data volume, performing edge computing labeling on the training data corresponding to the preset data volume to obtain a labeled training data set, and determining the unselected training data in the training data set as an unlabeled training data set;

[0020] Using the labeled training data set to train the first local feature extraction model to be trained to obtain a second local feature extraction model to be trained;

[0021] Inputting each unlabeled training data in the unlabeled training data set into the second local feature extraction model to be trained, determining the prediction confidence of each unlabeled training data, and numerically comparing the prediction confidence with a preset confidence threshold to obtain a comparison result;

[0022] The unlabeled training data whose comparison result shows that the prediction confidence is greater than the preset confidence threshold is added to the labeled training data set to obtain a target training data set;

[0023] Using the target training data set to train the second local feature extraction model to be trained to obtain a third local feature extraction model to be trained;

[0024] Performing point cloud reconstruction verification on the third to-be-trained local feature extraction model using a preset verification data set to obtain a point cloud reconstruction result output by the third to-be-trained local feature extraction model;

[0025] Based on the point cloud reconstruction result and the verification data set, parameters of the third local feature extraction model to be trained are updated to obtain the local feature extraction model.

[0026] Furthermore, the calculating and determining a similarity transformation matrix corresponding to the source point cloud data and the target point cloud data based on the local features and the global features includes:

[0027] Determining a local difference item between the source point cloud data and the target point cloud data under the local feature, and determining a global difference item between the source point cloud data and the target point cloud data under the global feature;

[0028] For each of the local features, determining first source point cloud data corresponding to the source point cloud data under the local feature, and determining first target point cloud data corresponding to the target point cloud data under the local feature;

[0029] For each of the global features, determining second source point cloud data corresponding to the source point cloud data under the global feature, and determining second target point cloud data corresponding to the target point cloud data under the global feature;

[0030] Determining a transformation matrix of the source point cloud data relative to the target point cloud data using a preset transformation function based on the local difference term, the global difference term, the first source point cloud data, the second source point cloud data, the first target point cloud data, and the second target point cloud data;

[0031] Based on the transformation matrix, the source point cloud data and the target point cloud data, a similarity transformation matrix corresponding to the source point cloud data and the target point cloud data is determined using a preset mutual information transformation formula.

[0032] Furthermore, the registering of a plurality of corresponding point cloud pairs between the source point cloud data and the target point cloud data based on the similarity transformation matrix to determine a target change matrix for each point cloud pair includes:

[0033] Based on the source point cloud data and the target point cloud data, determining initial change matrices corresponding to a plurality of point cloud pairs corresponding to the source point cloud data and the target point cloud data;

[0034] The initial change matrix is ​​optimized using the similarity transformation matrix to determine a target change matrix for each of the point cloud pairs.

[0035] Furthermore, determining initial change matrices corresponding to a plurality of point cloud pairs corresponding to the source point cloud data and the target point cloud data based on the source point cloud data and the target point cloud data includes:

[0036] Performing bidirectional feature matching on the source point cloud data and the target point cloud data to determine a plurality of point cloud pairs corresponding to the source point cloud data and the target point cloud data;

[0037] Determine the centroid coordinates of each point cloud pair using a preset centralization formula, and determine the covariance matrix of each point cloud pair based on the centroid coordinates and each point cloud pair;

[0038] Based on the covariance matrix, a target rotation matrix and a translation vector of each point cloud pair are determined respectively to determine an initial change matrix corresponding to each point cloud pair.

[0039] The present application also provides a device for splicing point cloud data of an autonomous driving vehicle, the device comprising:

[0040] A point cloud acquisition module is used to acquire source point cloud data collected by a laser radar provided in the vehicle for the target scene in the simulation test when performing a simulation test on the vehicle in the autonomous driving mode;

[0041] a feature extraction module, configured to determine, based on the source point cloud data and the target point cloud data, a similarity transformation matrix corresponding to the source point cloud data and the target point cloud data using a pre-trained feature model;

[0042] a point cloud registration module, configured to register a plurality of corresponding point cloud pairs between the source point cloud data and the target point cloud data based on the similarity transformation matrix, and determine a target change matrix for each of the point cloud pairs;

[0043] The point cloud stitching module is used to stitch and fuse the source point cloud data with the target point cloud data based on the target change matrix to obtain a three-dimensional model corresponding to the target scene.

[0044] Furthermore, when the point cloud acquisition module is used to obtain source point cloud data collected by a laser radar disposed in a vehicle for the target scene based on preset target point cloud data corresponding to the target scene in the simulation test, the point cloud acquisition module is used to:

[0045] Obtaining original source point cloud data collected by a laser radar installed in the vehicle for a target scene in a simulation test;

[0046] Performing denoising and filtering processing on the original source point cloud data to obtain source point cloud data to be processed;

[0047] Based on preset target point cloud data corresponding to the target scene, the source point cloud data to be processed is normalized to obtain source point cloud data corresponding to the target scene.

[0048] Furthermore, when the feature extraction module is used to determine the similarity transformation matrix corresponding to the source point cloud data and the target point cloud data using a pre-trained feature model based on the source point cloud data and the target point cloud data, the feature extraction module is used to:

[0049] Inputting the source point cloud data and the target point cloud data into a pre-trained local feature extraction model respectively, and obtaining a plurality of local features corresponding to the target scene output by the local feature extraction model;

[0050] Inputting the source point cloud data and the target point cloud data into a pre-trained global feature extraction model respectively, and obtaining a plurality of global features corresponding to the target scene output by the global feature extraction model;

[0051] Based on the local features and the global features, a similarity transformation matrix corresponding to the source point cloud data and the target point cloud data is calculated and determined.

[0052] Furthermore, when the feature extraction module is used to pre-train the local feature extraction model, the feature extraction module is used to:

[0053] Selecting training data from a preset training data set according to a preset data volume, performing edge computing labeling on the training data corresponding to the preset data volume to obtain a labeled training data set, and determining the unselected training data in the training data set as an unlabeled training data set;

[0054] Using the labeled training data set to train the first local feature extraction model to be trained to obtain a second local feature extraction model to be trained;

[0055] Inputting each unlabeled training data in the unlabeled training data set into the second local feature extraction model to be trained, determining the prediction confidence of each unlabeled training data, and numerically comparing the prediction confidence with a preset confidence threshold to obtain a comparison result;

[0056] The unlabeled training data whose comparison result shows that the prediction confidence is greater than the preset confidence threshold is added to the labeled training data set to obtain a target training data set;

[0057] Using the target training data set to train the second local feature extraction model to be trained to obtain a third local feature extraction model to be trained;

[0058] Performing point cloud reconstruction verification on the third to-be-trained local feature extraction model using a preset verification data set to obtain a point cloud reconstruction result output by the third to-be-trained local feature extraction model;

[0059] Based on the point cloud reconstruction result and the verification data set, parameters of the third local feature extraction model to be trained are updated to obtain the local feature extraction model.

[0060] Furthermore, when the feature extraction module is used to calculate and determine the similarity transformation matrix corresponding to the source point cloud data and the target point cloud data based on the local features and the global features, the feature extraction module is used to:

[0061] Determining a local difference item between the source point cloud data and the target point cloud data under the local feature, and determining a global difference item between the source point cloud data and the target point cloud data under the global feature;

[0062] For each of the local features, determining first source point cloud data corresponding to the source point cloud data under the local feature, and determining first target point cloud data corresponding to the target point cloud data under the local feature;

[0063] For each of the global features, determining second source point cloud data corresponding to the source point cloud data under the global feature, and determining second target point cloud data corresponding to the target point cloud data under the global feature;

[0064] Determining a transformation matrix of the source point cloud data relative to the target point cloud data using a preset transformation function based on the local difference term, the global difference term, the first source point cloud data, the second source point cloud data, the first target point cloud data, and the second target point cloud data;

[0065] Based on the transformation matrix, the source point cloud data and the target point cloud data, a similarity transformation matrix corresponding to the source point cloud data and the target point cloud data is determined using a preset mutual information transformation formula.

[0066] Furthermore, when the point cloud registration module is used to register a plurality of corresponding point cloud pairs between the source point cloud data and the target point cloud data based on the similarity transformation matrix and determine a target change matrix for each point cloud pair, the point cloud registration module is used to:

[0067] Based on the source point cloud data and the target point cloud data, determining initial change matrices corresponding to a plurality of point cloud pairs corresponding to the source point cloud data and the target point cloud data;

[0068] The initial change matrix is ​​optimized using the similarity transformation matrix to determine a target change matrix for each of the point cloud pairs.

[0069] Furthermore, when the point cloud registration module is used to determine, based on the source point cloud data and the target point cloud data, initial change matrices corresponding to a plurality of point cloud pairs corresponding to the source point cloud data and the target point cloud data, the point cloud registration module is used to:

[0070] Performing bidirectional feature matching on the source point cloud data and the target point cloud data to determine a plurality of point cloud pairs corresponding to the source point cloud data and the target point cloud data;

[0071] Determine the centroid coordinates of each point cloud pair using a preset centralization formula, and determine the covariance matrix of each point cloud pair based on the centroid coordinates and each point cloud pair;

[0072] Based on the covariance matrix, a target rotation matrix and a translation vector of each point cloud pair are determined respectively to determine an initial change matrix corresponding to each point cloud pair.

[0073] An embodiment of the present application also provides an electronic device, including: a processor, a memory, and a bus, wherein the memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor and the memory communicate through the bus, and when the machine-readable instructions are executed by the processor, the steps of the above-mentioned method for splicing point cloud data of autonomous driving vehicles are performed.

[0074] An embodiment of the present application also provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps of the method for splicing point cloud data of an autonomous driving vehicle as described above are executed.

[0075] The embodiments of the present application provide a method and device for splicing point cloud data of an autonomous driving vehicle. The splicing method includes: when performing a simulation test on a vehicle in an autonomous driving mode, based on preset target point cloud data corresponding to a target scene in the simulation test, obtaining source point cloud data collected by a laser radar arranged in the vehicle for the target scene; based on the source point cloud data and the target point cloud data, using a pre-trained feature model, determining the similarity transformation matrix corresponding to the source point cloud data and the target point cloud data; based on the similarity transformation matrix, aligning multiple point cloud pairs corresponding to the source point cloud data and the target point cloud data to determine a target change matrix for each point cloud pair; based on the target change matrix, splicing and fusing the source point cloud data with the target point cloud data to obtain a three-dimensional model corresponding to the target scene.

[0076] Compared with the method of directly collecting source point cloud data and performing point cloud splicing and registration through coarse registration and fine registration in the prior art, when conducting simulation tests on vehicles in autonomous driving mode, the target point cloud data corresponding to the target scene is used as a benchmark to obtain the source point cloud data collected by the lidar for the target scene, and the similarity transformation matrix corresponding to the source point cloud data and the target point cloud data is determined using a pre-trained feature model. The source point cloud data and the target point cloud data are then registered to determine the target change matrix, and then based on the target change matrix, the source point cloud data and the target point cloud data are fused to obtain a three-dimensional model corresponding to the target scene, thereby improving the accuracy and efficiency of splicing point cloud data of autonomous driving vehicles, and thereby improving the stability of point cloud data splicing.

[0077] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without creative work.

[0079] Figure 1A flowchart of a method for splicing point cloud data of an autonomous driving vehicle provided in an embodiment of the present application;

[0080] Figure 2 A schematic diagram of the structure of a device for splicing point cloud data of an autonomous driving vehicle provided in an embodiment of the present application;

[0081] Figure 3 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0082] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. The components of the embodiments of the present application generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the application for which protection is claimed, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, each other embodiment obtained by those skilled in the art without making creative work falls within the scope of protection of the present application.

[0083] Research has found that point cloud data, as a collection of points in three-dimensional space, can intuitively represent the shape and structure of an object. It is an important data form for three-dimensional reconstruction of vehicles during autonomous driving. Point cloud stitching technology is the core step in integrating the collected point cloud data into a three-dimensional model. The accuracy and stability of point cloud stitching directly affect the effect and application of three-dimensional reconstruction.

[0084] Among them, vehicle autonomous driving technology faces many challenges in its development process, and three-dimensional reconstruction technology has become a key supporting technology in this field of autonomous driving. Three-dimensional reconstruction provides an important basis for machine perception, navigation and decision-making by collecting and processing three-dimensional information of the environment or objects.

[0085] Currently, when stitching point cloud data collected by autonomous vehicles, the collected source point cloud data contains certain noise and outliers, and the source point cloud data and the point cloud data constructed by the vehicle do not have a unified coordinate system standard, which reduces the stability of the point cloud data stitching; in addition, the traditional point cloud stitching and registration method usually includes coarse registration and fine registration, but this method makes the stitched point cloud data inaccurate and inefficient.

[0086] During the actual operation of autonomous vehicles, point cloud data is usually collected by different devices (for example, laser scanners and binocular optical trackers, etc.). These devices correspond to different coordinate systems. If the point cloud data in the coordinate systems of different devices are directly aligned, large errors will be generated during splicing, affecting the splicing accuracy of the point cloud data. In addition, certain noise and outliers will be introduced during the collection process of point cloud data, which will interfere with the point cloud registration process and reduce the accuracy and efficiency of point cloud data splicing.

[0087] In addition, coarse registration and fine registration cannot guarantee the accuracy and efficiency of point cloud data docking at the same time. Coarse registration may not be able to accurately align the point cloud, and the fine registration process is time-consuming, making it difficult to meet the real-time and stability requirements of point cloud data docking.

[0088] Based on this, an embodiment of the present application provides a method for splicing point cloud data of an autonomous driving vehicle. When conducting a simulation test on a vehicle in an autonomous driving mode, the target point cloud data corresponding to the target scene is used as a reference to obtain the source point cloud data collected by the laser radar for the target scene, and the similarity transformation matrix corresponding to the source point cloud data and the target point cloud data is determined using a pre-trained feature model. The source point cloud data and the target point cloud data are aligned to determine the target change matrix, and then based on the target change matrix, the source point cloud data and the target point cloud data are fused to obtain a three-dimensional model corresponding to the target scene, thereby improving the accuracy and efficiency of splicing point cloud data of autonomous driving vehicles, and thereby improving the stability of point cloud data splicing.

[0089] See also Figure 1 , Figure 1 This is a flow chart of a method for splicing point cloud data of an autonomous driving vehicle provided in an embodiment of the present application. Figure 1 As shown in , the method for splicing point cloud data of an autonomous driving vehicle provided in an embodiment of the present application includes:

[0090] S101. When performing a simulation test on a vehicle in an autonomous driving mode, based on preset target point cloud data corresponding to a target scene in the simulation test, source point cloud data collected by a laser radar installed in the vehicle for the target scene is obtained.

[0091] It should be noted that the vehicle's autonomous driving mode refers to a system in which the vehicle can drive itself partially or completely without the intervention of a human driver; LiDAR (Light Detection and Ranging) plays a vital role in the vehicle's autonomous driving technology, creating a high-resolution 3D scene of the surrounding environment by emitting laser beams and measuring the time it takes for these beams to reflect back from objects.

[0092] In an embodiment of the present application, the target scene is a scene that needs to be virtually reconstructed in the vehicle's environment when performing a simulation test on a vehicle in autonomous driving mode. The scene may include but is not limited to the environment, objects, and elements.

[0093] Here, for the target scene, the vehicle will pre-set the corresponding target point cloud data for the target scene, and the laser radar installed in the vehicle will collect the source point cloud data of the target scene, and then splice the source point cloud data with the target point cloud data to obtain a three-dimensional model of the target scene.

[0094] Among them, the target point cloud data includes multiple point cloud coordinates corresponding to the target scene on the vehicle side; the source point cloud data includes multiple point cloud coordinates collected by the laser radar for the target scene.

[0095] In one embodiment of the present application, in a specific implementation, the step of obtaining source point cloud data collected by a laser radar disposed in a vehicle for the target scene based on preset target point cloud data corresponding to the target scene in the simulation test in step S101 may include:

[0096] S1011. Obtain original source point cloud data collected by a laser radar installed in a vehicle for a target scene in a simulation test.

[0097] In this step, during specific implementation, first, the laser radar removes dynamic objects in the target scene in the simulation test; then, after removing the dynamic objects, the laser radar collects the original source point cloud data of the target scene; finally, the original source point cloud data is obtained.

[0098] The original source point cloud data is a plurality of point cloud coordinates of the target scene collected by the lidar without data processing.

[0099] S1012: Perform denoising and filtering processing on the original source point cloud data to obtain source point cloud data to be processed.

[0100] In this step, during specific implementation, first, filtering methods including statistical filtering and radius filtering are used to remove noise points in the original source point cloud data; then, the filtered original source point cloud data is smoothed to reduce local fluctuations of the original source point cloud data; finally, the source point cloud data to be processed is obtained.

[0101] The source point cloud data to be processed are multiple point cloud coordinates of the target scene collected by the lidar and processed by denoising, filtering, smoothing and other data processing.

[0102] S1013 : Based on preset target point cloud data corresponding to the target scene, normalize the source point cloud data to be processed to obtain source point cloud data corresponding to the target scene.

[0103] In this step, during specific implementation, first, the coordinate system and point cloud scale of the preset target point cloud data corresponding to the target scene are determined; then, the source point cloud data to be processed is normalized according to the coordinate system and point cloud scale of the target point cloud data to eliminate the influence of scale differences; finally, the source point cloud data corresponding to the target scene after the normalization of the source point cloud data to be processed is obtained.

[0104] S102 : Based on the source point cloud data and the target point cloud data, using a pre-trained feature model, determine a similarity transformation matrix corresponding to the source point cloud data and the target point cloud data.

[0105] In an embodiment of the present application, for the splicing of source point cloud data and target point cloud data, a pre-trained feature model is used to extract local features and global features between the source point cloud data and the target point cloud data, and based on the extracted features, the corresponding similarity transformation matrix between the source point cloud data and the target point cloud data is determined to determine the similarity between the source point cloud data and the target point cloud data.

[0106] In one embodiment of the present application, during specific implementation, step S102 may include:

[0107] S1021. Input the source point cloud data and the target point cloud data into a pre-trained local feature extraction model respectively to obtain a plurality of local features corresponding to the target scene output by the local feature extraction model.

[0108] In an embodiment of the present application, the extracted local features mainly capture geometric information by analyzing the local neighborhood of each point cloud coordinate in the source point cloud data and the target point cloud data through a local feature extraction model.

[0109] Here, the local features include but are not limited to geometric features, texture features, context features, structural features and time features.

[0110] Among them, geometric features include at least point coordinates, vectors, curvature and main curvature direction; texture features include at least color RGB information, transparency information, reflection intensity and gradient; context features include at least semantic category, instance category and environmental information; structural features include at least neighborhood points and depth information; time features include at least timestamp and trajectory information.

[0111] In one embodiment of the present application, in a specific implementation, the step of pre-training the local feature extraction model in step S1021 may include:

[0112] S1021A. Select training data from a preset training data set according to a preset data volume, perform edge computing labeling on the training data corresponding to the preset data volume to obtain a labeled training data set, and determine the unselected training data in the training data set as an unlabeled training data set.

[0113] In an embodiment of the present application, in order to enable the local feature extraction model to accurately capture the geometric details in the point cloud data, the local feature extraction model is trained to focus on edge computing of the point cloud data.

[0114] In this step, training data corresponding to a preset data volume is selected from a preset training data set, the rate of change of each point cloud coordinate line in the training data is calculated, the point cloud coordinates whose rate of change is greater than a preset change rate threshold are recorded as edge points, and the point cloud coordinates are labeled to obtain a labeled training data set.

[0115] Furthermore, after obtaining the labeled training data set in the training data set, the unselected training data in the training data set is determined as the unlabeled training data set.

[0116] Among them, the preset training data set may include point cloud data corresponding to the original trajectory data of the autonomous driving vehicle; the data volume of the unlabeled training data set is extremely smaller than that of the labeled training data set.

[0117] In this way, the parameters of the local feature extraction model are adjusted according to the labeled training data set and the unlabeled training data set to improve the accuracy of local feature extraction.

[0118] S1021B: Use the labeled training data set to train the first local feature extraction model to be trained to obtain a second local feature extraction model to be trained.

[0119] In this step, a small amount of labeled training data set is used to guide the first local feature extraction model to be trained to learn the local features in the labeled training data set, thereby obtaining the second local feature extraction model to be trained.

[0120] S1021C. Input each unlabeled training data in the unlabeled training data set into the second local feature extraction model to be trained, determine the prediction confidence of each unlabeled training data, and numerically compare the prediction confidence with a preset confidence threshold to obtain a comparison result.

[0121] In this step, each unlabeled training data in the unlabeled training data set is reconstructed and optimized using each unlabeled training data in a large number of unlabeled training data sets to obtain a prediction result output by the second local feature extraction model to be trained, and based on the prediction result, the prediction confidence of each of the unlabeled training data is determined respectively.

[0122] Furthermore, the prediction confidence of each of the unlabeled training data is numerically compared with a preset confidence threshold to obtain a comparison result.

[0123] Here, since labeling point cloud data is time-consuming and computationally expensive, in an embodiment of the present application, a small amount of labeled training data sets are used to guide the first local feature extraction model to be trained to learn local features in the labeled training data sets, and each unlabeled training data in a large number of unlabeled training data sets is used to reconstruct and optimize each unlabeled training data in the unlabeled training data sets.

[0124] S1021D. Put the unlabeled training data whose comparison result shows that the prediction confidence is greater than a preset confidence threshold into the labeled training data set to obtain a target training data set.

[0125] In this step, the unlabeled training data with prediction confidence greater than the preset confidence threshold are added to the labeled training dataset to obtain the target training dataset ultimately used to train the model.

[0126] S1021E: Use the target training data set to train the second local feature extraction model to be trained to obtain a third local feature extraction model to be trained.

[0127] In this step, the target training data set is used to iteratively train the second local feature extraction model to be trained, and the parameters of the second local feature model to be trained are continuously updated so that the second local feature model to be trained gradually adapts to more local features in the target training data.

[0128] S1021F. Use a preset verification data set to perform point cloud reconstruction verification on the third local feature extraction model to be trained, and obtain a point cloud reconstruction result output by the third local feature extraction model to be trained.

[0129] In this step, during the specific implementation, first, the third local feature extraction model to be trained is used to extract local features in a preset verification data set; then, based on the local features, the point cloud is reconstructed; finally, the loss degree and mean square error of the reconstructed point cloud and the verification data set are verified to obtain the point cloud reconstruction result output by the third local feature extraction model to be trained.

[0130] In an embodiment of the present application, the point cloud reconstruction result includes at least point cloud data reconstructed based on local features, and the loss degree and mean square error in the verification data set.

[0131] Here, the expression of the loss function used to calculate the loss degree based on the reconstructed point cloud and the validation dataset is as follows.

[0132]

[0133] Among them, LCD represents the loss degree; P i represents the validation dataset; P j Represents the point cloud coordinate set of the reconstructed point cloud; n represents the number of point cloud coordinates in the point cloud coordinate set of the reconstructed point cloud; m represents the number of point cloud coordinates in the validation dataset.

[0134] S1021G. Based on the point cloud reconstruction result and the verification data set, update the parameters of the third local feature extraction model to be trained to obtain the local feature extraction model.

[0135] In this step, during the specific implementation, first, the loss degree and mean square error in the point cloud reconstruction result are regularized and weighted and summed to obtain the total loss value; then, based on the total loss value, the model parameters in the third local feature extraction model to be trained are updated using the gradient descent method; finally, after iterative training for a preset number of iterations, the local feature extraction model is obtained.

[0136] S1022. Input the source point cloud data and the target point cloud data into a pre-trained global feature extraction model respectively to obtain a plurality of global features corresponding to the target scene output by the global feature extraction model.

[0137] In an embodiment of the present application, the global features reflect the overall shape and structural information of the point cloud data; the global features include but are not limited to geometric features, contextual features, structural features and statistical features.

[0138] Among them, geometric features include at least bounding boxes, spatial distribution statistics and principal component analysis; contextual features include at least semantic categories, instance categories and environmental information; structural features include at least scene layout, key point detection and graph structure; statistical features include at least point cloud density, point cloud size and point cloud volume.

[0139] In this step, the source point cloud data and the target point cloud data are respectively input into the pre-trained global feature extraction model, and global feature classification is performed by integrating the global feature extraction model of multiple decision trees to process high-dimensional features and reduce overfitting, thereby obtaining multiple global features corresponding to the target scene output by the global feature extraction model.

[0140] The construction process of the decision tree of the pre-trained global feature extraction model includes training steps such as selecting the best splitting attribute, determining the splitting point, and recursively constructing subtrees.

[0141] Here, the pre-trained global feature extraction model may include at least one of a convolutional neural network model (CNN) or a graph neural network model (GNN).

[0142] S1023. Based on the local features and the global features, calculate and determine a similarity transformation matrix corresponding to the source point cloud data and the target point cloud data.

[0143] In an embodiment of the present application, the similarity between the source point cloud data and the target point cloud data is calculated based on local features and global features.

[0144] In one embodiment of the present application, during specific implementation, step S1023 may include:

[0145] S10231. Determine a local difference item between the source point cloud data and the target point cloud data under the local feature, and determine a global difference item between the source point cloud data and the target point cloud data under the global feature.

[0146] In an embodiment of the present application, the local difference item may include a supplementary descriptor matching item, such as cosine similarity, etc.; the global difference item may include a geometric constraint item, such as normal angle and curvature difference, etc.

[0147] S10232. For each of the local features, determine the first source point cloud data corresponding to the source point cloud data under the local feature, and determine the first target point cloud data corresponding to the target point cloud data under the local feature.

[0148] S10233. For each of the global features, determine the second source point cloud data corresponding to the source point cloud data under the global feature, and determine the second target point cloud data corresponding to the target point cloud data under the global feature.

[0149] In an embodiment of the present application, a preset feature matching algorithm is used to determine the first source point cloud data and the first target point cloud data under each local feature, and to determine the second source point cloud data and the second target point cloud data under each global feature.

[0150] S10234. Based on the local difference item, the global difference item, the first source point cloud data, the second source point cloud data, the first target point cloud data and the second target point cloud data, a preset transformation function is used to determine the transformation matrix of the source point cloud data relative to the target point cloud data.

[0151] In an embodiment of the present application, by calculating the transformation matrix of the source point cloud data relative to the target point cloud data, the best corresponding target point cloud coordinates are found in the target point cloud for each source point cloud coordinate in the source point cloud data to minimize the local loss of multiple constraints.

[0152] Here, the expression of the preset transformation function is as follows.

[0153]

[0154] Where T(i) represents the transformation matrix of each source point cloud coordinate in the source point cloud data relative to the corresponding target point cloud coordinate in the target point cloud data; Desc(i,j) represents the local difference term; Geo(i,j) represents the global difference term; f i represents the i-th local feature point in the source point cloud data; v i represents the i-th global feature point in the source point cloud data; g j represents the jth local feature point in the target point cloud data; w j represents the jth global feature point in the target point cloud data; λ1 represents the weight coefficient between the local feature difference and the global feature difference; λ2 represents the weight coefficient of the local feature difference; λ3 represents the weight coefficient of the global feature difference.

[0155] S10235. Based on the transformation matrix, the source point cloud data and the target point cloud data, determine the similarity transformation matrix corresponding to the source point cloud data and the target point cloud data using a preset mutual information transformation formula.

[0156] In an embodiment of the present application, based on the concept of mutual information in information theory, a similarity transformation matrix that maximizes the mutual information between the source point cloud data and the target point cloud data is determined. The expression of the preset mutual information transformation formula is shown below.

[0157]

[0158] Among them, Transform represents the corresponding similarity transformation matrix between the source point cloud data and the target point cloud data; MI represents the mutual information between the source point cloud data and the target point cloud data; P represents the source point cloud data; Q represents the target point cloud data; T represents the transformation matrix.

[0159] S103 . Based on the similarity transformation matrix, register a plurality of point cloud pairs corresponding to the source point cloud data and the target point cloud data, and determine a target change matrix for each point cloud pair.

[0160] In one embodiment of the present application, during specific implementation, step S103 may include:

[0161] S1031. Based on the source point cloud data and the target point cloud data, determine initial change matrices corresponding to a plurality of point cloud pairs corresponding to the source point cloud data and the target point cloud data.

[0162] In an embodiment of the present application, in order to quickly align the source point cloud data and the target point cloud data at a rough position, the initial change matrix is ​​calculated and determined using the least squares method based on multiple corresponding point cloud pairs between the source point cloud data and the target point cloud data.

[0163] Here, the initial change matrix is ​​used to roughly align the source point cloud data with the target point cloud data.

[0164] In one embodiment of the present application, during specific implementation, step S1031 may include:

[0165] S10311. Perform bidirectional feature matching on the source point cloud data and the target point cloud data to determine a plurality of point cloud pairs corresponding to the source point cloud data and the target point cloud data.

[0166] S10312. Determine the centroid coordinates of each point cloud pair using a preset centralization formula, and determine the covariance matrix of each point cloud pair based on the centroid coordinates and each point cloud pair.

[0167] S1033. Based on the covariance matrix, determine the target rotation matrix and translation vector of each point cloud pair respectively to determine the initial change matrix corresponding to each point cloud pair.

[0168] Furthermore, after determining the initial change matrix corresponding to each point cloud pair, the initial change matrix is ​​optimized using a preset optimization method to improve the accuracy of the registration.

[0169] S1032: Optimize the initial change matrix using the similarity transformation matrix to determine a target change matrix for each point cloud pair.

[0170] In this step, the initial change matrix is ​​optimized using the similarity transformation matrix to fine-tune the registration error in the initial change matrix and determine the target change matrix for each point cloud pair.

[0171] S104 : Based on the target change matrix, the source point cloud data and the target point cloud data are spliced ​​and fused to obtain a three-dimensional model corresponding to the target scene.

[0172] In this step, during specific implementation, first, based on the target change matrix, the source point cloud data and the target point cloud data are spliced ​​and fused using splicing and fusion methods such as weighted averaging and Poisson surface reconstruction to obtain spliced ​​and fused point cloud data; then, the spliced ​​and fused point cloud data is smoothed to optimize the surface of the three-dimensional model corresponding to the target scene and improve the visual effect of the three-dimensional model; finally, based on the optimized spliced ​​and fused point cloud data, the three-dimensional model corresponding to the target scene is obtained.

[0173] The method for splicing point cloud data of an autonomous driving vehicle provided in an embodiment of the present application obtains source point cloud data collected by a lidar for the target scene based on target point cloud data corresponding to the target scene when conducting a simulation test on a vehicle in an autonomous driving mode, uses a pre-trained feature model to determine the similarity transformation matrix corresponding to the source point cloud data and the target point cloud data, aligns the source point cloud data and the target point cloud data, determines the target change matrix, and then, based on the target change matrix, fuses the source point cloud data and the target point cloud data to obtain a three-dimensional model corresponding to the target scene, thereby improving the accuracy and efficiency of splicing point cloud data of autonomous driving vehicles, and thereby improving the stability of point cloud data splicing.

[0174] See also Figure 2 , Figure 2 This is a schematic diagram of the structure of a device for splicing point cloud data of an autonomous driving vehicle provided in an embodiment of the present application. Figure 2 As shown in , the splicing device 200 includes:

[0175] The point cloud acquisition module 210 is configured to acquire source point cloud data collected by a laser radar disposed in the vehicle for the target scene in the simulation test, based on preset target point cloud data corresponding to the target scene in the simulation test, when performing a simulation test on the vehicle in the autonomous driving mode;

[0176] A feature extraction module 220 is configured to determine, based on the source point cloud data and the target point cloud data, a similarity transformation matrix corresponding to the source point cloud data and the target point cloud data using a pre-trained feature model;

[0177] a point cloud registration module 230 for registering a plurality of corresponding point cloud pairs between the source point cloud data and the target point cloud data based on the similarity transformation matrix, and determining a target change matrix for each of the point cloud pairs;

[0178] The point cloud stitching module 240 is configured to stitch and fuse the source point cloud data with the target point cloud data based on the target change matrix to obtain a three-dimensional model corresponding to the target scene.

[0179] Furthermore, when the point cloud acquisition module 210 is used to obtain source point cloud data collected by a laser radar disposed in a vehicle for the target scene based on preset target point cloud data corresponding to the target scene in the simulation test, the point cloud acquisition module 210 is used to:

[0180] Obtaining original source point cloud data collected by a laser radar installed in the vehicle for a target scene in a simulation test;

[0181] Performing denoising and filtering processing on the original source point cloud data to obtain source point cloud data to be processed;

[0182] Based on preset target point cloud data corresponding to the target scene, the source point cloud data to be processed is normalized to obtain source point cloud data corresponding to the target scene.

[0183] Furthermore, when the feature extraction module 220 is used to determine the similarity transformation matrix corresponding to the source point cloud data and the target point cloud data using a pre-trained feature model based on the source point cloud data and the target point cloud data, the feature extraction module 220 is used to:

[0184] Inputting the source point cloud data and the target point cloud data into a pre-trained local feature extraction model respectively, and obtaining a plurality of local features corresponding to the target scene output by the local feature extraction model;

[0185] Inputting the source point cloud data and the target point cloud data into a pre-trained global feature extraction model respectively, and obtaining a plurality of global features corresponding to the target scene output by the global feature extraction model;

[0186] Based on the local features and the global features, a similarity transformation matrix corresponding to the source point cloud data and the target point cloud data is calculated and determined.

[0187] Furthermore, when the feature extraction module 220 is used to pre-train the local feature extraction model, the feature extraction module 220 is used to:

[0188] Selecting training data from a preset training data set according to a preset data volume, performing edge computing labeling on the training data corresponding to the preset data volume to obtain a labeled training data set, and determining the unselected training data in the training data set as an unlabeled training data set;

[0189] Using the labeled training data set to train the first local feature extraction model to be trained to obtain a second local feature extraction model to be trained;

[0190] Inputting each unlabeled training data in the unlabeled training data set into the second local feature extraction model to be trained, determining the prediction confidence of each unlabeled training data, and numerically comparing the prediction confidence with a preset confidence threshold to obtain a comparison result;

[0191] The unlabeled training data whose comparison result shows that the prediction confidence is greater than the preset confidence threshold is added to the labeled training data set to obtain a target training data set;

[0192] Using the target training data set to train the second local feature extraction model to be trained to obtain a third local feature extraction model to be trained;

[0193] Performing point cloud reconstruction verification on the third to-be-trained local feature extraction model using a preset verification data set to obtain a point cloud reconstruction result output by the third to-be-trained local feature extraction model;

[0194] Based on the point cloud reconstruction result and the verification data set, parameters of the third local feature extraction model to be trained are updated to obtain the local feature extraction model.

[0195] Furthermore, when the feature extraction module 220 is used to calculate and determine the similarity transformation matrix corresponding to the source point cloud data and the target point cloud data based on the local features and the global features, the feature extraction module 220 is used to:

[0196] Determining a local difference item between the source point cloud data and the target point cloud data under the local feature, and determining a global difference item between the source point cloud data and the target point cloud data under the global feature;

[0197] For each of the local features, determining first source point cloud data corresponding to the source point cloud data under the local feature, and determining first target point cloud data corresponding to the target point cloud data under the local feature;

[0198] For each of the global features, determining second source point cloud data corresponding to the source point cloud data under the global feature, and determining second target point cloud data corresponding to the target point cloud data under the global feature;

[0199] Determining a transformation matrix of the source point cloud data relative to the target point cloud data using a preset transformation function based on the local difference term, the global difference term, the first source point cloud data, the second source point cloud data, the first target point cloud data, and the second target point cloud data;

[0200] Based on the transformation matrix, the source point cloud data and the target point cloud data, a similarity transformation matrix corresponding to the source point cloud data and the target point cloud data is determined using a preset mutual information transformation formula.

[0201] Furthermore, when the point cloud registration module 230 is used to register a plurality of corresponding point cloud pairs between the source point cloud data and the target point cloud data based on the similarity transformation matrix and determine a target change matrix for each point cloud pair, the point cloud registration module 230 is used to:

[0202] Based on the source point cloud data and the target point cloud data, determining initial change matrices corresponding to a plurality of point cloud pairs corresponding to the source point cloud data and the target point cloud data;

[0203] The initial change matrix is ​​optimized using the similarity transformation matrix to determine a target change matrix for each of the point cloud pairs.

[0204] Furthermore, when the point cloud registration module 230 is used to determine, based on the source point cloud data and the target point cloud data, initial change matrices corresponding to a plurality of point cloud pairs corresponding to the source point cloud data and the target point cloud data, the point cloud registration module 230 is used to:

[0205] Performing bidirectional feature matching on the source point cloud data and the target point cloud data to determine a plurality of point cloud pairs corresponding to the source point cloud data and the target point cloud data;

[0206] Determine the centroid coordinates of each point cloud pair using a preset centralization formula, and determine the covariance matrix of each point cloud pair based on the centroid coordinates and each point cloud pair;

[0207] Based on the covariance matrix, a target rotation matrix and a translation vector of each point cloud pair are determined respectively to determine an initial change matrix corresponding to each point cloud pair.

[0208] The device for splicing point cloud data of an autonomous driving vehicle provided in an embodiment of the present application obtains source point cloud data collected by a laser radar for a target scene based on target point cloud data corresponding to the target scene when performing a simulation test on a vehicle in an autonomous driving mode, uses a pre-trained feature model to determine the similarity transformation matrix corresponding to the source point cloud data and the target point cloud data, aligns the source point cloud data and the target point cloud data, determines the target change matrix, and then fuses the source point cloud data with the target point cloud data based on the target change matrix to obtain a three-dimensional model corresponding to the target scene, thereby improving the accuracy and efficiency of splicing point cloud data of autonomous driving vehicles, and thereby improving the stability of point cloud data splicing.

[0209] See also Figure 3 , Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. Figure 3 As shown in FIG, the electronic device 300 includes a processor 310 , a memory 320 and a bus 330 .

[0210] The memory 320 stores machine-readable instructions executable by the processor 310. When the electronic device 300 is running, the processor 310 communicates with the memory 320 via the bus 330. When the machine-readable instructions are executed by the processor 310, the above-mentioned Figure 1 The steps of the method for splicing point cloud data of an autonomous driving vehicle in the method embodiment shown are specifically implemented in accordance with the method embodiment and will not be described in detail here.

[0211] The embodiment of the present application also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the computer program can execute the above-mentioned Figure 1 The steps of the method for splicing point cloud data of an autonomous driving vehicle in the method embodiment shown are specifically implemented in accordance with the method embodiment and will not be described in detail here.

[0212] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0213] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. There may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed may be through some communication interface, indirect coupling or communication connection of devices or units, which may be electrical, mechanical or other forms.

[0214] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0215] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0216] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a non-volatile computer-readable storage medium that is executable by a processor. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling 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 various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0217] Finally, it should be noted that the above-described embodiments are only specific implementation methods of the present application, which are used to illustrate the technical solutions of the present application, rather than to limit them. The scope of protection of the present application is not limited thereto. Although the present application has been described in detail with reference to the above-mentioned embodiments, those skilled in the art should understand that any person skilled in the art can modify or easily conceive of changes to the technical solutions described in the above-mentioned embodiments within the technical scope disclosed in the present application, or perform equivalent replacements for some of the technical features thereof. These modifications, changes, or replacements do not deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A method for splicing point cloud data of an autonomous driving vehicle, characterized in that: The splicing method comprises: When conducting a simulation test on a vehicle in an autonomous driving mode, obtaining source point cloud data of a target scene in the simulation test, collected by a laser radar disposed in the vehicle, based on preset target point cloud data corresponding to the target scene; Based on the source point cloud data and the target point cloud data, using a pre-trained feature model, determining a similarity transformation matrix corresponding to the source point cloud data and the target point cloud data; Based on the similarity transformation matrix, registering a plurality of corresponding point cloud pairs between the source point cloud data and the target point cloud data, and determining a target change matrix for each of the point cloud pairs; Based on the target change matrix, the source point cloud data and the target point cloud data are spliced ​​and fused to obtain a three-dimensional model corresponding to the target scene.

2. The method according to claim 1, characterized in that The method of obtaining the preset target point cloud data corresponding to the target scene in the simulation test and acquiring the source point cloud data collected by the laser radar disposed in the vehicle for the target scene includes: Obtaining original source point cloud data collected by a laser radar installed in the vehicle for a target scene in a simulation test; Performing denoising and filtering processing on the original source point cloud data to obtain source point cloud data to be processed; Based on preset target point cloud data corresponding to the target scene, the source point cloud data to be processed is normalized to obtain source point cloud data corresponding to the target scene.

3. The method according to claim 1, characterized in that The determining, based on the source point cloud data and the target point cloud data, of a similarity transformation matrix corresponding to the source point cloud data and the target point cloud data using a pre-trained feature model, includes: Inputting the source point cloud data and the target point cloud data into a pre-trained local feature extraction model respectively, and obtaining a plurality of local features corresponding to the target scene output by the local feature extraction model; Inputting the source point cloud data and the target point cloud data into a pre-trained global feature extraction model respectively, and obtaining a plurality of global features corresponding to the target scene output by the global feature extraction model; Based on the local features and the global features, a similarity transformation matrix corresponding to the source point cloud data and the target point cloud data is calculated and determined.

4. The method according to claim 3, characterized in that The local feature extraction model is pre-trained by the following steps: Selecting training data from a preset training data set according to a preset data volume, performing edge computing labeling on the training data corresponding to the preset data volume to obtain a labeled training data set, and determining the unselected training data in the training data set as an unlabeled training data set; Using the labeled training data set to train the first local feature extraction model to be trained to obtain a second local feature extraction model to be trained; Inputting each unlabeled training data in the unlabeled training data set into the second local feature extraction model to be trained, determining the prediction confidence of each unlabeled training data, and numerically comparing the prediction confidence with a preset confidence threshold to obtain a comparison result; The unlabeled training data whose comparison result shows that the prediction confidence is greater than the preset confidence threshold is added to the labeled training data set to obtain a target training data set; Using the target training data set to train the second local feature extraction model to be trained to obtain a third local feature extraction model to be trained; Performing point cloud reconstruction verification on the third to-be-trained local feature extraction model using a preset verification data set to obtain a point cloud reconstruction result output by the third to-be-trained local feature extraction model; Based on the point cloud reconstruction result and the verification data set, parameters of the third local feature extraction model to be trained are updated to obtain the local feature extraction model.

5. The method according to claim 3, characterized in that The calculating and determining, based on the local features and the global features, a similarity transformation matrix corresponding to the source point cloud data and the target point cloud data, includes: Determining a local difference item between the source point cloud data and the target point cloud data under the local feature, and determining a global difference item between the source point cloud data and the target point cloud data under the global feature; For each of the local features, determining first source point cloud data corresponding to the source point cloud data under the local feature, and determining first target point cloud data corresponding to the target point cloud data under the local feature; For each of the global features, determining second source point cloud data corresponding to the source point cloud data under the global feature, and determining second target point cloud data corresponding to the target point cloud data under the global feature; Determining a transformation matrix of the source point cloud data relative to the target point cloud data using a preset transformation function based on the local difference term, the global difference term, the first source point cloud data, the second source point cloud data, the first target point cloud data, and the second target point cloud data; Based on the transformation matrix, the source point cloud data and the target point cloud data, a similarity transformation matrix corresponding to the source point cloud data and the target point cloud data is determined using a preset mutual information transformation formula.

6. The method according to claim 1, characterized in that The registering a plurality of corresponding point cloud pairs between the source point cloud data and the target point cloud data based on the similarity transformation matrix, and determining a target change matrix for each point cloud pair, includes: Based on the source point cloud data and the target point cloud data, determining initial change matrices corresponding to a plurality of point cloud pairs corresponding to the source point cloud data and the target point cloud data; The initial change matrix is ​​optimized using the similarity transformation matrix to determine a target change matrix for each of the point cloud pairs.

7. The method according to claim 6, characterized in that The determining, based on the source point cloud data and the target point cloud data, initial change matrices corresponding to a plurality of point cloud pairs corresponding to the source point cloud data and the target point cloud data, includes: Performing bidirectional feature matching on the source point cloud data and the target point cloud data to determine a plurality of point cloud pairs corresponding to the source point cloud data and the target point cloud data; Determine the centroid coordinates of each point cloud pair using a preset centralization formula, and determine the covariance matrix of each point cloud pair based on the centroid coordinates and each point cloud pair; Based on the covariance matrix, a target rotation matrix and a translation vector of each point cloud pair are determined respectively to determine an initial change matrix corresponding to each point cloud pair.

8. A device for stitching point cloud data of an autonomous driving vehicle, characterized in that: The splicing device comprises: A point cloud acquisition module is used to acquire source point cloud data collected by a laser radar provided in the vehicle for the target scene in the simulation test when performing a simulation test on the vehicle in the autonomous driving mode; a feature extraction module, configured to determine, based on the source point cloud data and the target point cloud data, a similarity transformation matrix corresponding to the source point cloud data and the target point cloud data using a pre-trained feature model; a point cloud registration module, configured to register a plurality of corresponding point cloud pairs between the source point cloud data and the target point cloud data based on the similarity transformation matrix, and determine a target change matrix for each of the point cloud pairs; The point cloud stitching module is used to stitch and fuse the source point cloud data with the target point cloud data based on the target change matrix to obtain a three-dimensional model corresponding to the target scene.

9. An electronic device, characterized in that: include: A processor, a memory and a bus, wherein the memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor and the memory communicate through the bus. When the processor is running, the machine-readable instructions execute the steps of the method for splicing point cloud data of an autonomous driving vehicle as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, which, when executed by a processor, executes the steps of the method for stitching point cloud data of an autonomous driving vehicle according to any one of claims 1 to 7.

Images