Target detection model generalization enhancement method and device considering point cloud distribution structure

Through voxelization filtering and data unification of lidar point cloud center, the problem of poor performance of deep learning models under different data sets is solved, the generalization and real-time performance of the three-dimensional object detection model is improved, and data acquisition work is reduced. Promotes the application in complex traffic scenarios.

CN118967918APending Publication Date: 2024-11-15YANSHAN UNIV
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202410936585.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-12
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

Deep learning models perform poorly when facing the data distribution structure of point clouds of lidars in the dataset, resulting in insufficient generalization capabilities in actual scenarios, especially due to the differences in horizontal and vertical resolutions of different models of lidars. The problem of heterogeneity of point cloud data.

Method used

Through the voxelization filtering of the center of the lidar point cloud, the inconsistent horizontal and vertical resolution structure of point clouds is broken, the reflection intensity dimension is removed, and the coordinates of the cloud origin are unified to form a data set with consistent horizontal and vertical resolutions, which is used to train three-dimensional target detection Model, improve the generalization of the model.

Benefits of technology

On the basis of ensuring real-time performance, the generalization of the three-dimensional target detection model of lidar is improved, the human data collection and labeling work is reduced, the barrier between the autonomous driving simulation data set and the natural driving scene data set is reduced, and complex traffic is promoted Application of three-dimensional object detection technology in scenarios.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118967918A_ABST
    Figure CN118967918A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of three-dimensional target detection, in particular to a target detection model generalization enhancement method and device considering a point cloud distribution structure, and can solve the problem that an existing deep learning model is not suitable for a laser radar point cloud data distribution structure in a data set to a certain extent. And the actual scene generalization ability is insufficient. The method comprises the following steps: acquiring laser radar point cloud data in a public simulation automatic driving data set and a natural driving data set with a three-dimensional target detection truth value; obtaining a plurality of point cloud data sets through laser radar point cloud voxelization filtering; highly unifying the origin coordinates of the plurality of point cloud data sets, and removing the point cloud reflection intensity dimension; fusing the plurality of processed point cloud data sets into a training set, a test set and a verification set; and based on the training set, an existing three-dimensional target detection network is used for training, and a high-generalization three-dimensional target detection model is obtained.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the technical field of three-dimensional target detection, and more specifically, to a method and device for enhancing the generalization of a target detection model taking into account a point cloud distribution structure. Background Art

[0002] With the continuous development of modern artificial intelligence technology, autonomous driving has gradually emerged and become one of the effective solutions to problems such as traffic congestion and traffic accidents.

[0003] In order to meet the needs of autonomous driving technology in complex traffic scenarios, scholars at home and abroad have continuously innovated and proposed a series of advanced environmental perception algorithms and technologies to adapt to the data of different perception sensors. Among them, the three-dimensional target detection technology based on lidar plays an important role in the fields of autonomous driving and smart transportation because it can provide autonomous driving vehicles with information such as the category, location, and size of obstacles in traffic scenes. In recent years, deep learning models have been widely used by scholars at home and abroad because they can effectively capture potential features in the surrounding environment from lidar point cloud data.

[0004] However, deep learning models are often limited by the structure of the lidar point cloud in the dataset. This limitation stems from the differences in the horizontal and vertical resolutions of different models of lidars, which leads to significantly different distribution structures of the point cloud data. Therefore, in cases where the distribution structure of the lidar point cloud data is different from that in the dataset, the heterogeneity of the lidar point cloud makes the performance of deep learning models on test data often unsatisfactory. This difference may indirectly affect the generalization ability of the model in actual scenarios, further highlighting the challenges faced by the model when facing novel data. Summary of the invention

[0005] In order to solve the problem that the heterogeneity of lidar point clouds in existing deep learning models often makes the performance of deep learning models on test data unsatisfactory when the distribution structure of lidar point cloud data is different from that in the dataset, resulting in insufficient generalization ability in actual scenarios, the present application provides a method and device for enhancing the generalization of a target detection model taking into account the point cloud distribution structure.

[0006] The embodiment of the present application is implemented as follows:

[0007] In a first aspect, the present application provides a method for enhancing the generalization of an object detection model considering the point cloud distribution structure, comprising:

[0008] Obtain lidar point cloud data from public simulated autonomous driving datasets and natural driving datasets with true 3D object detection values;

[0009] Through voxel filtering at the center of the LiDAR point cloud, the inconsistent structure of the horizontal and vertical resolutions of the point clouds in different models of LiDAR data sets is broken, and multiple LiDAR point cloud data sets with consistent horizontal and vertical resolutions are obtained. The dimension of the reflection intensity of the LiDAR point cloud is removed, and the coordinates of the point cloud origins of multiple point cloud data sets are highly unified, so as to obtain multiple LiDAR point cloud data sets with a unified point cloud distribution structure.

[0010] The processed multiple point cloud data sets are fused and divided into a training set, a test set and a validation set;

[0011] Based on the training set, an existing three-dimensional object detection network is used for training to obtain a strong generalization three-dimensional object detection model.

[0012] In a possible implementation, the method further includes:

[0013] The three-dimensional object detection model is tested based on the test set and the validation set.

[0014] In a possible implementation, the step of obtaining the laser radar point cloud data in the public simulated driving dataset and the natural driving dataset with the true value of three-dimensional target detection further includes:

[0015] Obtain point cloud data of different types of lidars in the current public natural driving data set to obtain a natural driving point cloud data set;

[0016] The point cloud data of the laser radar in the currently public simulated autonomous driving data set is obtained to obtain the autonomous driving point cloud data set.

[0017] In a possible implementation, the voxel filtering of the laser radar point cloud center breaks the inconsistent structure of the horizontal and vertical resolutions of the point clouds in the laser radar data sets of different models to obtain multiple laser radar point cloud data sets with consistent horizontal and vertical resolutions, further comprising:

[0018] The point cloud voxel filtering method is used to break the data distribution structure unique to the lidar point cloud in the natural driving point cloud dataset and the autonomous driving point cloud dataset, and obtain multiple point cloud datasets.

[0019] In a possible implementation, removing the dimension of the laser radar point cloud reflection intensity and highly unifying the point cloud origin coordinates of multiple point cloud data sets to obtain multiple laser radar point cloud data sets with a unified point cloud distribution structure further includes:

[0020] Divide the point cloud into a 3D voxel grid of a certain size according to its range;

[0021] For each 3D voxel grid, if there is at least one point in the voxel grid, the center point of the voxel grid is used as the key feature point in this voxel grid;

[0022] Among them, the unique horizontal and vertical resolutions of the original point cloud are reconstructed into a voxel grid point cloud structure, breaking the problem of inconsistent point cloud data structure caused by different lidar models.

[0023] In a possible implementation, the size of the three-dimensional voxel grid ranges from 0.1 to 0.3 m.

[0024] In a possible implementation, the natural driving point cloud dataset includes the 32-line mechanical rotating lidar 3D target detection dataset NuScenes, the 40-line mechanical rotating lidar 3D target detection dataset Lyft, the 64-line mechanical rotating lidar 3D target detection dataset Kitti and Waymo, the solid-state lidar 3D target detection dataset SimoSet, and the 128-line mechanical rotating lidar 3D target detection dataset DAIR-V2X.

[0025] In a possible implementation, the autonomous driving point cloud dataset includes a SHIFT three-dimensional object detection dataset, a Livox three-dimensional object detection dataset, and a 51WORLD three-dimensional object detection dataset.

[0026] In a second aspect, the present application provides a device for enhancing the generalization of a target detection model considering a point cloud distribution structure, comprising:

[0027] A data acquisition module, used to acquire lidar point cloud data from a public simulated autonomous driving dataset and a natural driving dataset with true values ​​of three-dimensional target detection;

[0028] The point cloud distribution structure unification module is used to break the inconsistent horizontal and vertical resolution structure of point clouds in different models of lidar data sets through voxel filtering at the center of the lidar point cloud, obtain multiple lidar point cloud data sets with consistent horizontal and vertical resolutions, remove the dimension of the lidar point cloud reflection intensity, and highly unify the point cloud origin coordinates of multiple point cloud data sets to obtain multiple lidar point cloud data sets with unified point cloud distribution structure.

[0029] A data fusion module, used to fuse the processed multiple point cloud data sets into a training set, a test set and a validation set;

[0030] A model training module, used to train an existing three-dimensional object detection network based on the training set to obtain a strong generalization three-dimensional object detection model;

[0031] In a possible implementation, the point cloud distribution structure unification module further includes:

[0032] The voxel center filtering module is used to break the inconsistent horizontal and vertical resolution structure of point clouds in different models of lidar data sets, and obtain multiple lidar point cloud data sets with consistent horizontal and vertical resolutions;

[0033] The reflection intensity dimension reduction module is used to delete the reflection intensity features of the lidar point cloud and obtain multiple lidar point cloud data sets without the reflection intensity dimension;

[0034] The LiDAR installation height unification module is used to unify the LiDAR installation heights in different LiDAR data sets to obtain multiple LiDAR point cloud data sets with unified point cloud distribution structures.

[0035] The technical solution provided by this application can at least achieve the following beneficial effects:

[0036] The method for enhancing the generalization of the target detection model considering the point cloud distribution structure provided in this application not only unifies the data distribution structure of the lidar point cloud in the public data set on the basis of ensuring the real-time performance of the three-dimensional target detection model, but also improves the generalization of the lidar three-dimensional target detection model. More importantly, it helps to reduce manual data collection and labeling work, reduces the gap between the autonomous driving simulation data set and the natural driving scene data set, and provides an effective solution for accelerating the test operation and demonstration application of three-dimensional target detection technology in complex traffic scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0038] Figure 1 It is a flowchart of a method and device for enhancing the generalization of a target detection model considering the point cloud distribution structure, shown in an exemplary embodiment of the present application;

[0039] Figure 2 is a schematic structural diagram of a device for enhancing the generalization of a target detection model taking into account a point cloud distribution structure, shown in an exemplary embodiment of the present application;

[0040] Figure 3 It is a structural diagram of a point cloud distribution structure unification module shown in an exemplary embodiment of the present application.

[0041] Reference numerals:

[0042] 1. Data acquisition module; 2. Point cloud distribution structure unification module; 21. Voxel center filtering module; 22. Reflection intensity dimensionality reduction module; 23. LiDAR installation height unification module; 3. Data fusion module; 4. Model training module. DETAILED DESCRIPTION

[0043] In order to make the purpose, implementation mode and advantages of the present application clearer, the exemplary implementation mode of the present application will be clearly and completely described below in conjunction with the drawings in the exemplary embodiments of the present application. Obviously, the described exemplary embodiments are only part of the embodiments of the present application, not all of the embodiments. It should be understood that the specific embodiments described here are only used to explain the present application and are not used to limit the present application.

[0044] It should be noted that the brief description of terms in this application is only for the convenience of understanding the embodiments described below, and is not intended to limit the embodiments of this application. Unless otherwise specified, these terms should be understood according to their common and usual meanings.

[0045] The terms "first", "second", "third", etc. in the specification and claims of this application and the above drawings are used to distinguish similar or similar objects or entities, and do not necessarily mean to limit a specific order or sequence, unless otherwise noted. It should be understood that the terms used in this way can be interchangeable under appropriate circumstances.

[0046] The terms "comprises," "comprising," and "having," and any variations thereof, are intended to cover but not exclude inclusion, for example, a product or device comprising a list of components is not necessarily limited to all the components expressly listed but may include other components not expressly listed or inherent to such product or device.

[0047] Before explaining the method for enhancing the generalization of the target detection model considering the point cloud distribution structure provided in the embodiment of the present application, the application scenario and implementation environment of the embodiment of the present application are first introduced.

[0048] With the continuous development of modern artificial intelligence technology, autonomous driving has gradually emerged and become one of the effective solutions to problems such as traffic congestion and traffic accidents.

[0049] In order to meet the needs of autonomous driving technology in complex traffic scenarios, scholars at home and abroad have continuously innovated and proposed a series of advanced environmental perception algorithms and technologies to adapt to the data of different perception sensors. Among them, the three-dimensional target detection technology based on lidar plays an important role in the fields of autonomous driving and smart transportation because it can provide autonomous driving vehicles with information such as the category, location, and size of obstacles in traffic scenes. In recent years, deep learning models have been widely used by scholars at home and abroad because they can effectively capture potential features in the surrounding environment from lidar point cloud data.

[0050] However, deep learning models are often limited by the structure of the lidar point cloud in the dataset. This limitation stems from the differences in the horizontal and vertical resolutions of different models of lidars, which leads to significantly different distribution structures of the point cloud data. Therefore, in cases where the distribution structure of the lidar point cloud data is different from that in the dataset, the heterogeneity of the lidar point cloud makes the performance of deep learning models on test data often unsatisfactory. This difference may indirectly affect the generalization ability of the model in actual scenarios, further highlighting the challenges faced by the model when facing novel data.

[0051] Based on this, the present application provides a method and device for enhancing the generalization of a target detection model by considering the point cloud distribution structure. The point cloud data is reconstructed by voxelizing the lidar point cloud, breaking the distribution structure of the point cloud data of lidars of different scenes and models, and is used for the generalization research of three-dimensional target detection by vehicle-mounted lidar based on deep learning.

[0052] It can be seen that some embodiments of the present application aim to reduce the gap between the lidar point cloud data structure in the autonomous driving simulation data set and the natural driving data set, avoid the situation where the three-dimensional target detection model is limited to the point cloud data structure of the data set and reduce the generalization of the model, reduce a large amount of manual data collection and data labeling, and provide an effective solution for accelerating the test operation and demonstration application of three-dimensional target detection technology in complex traffic scenarios.

[0053] Next, the technical solution of the present application and how the technical solution of the present application solves the above-mentioned technical problems will be described in detail through embodiments and in combination with the accompanying drawings. The embodiments may be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. Obviously, the described embodiments are part of the embodiments of the present application, not all of them.

[0054] Figure 1 It is a flow chart of a method and device for enhancing the generalization of a target detection model considering the point cloud distribution structure, shown as an exemplary embodiment of the present application.

[0055] In an exemplary embodiment, Figure 1As shown, a method for enhancing the generalization of a target detection model considering the point cloud distribution structure is provided. In this embodiment, the method may include the following steps:

[0056] Step 100: Obtain lidar point cloud data from a public simulated autonomous driving dataset and a natural driving dataset with true three-dimensional target detection values.

[0057] Step 200: Through voxel filtering at the center of the lidar point cloud, the inconsistent horizontal and vertical resolution structures of the point clouds in the lidar data sets of different models are broken to obtain multiple lidar point cloud data sets with consistent horizontal and vertical resolutions, the dimension of the lidar point cloud reflection intensity is removed, and the point cloud origin coordinates of the multiple point cloud data sets are highly unified to obtain multiple lidar point cloud data sets with unified point cloud distribution structures.

[0058] Step 300: Fusing the processed multiple point cloud data sets into a training set, a test set and a validation set.

[0059] Step 400: Based on the training set, use an existing three-dimensional object detection network for training to obtain a three-dimensional object detection model.

[0060] In a possible implementation, the specific implementation of the generalization enhancement method of the three-dimensional object detection model is as follows:

[0061] Obtain point cloud data of different types of lidars in the currently public natural driving datasets, including but not limited to the 32-line mechanical rotating lidar 3D target detection dataset NuScenes, the 40-line mechanical rotating lidar 3D target detection dataset Lyft, the 64-line mechanical rotating lidar 3D target detection dataset Kitti and Waymo, the solid-state lidar 3D target detection dataset SimoSet, the 128-line mechanical rotating lidar 3D target detection dataset DAIR-V2X, etc.

[0062] Obtain the lidar point cloud data in the currently public simulated autonomous driving dataset, which may include but is not limited to the SHIFT 3D target detection dataset, the Livox 3D target detection dataset, and the 51WORLD 3D target detection dataset.

[0063] Through voxel filtering at the center of the lidar point cloud, the inconsistent horizontal and vertical resolution structures of the point clouds in different models of lidar data sets are broken; the dimension of the reflection intensity of the lidar point cloud is removed, and the point cloud origin coordinates of multiple point cloud data sets are highly unified to obtain multiple lidar point cloud data sets with unified point cloud distribution structures.

[0064] The point cloud voxelization in the current deep learning-based vehicle-mounted LiDAR 3D target detection network model is often divided into 3D voxel grids of a certain size according to the range of the LiDAR point cloud, and the LiDAR point cloud data is placed in the 3D voxel grid by projection, and random downsampling or upsampling is performed in each 3D voxel grid. However, this point cloud voxelization method often depends on the horizontal and vertical resolution of the LiDAR, and the model test effect is poor in point clouds with horizontal and vertical resolutions different from the LiDAR used in the data set.

[0065] Different from the existing point cloud voxelization, the point cloud center voxelization filtering in some embodiments of the present application refers to dividing the point cloud range into three-dimensional voxel grids of a certain size. For each voxel grid, if there is at least one point in the voxel grid, the center point of the voxel grid is used as the key feature point in this voxel grid. The horizontal and vertical resolutions unique to the original point cloud are reconstructed into a voxel grid point cloud structure, thereby breaking the problem of inconsistent horizontal and vertical resolutions of the point cloud caused by different LiDAR models.

[0066] The size of the 3D voxel grid is in the range of 0.1-0.3m, and the 3D target generalization performance is good. If the 3D voxel grid is large, it is easy to lose the key point features of traffic participants in the traffic scene; if the 3D voxel grid is small, the generalization of the model in the low-beam lidar dataset (such as the 32-line lidar 3D target detection dataset NuScenes) is less affected, because the reconstruction rate of the small voxel grid in the low-beam point cloud structure is low.

[0067] The processed multiple lidar data sets are fused and divided into training, testing, and validation sets.

[0068] Based on the training set, the existing 3D object detection network is used for training to obtain a 3D object detection model.

[0069] The 3D object detection network model is tested on the test set and validation set to prove the generalization of the model.

[0070] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the instructions, these steps are not necessarily executed in sequence according to the order of the instructions. Unless there is a clear explanation in this article, the execution of these steps is not strictly limited in order, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flowcharts involved in the above-mentioned embodiments may include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of the steps or stages in other steps.

[0071] Corresponding to the aforementioned embodiment of the method for enhancing the generalization of a target detection model considering the point cloud distribution structure, the present application also provides an embodiment of a device for enhancing the generalization of a target detection model considering the point cloud distribution structure, adopting the same technical concept.

[0072] Figure 2 It is a structural schematic diagram of a device for enhancing the generalization of a target detection model taking into account the point cloud distribution structure, as shown in an exemplary embodiment of the present application.

[0073] In an exemplary embodiment, Figure 2 As shown, the generalization enhancement device of the target detection model considering the point cloud distribution structure includes:

[0074] The data acquisition module 1 is used to acquire the lidar point cloud data in the public simulated autonomous driving dataset and the natural driving dataset with three-dimensional target detection true value.

[0075] The point cloud distribution structure unification module 2 is used to break the inconsistent horizontal and vertical resolution structure of point clouds in different models of laser radar data sets through voxel filtering at the center of the laser radar point cloud, and obtain multiple laser radar point cloud data sets with consistent horizontal and vertical resolutions. The point cloud distribution structure unification module is used to remove the dimension of the laser radar point cloud reflection intensity and highly unify the point cloud origin coordinates of multiple point cloud data sets to obtain multiple laser radar point cloud data sets with unified point cloud distribution structures.

[0076] The data fusion module 3 is used to fuse the processed multiple point cloud data sets and divide them into a training set, a test set and a validation set.

[0077] The model training module 4 is used to perform training based on the training set using an existing three-dimensional object detection network to obtain a highly generalized three-dimensional object detection model.

[0078] In a possible implementation, the method further includes:

[0079] A model verification module is used to test the three-dimensional object detection model based on the test set and the verification set.

[0080] Figure 3 It is a structural diagram of a point cloud distribution structure unification module shown in an exemplary embodiment of the present application.

[0081] In one possible implementation, Figure 3 As shown, the point cloud distribution structure unified module also includes:

[0082] The voxel center filtering module 21 is used to break the inconsistent horizontal and vertical resolution structure of point clouds in different models of laser radar data sets, and obtain multiple laser radar point cloud data sets with consistent horizontal and vertical resolutions.

[0083] The reflection intensity dimension reduction module 22 is used to delete the reflection intensity features of the laser radar point cloud to obtain multiple laser radar point cloud data sets without reflection intensity dimensions.

[0084] The laser radar installation height unification module 23 is used to unify the laser radar installation heights in different laser radar data sets to obtain multiple laser radar point cloud data sets with unified point cloud distribution structures.

[0085] For the specific definition of the device for enhancing the generalization of the target detection model considering the point cloud distribution structure, please refer to the definition of the method for enhancing the generalization of the target detection model considering the point cloud distribution structure mentioned above, which will not be repeated here. Each module in the above-mentioned device for enhancing the generalization of the target detection model considering the point cloud distribution structure can be implemented in whole or in part through software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0086] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0087] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. A method for enhancing the generalization of a target detection model considering the point cloud distribution structure, characterized in that: include: Obtain lidar point cloud data from public simulated autonomous driving datasets and natural driving datasets with true 3D object detection values; Through voxel filtering at the center of the LiDAR point cloud, the inconsistent structure of the horizontal and vertical resolutions of the point clouds in different models of LiDAR data sets is broken, and multiple LiDAR point cloud data sets with consistent horizontal and vertical resolutions are obtained. The dimension of the reflection intensity of the LiDAR point cloud is removed, and the coordinates of the point cloud origins of multiple point cloud data sets are highly unified, so as to obtain multiple LiDAR point cloud data sets with a unified point cloud distribution structure. The processed multiple point cloud data sets are fused and divided into a training set, a test set and a validation set; Based on the training set, an existing three-dimensional object detection network is used for training to obtain a strong generalization three-dimensional object detection model.

2. The method for enhancing the generalization of a target detection model considering the point cloud distribution structure according to claim 1, characterized in that: Also includes: The three-dimensional object detection model is tested based on the test set and the validation set.

3. The method for enhancing the generalization of a target detection model considering the point cloud distribution structure according to claim 1, characterized in that: The step of obtaining the laser radar point cloud data in the public simulated driving dataset and the natural driving dataset with the true value of three-dimensional target detection further includes: Obtain point cloud data of different types of lidars in the current public natural driving data set to obtain a natural driving point cloud data set; The point cloud data of the laser radar in the currently public simulated autonomous driving data set is obtained to obtain the autonomous driving point cloud data set.

4. The method for enhancing the generalization of a target detection model considering the point cloud distribution structure according to claim 3, characterized in that: The method of filtering the laser radar point cloud center voxelization to break the inconsistent structure of the point clouds in different models of laser radar data sets in horizontal and vertical resolutions, and obtaining multiple laser radar point cloud data sets with consistent horizontal and vertical resolutions, further includes: The point cloud voxel filtering method is used to break the data distribution structure unique to the lidar point cloud in the natural driving point cloud dataset and the autonomous driving point cloud dataset, and obtain multiple point cloud datasets.

5. The method for enhancing the generalization of a target detection model considering the point cloud distribution structure according to claim 1, characterized in that: The step of removing the dimension of the laser radar point cloud reflection intensity and highly unifying the point cloud origin coordinates of the multiple point cloud data sets to obtain multiple laser radar point cloud data sets with a unified point cloud distribution structure further includes: Divide the point cloud into a 3D voxel grid of a certain size according to its range; For each 3D voxel grid, if there is at least one point in the voxel grid, the center point of the voxel grid is used as the key feature point in this voxel grid; Among them, the unique horizontal and vertical resolutions of the original point cloud are reconstructed into a voxel grid point cloud structure, breaking the problem of inconsistent point cloud data structure caused by different lidar models.

6. The method for enhancing the generalization of a target detection model considering the point cloud distribution structure according to claim 5, characterized in that: The size of the three-dimensional voxel grid ranges from 0.1 to 0.3 m.

7. The method for enhancing the generalization of a target detection model considering the point cloud distribution structure according to claim 3, characterized in that: The natural driving point cloud datasets include the 32-line mechanical rotating lidar 3D target detection dataset NuScenes, the 40-line mechanical rotating lidar 3D target detection dataset Lyft, the 64-line mechanical rotating lidar 3D target detection dataset Kitti and Waymo, the solid-state lidar 3D target detection dataset SimoSet, and the 128-line mechanical rotating lidar 3D target detection dataset DAIR-V2X.

8. The method for enhancing the generalization of a target detection model considering the point cloud distribution structure according to claim 3, characterized in that: The autonomous driving point cloud dataset includes the SHIFT 3D target detection dataset, the Livox 3D target detection dataset, and the 51WORLD 3D target detection dataset.

9. A device for enhancing the generalization of a target detection model considering the point cloud distribution structure, characterized in that: include: A data acquisition module, used to acquire lidar point cloud data from a public simulated autonomous driving dataset and a natural driving dataset with true values ​​of three-dimensional target detection; The point cloud distribution structure unification module is used to break the inconsistent horizontal and vertical resolution structure of point clouds in different models of laser radar data sets through voxel filtering of the laser radar point cloud center, obtain multiple laser radar point cloud data sets with consistent horizontal and vertical resolutions, remove the dimension of laser radar point cloud reflection intensity, and highly unify the point cloud origin coordinates of multiple point cloud data sets to obtain multiple laser radar point cloud data sets with unified point cloud distribution structure; A data fusion module, used to fuse the processed multiple point cloud data sets into a training set, a test set and a validation set; The model training module is used to train the existing three-dimensional object detection network based on the training set to obtain a strong generalization three-dimensional object detection model.

10. The device for enhancing the generalization of a target detection model taking into account the point cloud distribution structure according to claim 9, characterized in that: The point cloud distribution structure unification module further includes: The voxel center filtering module is used to break the inconsistent horizontal and vertical resolution structure of point clouds in different models of lidar data sets, and obtain multiple lidar point cloud data sets with consistent horizontal and vertical resolutions; The reflection intensity dimension reduction module is used to delete the reflection intensity features of the lidar point cloud and obtain multiple lidar point cloud data sets without the reflection intensity dimension; The LiDAR installation height unification module is used to unify the LiDAR installation heights in different LiDAR data sets to obtain multiple LiDAR point cloud data sets with unified point cloud distribution structures.

Citation Information

Cited By

  • Method and system for generating laser radar point cloud test data based on interpretable analysis

    CN122453830A