Radar data generation and automatic labeling method and related apparatus

Abstract

The application discloses a radar data generation and automatic labeling method and related devices, by acquiring the original radar data and the corresponding auxiliary data obtained by the radar equipment collecting the real target scene, and inputting the original radar data and the auxiliary data into the first coding network to obtain the first variable data, and applying physical consistency constraints to the first variable data to obtain the second variable data aligned with the actual physical quantity, and generating simulated radar data according to the second variable data, and outputting the dynamic state information, existence information and / or trajectory identification information label of the target according to the effective scene state variable in the second variable data, and screening the label according to the credibility score, so that the data authenticity of the simulated radar data and the labeling reliability can be considered without relying on the existing labeling of the original radar data.

Description

Technical Field

[0001] This invention relates to the field of radar data generation and automatic annotation technology, and in particular to a radar data generation and automatic annotation method and related apparatus. Background Technology

[0002] Millimeter-wave radar possesses characteristics such as resistance to changes in lighting conditions, all-weather operation, and the ability to directly measure radial velocity, making it highly valuable in fields such as advanced driver assistance systems, autonomous driving, intelligent transportation, and mobile robotics. The raw data acquired by radar systems needs to be processed by radar algorithms to be converted into effective information that humans can recognize, such as target location, velocity, and type. Developing radar algorithms requires not only a large amount of radar data as a training set but also annotation of the radar data to provide training basis and evaluation criteria for the algorithm. However, current radar algorithm development has long been constrained by data and annotation bottlenecks, resulting in low model verification efficiency and insufficient scene coverage for radar algorithm validation.

[0003] In existing technologies, there are generally three methods for acquiring radar data for radar algorithm training and / or for intelligent annotation of radar data: traditional physical simulation, purely generative data fitting, and cross-modal annotation or conversion schemes. Each of these methods has its advantages and disadvantages, but none can simultaneously ensure data authenticity, physical interpretability, and annotation credibility.

[0004] Traditional physical simulation methods typically generate radar data using physical simulation models built with software such as MATLAB (MATrix Laboratory) to quickly verify the feasibility of algorithms. While this approach offers high interpretability, its drawbacks include overly idealized data that struggles to fully simulate real-world application scenarios, such as multipath propagation, clutter, obstruction, material differences, and channel inconsistencies. This results in a significant gap between simulation and reality. Many scenarios that pass in simulation may still fail in actual deployment.

[0005] Purely generative data fitting schemes, typically including Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), and Diffusion models, can generate data that approximates the true distribution, but the latent variables often lack stable physical semantics. While these methods may produce realistic outputs, they struggle to directly derive labels with clear physical meanings, such as position, velocity, radial velocity, or trajectory, resulting in insufficient interpretability.

[0006] Cross-modal annotation or conversion schemes: These schemes evaluate algorithm performance in real-world environments by collecting radar data from actual roads. However, such data often lacks readily available ground truth annotations, relying primarily on human interpretation or additional sensors such as cameras and lidar for cross-modal alignment. This approach is not only costly but also susceptible to issues such as timestamp discrepancies, extrinsic parameter calibration errors, and differences in physical observations, making it difficult to stably and quantitatively evaluate the algorithm's performance.

[0007] In summary, existing technologies suffer from two shortcomings: "simulated data is not realistic enough" and "real road test data lacks high-quality annotation." Therefore, there is a need to propose a new radar data generation and automatic annotation method and related devices to solve these problems. Summary of the Invention

[0008] In view of the above problems, the purpose of this invention is to provide a radar data generation and automatic annotation method and related apparatus, so as to balance the data authenticity and annotation reliability of simulated radar data without relying on the existing annotations of the original radar data.

[0009] According to one aspect of the present invention, a radar data generation and automatic annotation method is provided, comprising: acquiring raw radar data and corresponding auxiliary data obtained by radar equipment collecting real target scenes; inputting the raw radar data and auxiliary data into a first encoding network to obtain first variable data; applying physical consistency constraints to the first variable data to obtain second variable data aligned with actual physical quantities, the second variable data including at least effective scene state variables; and generating simulated radar data based on the second variable data, and annotating the target's dynamic state information, existence information, and / or trajectory identification information based on the effective scene state variables, wherein the raw radar data and its corresponding auxiliary data are obtained by radar equipment collecting target scenes.

[0010] Optionally, the method further includes calculating the credibility score of the corresponding simulated radar data based on an evaluation mechanism based on multi-dimensional physical constraints and self-consistency verification, and classifying the corresponding labels into high-credibility labels, weakly supervised labels, ignored labels, or labels to be manually reviewed according to the credibility score.

[0011] Optionally, the method further includes inputting the simulated radar data into a second encoding network to obtain inverse scene state variables; and establishing reversible consistency constraints based on the difference between the valid scene state variables in the second variable data and the inverse scene state variables.

[0012] Optionally, the evaluation mechanism based on multi-dimensional physical constraints and self-consistency verification calculates the credibility score of the corresponding simulated radar data by calculating the credibility score based on at least one of physical consistency, temporal stability, reversible consistency and cross-representation alignment.

[0013] Optionally, the second variable data may also include effective echo characteristic variables and effective sensor characteristic variables.

[0014] Optionally, generating simulated radar data based on the second variable data includes inputting the second variable data into a conditional generation network to obtain simulated radar data; or reconfiguring at least one variable in the second variable data and inputting the configured second variable data into a conditional generation network to generate simulated radar data under different scenarios and / or different imaging conditions.

[0015] Optionally, the method further includes generating non-real target echo information based on the effective echo characteristic variables and / or effective sensor characteristic variables, and outputting non-real target echo information annotations, as well as outputting the target's 2D / 3D bounding box, category, occupied grid, and semantic region annotations based on the effective scene state variables. The non-real target echo information includes multipath reflection, ghosting, sidelobes, clutter, or imaging anomaly information, and the non-real target echo information annotations and target-related annotations are managed separately.

[0016] Optionally, the first coding network is a structured coding network, and the second coding network is an inversion network or a structured coding network.

[0017] Optionally, the raw radar data includes range-Doppler sequences and range-azimuth sequences; the auxiliary data includes timestamps, vehicle speed, angular velocity, and inertial measurement units; the target's dynamic state information includes at least one of position, velocity vector, radial velocity, and trajectory identification information.

[0018] According to a second aspect of the present invention, a radar data generation and automatic annotation system is provided, comprising: a first encoding network for receiving raw radar data and auxiliary data, and generating first variable data; a physical consistency constraint module for applying physical consistency constraints to the first variable data to obtain second variable data aligned with actual physical quantities, the second variable data including at least valid scene state variables; a data configuration module for reconfiguring at least one variable in the second variable data and outputting it, or directly outputting the second variable data; a conditional generation network for receiving the output signal of the data configuration module to generate simulated radar data; a second encoding network for obtaining reverse scene state variables based on the simulated radar data; a credibility evaluation unit for establishing reversible consistency constraints based on the differences between the valid scene state variables and the reverse scene state variables in the second variable data, and generating a credibility score based on at least one of physical consistency, temporal stability, reversible consistency, and cross-representation alignment; an information output unit for outputting dynamic state information, existence information, and / or trajectory identification information of a target based on the valid scene state variables; and an annotation unit for annotating the output signals of the information output unit and the credibility evaluation unit.

[0019] According to a third aspect of the present invention, an electronic device is provided, including a processor and a storage device, wherein the storage device stores program instructions that, when executed by the processor, implement the steps of the method described above.

[0020] According to a fourth aspect of the present invention, a storage medium is provided that stores a computer program or instructions, which, when executed by a processor, implement the steps of the method described above.

[0021] The radar data generation and automatic annotation method and related apparatus provided by this invention obtain first variable data based on the original radar data, giving the variable data used to generate simulated radar data a clear physical semantics, facilitating the extraction of annotation information such as position, velocity, and trajectory. Furthermore, through physical consistency constraints (such as temporal dynamic constraints, vehicle motion compensation constraints, and radar geometric projection constraints) and reversible consistency verification, the generated simulated radar data can be ensured to be physically correct and self-consistent, improving the credibility of the annotation. The second variable data obtained based on the physical consistency constraints can directly output the target's dynamic state information (such as position and velocity) and existence information, as well as annotations of non-real target echo information (such as multipath reflections, ghosting, and other non-real target echoes), without relying on additional sensors such as cameras and lidar, reducing annotation costs and avoiding cross-modal alignment errors.

[0022] In a preferred embodiment, by reconfiguring at least one variable in the first variable data, simulated radar data under different scene states and imaging conditions can be obtained, thereby satisfying the algorithm training requirements for diversity and realism, and solving the problem that simulated radar data is not realistic enough. Attached Figure Description

[0023] The above and other objects, features and advantages of the present invention will become more apparent from the following description of embodiments of the invention with reference to the accompanying drawings, in which:

[0024] Figure 1 A flowchart illustrating a radar data generation and automatic annotation method according to an embodiment of the present invention is shown;

[0025] Figure 2 A schematic diagram of a radar data generation and automatic annotation system according to an embodiment of the present invention is shown;

[0026] Figure 3 A schematic diagram illustrating physical consistency constraints on the first variable data according to an embodiment of the present invention is shown;

[0027] Figure 4 A schematic diagram of a reversible closed-loop process for forward generation and reverse parsing of simulated radar data according to an embodiment of the present invention is shown.

[0028] Figure 5 A schematic diagram illustrating the process of calculating and classifying the credibility score of annotations according to an embodiment of the present invention is shown;

[0029] Figure 6 A schematic diagram illustrating the data configuration of the second variable data and the generation of the simulated radar dataset according to an embodiment of the present invention is shown. Detailed Implementation

[0030] Various embodiments of the invention will now be described in more detail with reference to the accompanying drawings. In the various drawings, the same elements or modules are indicated by the same or similar reference numerals. For clarity, the various parts in the drawings are not drawn to scale.

[0031] It should be understood that, in the following description, "circuit" may include single or combined hardware circuits, programmable circuits, state machine circuits, and / or elements capable of storing instructions executed by the programmable circuit. When an element or circuit is said to be "connected" to another element or "connected" between two nodes, it may be directly coupled or connected to the other element, or there may be intermediate elements; the connection between elements may be physical, logical, or a combination thereof. Conversely, when an element is said to be "directly coupled to" or "directly connected" to another element, it means that there are no intermediate elements between them.

[0032] Furthermore, certain terms are used in this patent specification and claims to refer to specific components. Those skilled in the art will understand that hardware manufacturers may use different names to refer to the same component. This patent specification and claims do not distinguish components based on differences in name, but rather on differences in function.

[0033] Furthermore, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0034] Figure 1 A flowchart illustrating a radar data generation and automatic annotation method according to an embodiment of the present invention is shown.

[0035] See Figure 1 The radar data generation and automatic annotation method provided in this embodiment of the invention includes steps S01-S07.

[0036] In step S01, the raw radar data and corresponding auxiliary data are acquired.

[0037] The original radar data can be unlabeled; if the original radar data has existing labels, this method can also omit the existing labels and generate corresponding labels through subsequent scene state variables, physical consistency constraints, reversible consistency verification, and credibility assessment mechanisms. The original radar data and its corresponding auxiliary data are obtained by radar equipment collecting data from the target scene. The radar is preferably a millimeter-wave radar. The original radar data includes range-Doppler (RD) sequences and range-azimuth (RA) sequences. Auxiliary data includes timestamps, vehicle speed, angular velocity, and inertial measurement unit (IMU) data.

[0038] The target scenario is a real-world scenario such as a road or water area. The raw radar data can be obtained by continuously acquiring N (N>1) frames of the target scenario using a millimeter-wave radar device. The raw radar data includes the RD sequence and RA sequence corresponding to each frame, and the corresponding auxiliary data includes the timestamp, vehicle speed, angular velocity, and IMU data for each frame.

[0039] Furthermore, the raw radar data also includes ADC raw data, radar echo (RE) data, point cloud, and target list (Track List) data. Among them, the ADC raw data is obtained by sampling the analog echo signal received by the radar through an analog-to-digital converter.

[0040] In step S02, the raw radar data and auxiliary data are input into the first coding network to obtain the first variable data.

[0041] The first encoding network employs a structured encoding network. For example, the first encoding network can be a learnable encoding network capable of extracting features from radar data sequences and performing spatiotemporal modeling. This first encoding network can be implemented using convolutional neural networks, Transformers or other attention networks, state-space models, recurrent neural networks, graph neural networks, self-supervised encoding networks, generative encoding networks, or combinations thereof; it can also employ other neural network architectures capable of outputting structured scene state variables. This application does not limit the specific backbone network of the first encoding network, but focuses on its structured output of scene state variables corresponding to physical quantities.

[0042] Furthermore, the first variable data includes the scene state variable s_t, and optionally includes the echo characteristic variable z_echo,t, and the sensor characteristic variable z_sensor,t. The echo characteristic variable z_echo,t characterizes scattering intensity, micro-Doppler, multipath, occlusion, or imaging conditions, while the sensor characteristic variable z_sensor,t characterizes sensor noise, waveform settings, channel differences, device deviation, or external parameter errors.

[0043] For example, in this application, the scene state variable s_t={ o_{t,k}}, where t represents the time index or frame index, and k = 1 ... K.

[0044] Where k represents the target slot or target index. o_{t,k}= ( r_{t,k}, a_{t,k}, v_{r,t,k},v_{x,t,k}, v_{y,t,k}, e_{t,k}, id_{t,k} ). r_{t,k} represents the distance or position parameter, a_{t,k} represents the azimuth parameter, v_{r,t,k} represents the radial velocity, v_{x,t,k} and v_{y,t,k} represent the planar velocity components, e_{t,k} represents the existence score, and id_{t,k} represents the trajectory recognition information or trajectory token.

[0045] In step S03, a physical consistency constraint is applied to the first variable data to obtain second variable data that is aligned with the actual physical quantity.

[0046] Physical consistency constraints include at least one of the following: temporal dynamics constraints, vehicle motion compensation constraints, and radar geometric projection constraints. For example, ... Figure 3 As shown, the physical consistency constraints include temporal dynamic constraints, vehicle motion compensation constraints, and radar geometric projection constraints.

[0047] In step S04, simulated radar data is generated based on the second variable data.

[0048] The simulated radar data includes a time-series RD sequence RD_t and a time-series RA sequence RA_t. The time-series RD sequence RD_t includes the RD matrix of each frame arranged sequentially according to time frames, and the time-series RA sequence RA_t includes the RA matrix of each frame arranged sequentially according to time frames. The second variable data includes the effective scene state variable s1_t.

[0049] Furthermore, the second variable data also includes the effective echo characteristic variable z1_echo,t and the effective sensor characteristic variable z1_sensor,t.

[0050] The step of generating simulated radar data based on the second variable data includes: directly inputting the second variable data into a conditional generation network to obtain simulated radar data, or as follows: Figure 6 The method involves reconfiguring at least one variable in the second variable data and inputting the configured second variable data into a conditional generation network to generate simulated radar data under different scenarios and / or different imaging conditions. For example, reconfiguring at least one variable in the second variable data includes editing the effective scenario state variable s1_t to form an effective scenario state variable s1_t corresponding to a scenario in which a specific target is inserted, deleted, translated, accelerated, decelerated, or occluded in the target scenario.

[0051] For example, reconfiguring at least one variable in the second variable data includes: resampling or perturbing the effective echo characteristic variable z1_echo,t and / or the effective sensor characteristic variable z1_sensor,t to change the imaging conditions.

[0052] In step S05, the dynamic state information, existence information and / or trajectory recognition information of the target are labeled according to the effective scene state variable s1_t.

[0053] For example, dynamic state information, existence information, and trajectory recognition information of a target can be obtained through at least one of the analytical function, projection function, and mapping function.

[0054] In step S05, the valid scene state variable s1_t includes both edited and unedited valid scene state variables s1_t. Dynamic state information includes at least one of position, velocity vector, radial velocity, and trajectory recognition information. Position is obtained by coordinate transformation of r_{t,k} and a_{t,k}; radial velocity is directly obtained from v_{r,t,k}; planar velocity is derived from v_{x,t,k}, v_{y,t,k}, or multi-frame relationships; existence information is obtained by thresholding e_{t,k}; and trajectory recognition information is obtained from id_{t,k} or cross-frame association results.

[0055] In addition, based on the effective scene state variable s1_t, the target's 2D / 3D bounding box, category, occupied grid, and semantic region annotation can also be output.

[0056] In step S06, simulated radar data is input into the second encoding network to obtain the reverse scene state variable s2_t, and a reversible consistency constraint is established based on the difference between the valid scene state variable s1_t and the reverse scene state variable s2_t in the second variable data.

[0057] For example, the second coding network can be an inversion network. For example, when the first coding network is a structured coding network, simulated radar data can also be input back into the structured coding network, in which case the second coding network is a structured coding network.

[0058] The process of establishing reversible consistency constraints is as follows: Figure 4As shown, the original radar data and auxiliary data are forward encoded by the first encoding network to obtain the first variable data, and then subjected to physical consistency constraints to obtain the second variable data including the effective scene state variable s1_t. After the second variable data is configured, it is input into the conditional generation network for conditional generation to obtain simulated radar data. Then, the simulated radar data is reverse encoded by the second encoding network to obtain the reverse scene state variable s2_t. Finally, a reversible consistency constraint is established based on the effective scene state variable s1_t and the reverse scene state variable s2_t, and the reversible consistency loss is calculated.

[0059] In step S07, the credibility score of the simulated radar data is calculated based on the evaluation mechanism of multi-dimensional physical constraints and self-consistency verification, and the corresponding labels are divided into high credibility labels, weak supervision labels, ignored labels, or labels to be manually reviewed according to the credibility score.

[0060] The evaluation mechanism based on multi-dimensional physical constraints and self-consistency verification calculates the credibility score of simulated radar data by calculating the credibility score based on at least one of physical consistency, temporal stability, reversible consistency, and cross-representation alignment. Cross-representation alignment characterizes the consistency between the projection position, peak response, characteristic response, or target association results of the same target or scene state across different radar representations such as range-Doppler sequences, range-azimuth sequences, radar echoes, point clouds, or target lists.

[0061] Furthermore, such as Figure 5 As shown, physical consistency score, reversible consistency score, temporal stability score, and cross-representation alignment score can be calculated based on annotation information (including target-related annotation information and annotation information of non-real target echoes). The corresponding annotations are then classified into high-confidence annotations, weakly supervised annotations, ignored annotations, or annotations requiring manual review based on the confidence score. For example, when the confidence score is greater than a first preset value, the simulated radar data is labeled as a high-confidence annotation; when the confidence score is less than or equal to the first preset value but greater than a second preset value, the simulated radar data is labeled as a weakly supervised annotation; when the confidence score is less than or equal to the second preset value but greater than a third preset value, and the simulated radar data and / or the annotation is classified as Class I data, the simulated radar data is labeled as an annotation requiring manual review; when the confidence score is less than or equal to the third preset value, or the simulated radar data and / or the annotation is classified as Class II data, the simulated radar data is labeled as an ignored annotation. For annotations requiring manual review, the annotation information can be updated through rule correction or retraining after manual review.

[0062] Furthermore, the first category of data includes at least one of the following: simulated radar data and / or the scene corresponding to the annotation belongs to a rare scene, a boundary scene, a difficult scene, or a cross-representation score conflict but still has data value; the second category of data includes at least one of the following: simulated radar data and / or annotation has serious physical inconsistencies, cannot establish target association, the generated result cannot be backed up to the effective scene state, or the annotation is obviously unusable.

[0063] For example, the confidence score score_{t,k} can be obtained by weighting at least two of the following: geometric projection score_phy, temporal dynamics score_tmp, invertible consistency score_inv, cross-representation alignment score_xview, vehicle motion compensation consistency score_ego, and existence and sparsity score_exist.

[0064] Preferably, the confidence score score_{t,k} can be obtained by weighting the geometric projection score score_phy, the temporal dynamics score score_tmp, the reversible consistency score score_inv, and the cross-representation alignment score_xview, i.e., score_{t,k}=α1·score_phy+α2·score_tmp+α3·score_inv+α4·score_xview.

[0065] Among them, α1-α4 are adjustable parameters. The geometric projection score_phy reflects whether the geometric projection, vehicle compensation, and dynamic residual are sufficiently small; the temporal dynamics score_tmp reflects whether the state changes between adjacent frames are smooth and reasonable; the reversible consistency score_inv reflects whether the difference between the preset scene state variable s1_t and the reverse scene state variable s2_t is lower than the preset value; and the cross-representation alignment score_xview reflects whether the correspondence of the same target in the RD sequence and RA sequence is consistent.

[0066] Furthermore, the radar data generation and automatic annotation method provided in this embodiment of the invention also includes: generating non-real target echo information based on the effective echo characteristic variable z1_echo,t and / or the effective sensor characteristic variable z1_sensor,t, and outputting annotations for the non-real target echo information. The non-real target echo information includes ghost images, multipath artifacts, clutter responses, and sidelobe remnants.

[0067] Furthermore, non-real target echo information annotations and target-related annotations are managed separately. Target-related annotations include the target's dynamic state information, existence information, as well as 2D / 3D bounding boxes, categories, occupied grids, semantic region annotations, etc.

[0068] Furthermore, such as Figure 6 As shown, simulated radar data can be combined with its corresponding annotations to form a simulated radar dataset before being output.

[0069] Figure 2 A schematic diagram of a radar data generation and automatic annotation system according to an embodiment of the present invention is shown.

[0070] See Figure 2 The radar data generation and automatic annotation system 100 provided in this embodiment of the invention includes a first encoding network 110, a physical consistency constraint module 120, a data configuration module 130, a conditional generation network 140, a second encoding network 150, a credibility evaluation unit 160, an information output unit 170, and an annotation unit 180.

[0071] The first encoding network 110 is used to receive raw radar data and auxiliary data, and generate first variable data. The first variable data includes scene state variable s_t, and optionally includes echo characteristic variable z_echo,t, sensor characteristic variable z_sensor,t, etc.

[0072] The physical consistency constraint module 120 is used to apply physical consistency constraints to the first variable data to obtain the second variable data. The second variable data includes the scene state variable s1_t, and optionally includes the echo characteristic variable z1_echo,t, the sensor characteristic variable z1_sensor,t, etc.

[0073] The data configuration module 130 is used to reconfigure at least one variable in the second variable data and then output it, or to directly output the second variable data.

[0074] The conditional generation network 140 is used to receive the output of the data configuration module 130 to generate simulated radar data.

[0075] The second coding network 150 is used to obtain the reverse scene state variable s2_t based on the simulated radar data.

[0076] The credibility assessment unit 160 is used to establish reversible consistency constraints based on the difference between the effective scenario state variable s1_t and the reverse scenario state variable s2_t, and generate a credibility score based on at least one of temporal stability, physical consistency, reversible consistency and cross-representation alignment.

[0077] The information output unit 170 is used to output the target's dynamic state information and / or existence information and / or non-real target echo information based on the second variable data.

[0078] The annotation unit 180 is used to annotate the output signals of the credibility assessment unit 160 and the information output unit 170.

[0079] Furthermore, the present invention also provides an electronic device, including a processor and a storage device, wherein the storage device stores program instructions that, when executed by the processor, implement the steps of the method described above.

[0080] Furthermore, the present invention also provides a storage medium storing a computer program or instructions, which, when executed by a processor, implement the steps of the method described above.

[0081] The radar data generation and automatic annotation method and related apparatus provided by this invention obtain first variable data based on the original radar data, giving the variable data used to generate simulated radar data a clear physical semantics, facilitating the extraction of annotation information such as position, velocity, and trajectory. Furthermore, through physical consistency constraints (such as temporal dynamic constraints, vehicle motion compensation constraints, and radar geometric projection constraints) and reversible consistency verification, the generated simulated radar data can be ensured to be physically correct and self-consistent, improving the credibility of the annotation. The second variable data obtained based on the physical consistency constraints can directly output the target's dynamic state information (such as position and velocity) and existence information, as well as annotations of non-real target echo information (such as multipath reflections, ghosting, and other non-real target echoes), without relying on additional sensors such as cameras and lidar, reducing annotation costs and avoiding cross-modal alignment errors.

[0082] In a preferred embodiment, by reconfiguring at least one variable in the first variable data, simulated radar data under different scene states and imaging conditions can be obtained, thereby satisfying the algorithm training requirements for diversity and realism, and solving the problem that simulated radar data is not realistic enough.

[0083] As described above, these embodiments of the present invention do not exhaustively describe all details, nor do they limit the invention to specific embodiments. Clearly, many modifications and variations can be made based on the above description. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to effectively utilize the invention and its modifications. The scope of protection of this invention should be determined by the scope defined in the claims and their equivalents.

Claims

1. A method for generating and automatically labeling radar data, comprising: Acquire raw radar data and corresponding auxiliary data obtained from real target scenes collected by radar equipment, and input the raw radar data and auxiliary data into a first coding network to obtain first variable data; Physical consistency constraints are applied to the first variable data to obtain second variable data that is aligned with the actual physical quantities. The second variable data includes at least valid scenario state variables. as well as Simulated radar data is generated based on the second variable data, and the target's dynamic state information, existence information, and / or trajectory identification information are labeled according to the effective scene state variables.

2. The method according to claim 1, further comprising: An evaluation mechanism based on multi-dimensional physical constraints and self-consistency verification calculates the credibility score of the corresponding simulated radar data, and classifies the corresponding labels into high-credibility labels, weakly supervised labels, ignored labels, or labels to be manually reviewed based on the credibility score.

3. The method according to claim 2, further comprising: The simulated radar data is input into the second coding network to obtain the reverse scene state variables; as well as A reversible consistency constraint is established based on the difference between the valid scenario state variables in the second variable data and the reverse scenario state variables.

4. The method according to claim 3, wherein, The evaluation mechanism based on multi-dimensional physical constraints and self-consistency verification calculates the credibility score of the corresponding simulated radar data, including: The credibility score is calculated based on at least one of physical consistency, temporal stability, reversible consistency, and cross-representation alignment.

5. The method according to claim 1, wherein, The second variable data also includes effective echo characteristic variables and effective sensor characteristic variables.

6. The method according to claim 5, wherein, The step of generating simulated radar data based on the second variable data includes: The second variable data is input into a conditional generation network to obtain simulated radar data; or At least one variable in the second variable data is reconfigured, and the reconfigured second variable data is input into a conditional generation network to generate simulated radar data under different scenarios and / or different imaging conditions.

7. The method according to claim 5, further comprising: Based on the effective echo characteristic variables and / or effective sensor characteristic variables, non-real target echo information is generated, and a label for the non-real target echo information is output. Based on the valid scene state variables, output the target's 2D / 3D bounding box, category, occupied grid, and semantic region annotation. The non-real target echo information includes multipath reflection, ghosting, sidelobes, clutter, or imaging anomaly information. The non-real target echo information annotation and the target-related annotation are managed separately.

8. The method according to claim 3, wherein, The first coding network is a structured coding network, and the second coding network is an inversion network or a structured coding network.

9. The method according to claim 1, wherein, The raw radar data includes range-Doppler sequences and range-azimuth sequences; the auxiliary data includes timestamps, vehicle speed, angular velocity, and inertial measurement units; the target's dynamic state information includes at least one of position, velocity vector, radial velocity, and trajectory identification information.

10. A radar data generation and automatic annotation system, comprising: The first coding network is used to receive raw radar data and auxiliary data, and generate first variable data; The physical consistency constraint module is used to apply physical consistency constraints to the first variable data to obtain second variable data that is aligned with the actual physical quantities. The second variable data includes at least valid scenario state variables. The data configuration module is used to reconfigure at least one variable in the second variable data and then output it, or to directly output the second variable data. A conditional generation network is used to receive the output signal from the data configuration module to generate simulated radar data; The second encoding network is used to obtain reverse scene state variables based on the simulated radar data; The credibility assessment unit establishes reversible consistency constraints based on the differences between the valid scenario state variables and the reverse scenario state variables in the second variable data, and generates a credibility score based on at least one of physical consistency, temporal stability, reversible consistency and cross-representation alignment. The information output unit is used to output the target's dynamic state information, existence information, and / or trajectory recognition information based on the effective scene state variables. The annotation unit is used to annotate the output signals of the information output unit and the credibility evaluation unit.

11. An electronic device comprising a processor and a storage device, the storage device storing program instructions that, when executed by the processor, implement the steps of the method as described in any one of claims 1 to 9.

12. A storage medium storing a computer program or instructions that, when executed by a processor, implement the steps of the method as described in any one of claims 1 to 9.

Images