Automatic driving simulation annotation data set generation method

Abstract

The invention relates to the technical field of data annotation, and particularly discloses an automatic driving simulation annotation data set generation method which comprises the following steps: dynamically loading an executable file of a data set generation tool through a plug-in interface of an automatic driving simulation platform, and importing the executable file through a high-precision map, a complex road network model containing multiple lanes, traffic signs and signal lamps is constructed, a simulation data set is constructed through simulation software, and the simulation software can generate various simulation scenes of different road conditions, weather conditions, traffic flow and the like and conditions which are difficult to encounter or too dangerous in the real world. Not only can the cost and difficulty of real vehicle data acquisition be reduced, but also the real-time position data of the dynamic target object acquired by the laser radar sensor in the simulation process can be corrected by combining hardware parameters, software dynamic response parameters and simulation platform operation parameters; the accuracy of data annotation can be improved, and higher data annotation efficiency can be ensured.

Description

Technical Field

[0001] This invention relates to the field of data annotation technology, specifically to a method for generating an autonomous driving simulation annotation dataset. Background Technology

[0002] Currently, there is a close relationship between autonomous driving algorithms and labeled datasets. Autonomous driving algorithms are based on deep learning models trained on a large amount of labeled data, so they need to be based on labels that cover various scenarios and conditions to ensure that the algorithms can cope with various complex situations in practical applications.

[0003] In existing technologies, simulation-annotated datasets have emerged to improve the accuracy of data annotation. Simulation data can be generated by computers, eliminating the need for tedious manual collection and annotation, significantly saving time and labor costs. Furthermore, simulation data inherently contains annotation information generated by computer programs, resulting in high annotation accuracy and good consistency. Simulation data can also easily generate various complex scenarios and extreme situations that may be difficult to obtain during actual data collection, thus providing accurate annotation data for autonomous driving algorithms.

[0004] However, in existing technologies, simulation annotation datasets rely on simulation platforms. During the operation of these platforms, abnormalities in their running status and hardware and software characteristics can cause delays in data transmission and processing, resulting in lag in the data acquired by LiDAR. Under these circumstances, errors can occur in the annotation data of dynamic targets, affecting the accuracy of the annotation data of dynamic targets. Summary of the Invention

[0005] The purpose of this invention is to provide a method for generating an autonomous driving simulation annotation dataset, thereby solving the following technical problems:

[0006] How to improve the accuracy of dynamic target object annotation data.

[0007] The objective of this invention can be achieved through the following technical solutions:

[0008] A method for generating an autonomous driving simulation annotation dataset, the method comprising:

[0009] S1: Dynamically load the executable file of the dataset generation tool through the plug-in interface of the autonomous driving simulation platform, and import it through high-precision maps to build a complex road network model containing multiple lanes, traffic signs, and traffic lights.

[0010] S2: Based on the modeling data of autonomous driving scenarios, add multiple simulated sensors to the simulated vehicle, including cameras, LiDAR, and GPS;

[0011] S3: Complete the configuration of the *.json file for dataset parameter information to support dynamic response analysis, providing hardware parameters, software dynamic response parameters, and simulation platform operating parameters;

[0012] S4: Start the simulation and, during the simulation run, use the dataset generation tool to extract the required dynamic target and static traffic sign information from the data generated by each sensor, the detected target information, and the road network file of the simulation platform.

[0013] S5: By combining hardware parameters, software dynamic response parameters and simulation platform operating parameters, the real-time position data of dynamic targets collected by the lidar sensor during the simulation process is corrected.

[0014] S6: Combine the real-time location data of the corrected dynamic target objects and save the data to the dynamic target object dataset path specified in the configuration file at each time step;

[0015] S7: After the simulation is completed, the extracted static traffic sign information will be saved in the static traffic sign dataset path specified in the configuration file.

[0016] Furthermore, the data provided in S3 includes:

[0017] The hardware parameters, software dynamic response parameters, and simulation platform operating parameters include hardware latency, software dynamic response parameters including software latency, and simulation platform operating parameters including filtering algorithm processing time, bandwidth, and data transmission volume.

[0018] Furthermore, the correction process in S5 includes:

[0019] Through formula Calculate the corrected bandwidth size for the i-th data extraction from the simulation platform. ;

[0020] Where i represents any data extraction from the simulation platform. This represents the total number of data extractions performed by the simulation platform during the i-th data extraction. Let be the bandwidth of the simulation platform during the i-th data extraction. For all The average value, This is a proportionality coefficient, set based on empirical fitting.

[0021] Furthermore, the correction process in S5 also includes:

[0022] Through formula Calculate the transmission delay during the i-th data extraction from the simulation platform. ;

[0023] in, This represents the amount of data transmitted during the i-th data extraction by the simulation platform.

[0024] Furthermore, the correction process in S5 also includes:

[0025] Through formula Calculate the response delay risk coefficient for the i-th data extraction from the simulation platform. ;

[0026] in, The hardware latency during the i-th data extraction by the simulation platform. The software delay during the i-th data extraction by the simulation platform. The preset software delay, To define a function, if Then let Otherwise, let , This is the standard value for delay. , and The weighting coefficients are set based on empirical fitting. The preset bandwidth size, This represents the processing time of the filtering algorithm during the i-th data extraction in the simulation platform. For all The average value.

[0027] Furthermore, the correction process in S5 also includes:

[0028] By obtaining the response delay risk coefficient of the simulation platform during the i-th data extraction. Compared with the preset response delay risk coefficient threshold Perform a comparison;

[0029] like It was determined that there was a serious response delay during the i-th data extraction of the simulation platform, and the dynamic target calibration position during the i-th data extraction needed to be corrected.

[0030] like It is determined that there is no response delay during the i-th data extraction of the simulation platform, and there is no need to correct the dynamic target calibration position during the i-th data extraction.

[0031] Furthermore, the correction process in S5 also includes:

[0032] Through formula Calculate the position error distance of the a-th dynamic target object during the i-th data extraction on the simulation platform. ;

[0033] Where 'a' is any dynamic target object. The distance between the a-th dynamic target and the simulated vehicle during the i-th data extraction by the simulation platform. The adjustment coefficient lookup table function has a range of values ​​that are related to... The numerical values ​​correspond one-to-one.

[0034] Furthermore, the dynamic loading process in S1 includes:

[0035] During the simulation, the dataset generation tool uses the SDK provided by the simulation platform. The executable file of the dataset generation tool is installed in the corresponding folder and runs together with the official model and tools of the simulation platform as a module.

[0036] Furthermore, the extracted data in S4 includes:

[0037] The data generated by each sensor, the information of the detected target objects, the road logic information in the road network file of the simulation platform, and the position and attitude data of the simulated vehicle in the world coordinate system.

[0038] Furthermore, the process of saving data to the dynamic target object dataset path specified in the configuration file in S6 includes:

[0039] According to the frequency specified in the configuration file, the dynamic target tag information and timestamps are saved in chronological order from top to bottom to a tag file in *.txt format.

[0040] Furthermore, the saving process in S7 includes:

[0041] S71: Obtain static element information from the interface of static facility targets based on simulation software;

[0042] S72: Based on the static feature attribute file, extract the static feature information contained in the file and redefine the coordinates of the static features;

[0043] S73: Based on different feature categories, static feature information is processed into different *.json format files.

[0044] The beneficial effects of this invention are:

[0045] (1) The present invention constructs a simulation dataset through simulation software. The simulation software can generate various simulation scenarios such as different road conditions, weather conditions, traffic flow, etc., as well as situations that are difficult to encounter or too dangerous in the real world. This can not only reduce the cost and difficulty of real vehicle data collection, but also improve the accuracy of data labeling and ensure higher data labeling efficiency by combining hardware parameters, software dynamic response parameters and simulation platform operation parameters to correct the real-time position data of dynamic targets collected by the lidar sensor during the simulation process.

[0046] (2) The present invention calculates the bandwidth fluctuation value of the simulation platform from the first data extraction to the i-th data extraction. Since the bandwidth may fluctuate dynamically due to other traffic, the data acquisition interval is uneven. Uneven sampling will introduce false frequency components, resulting in the amplitude calculation being too high, which in turn leads to the acquired data being too high. Therefore, this data is introduced to correct the bandwidth size of the simulation platform at the i-th data extraction, which can improve the accuracy and reliability of the bandwidth size data, thereby providing accurate data support for subsequent analysis of the delay status during the operation of the simulation platform.

[0047] (3) This invention obtains the response delay risk coefficient of the simulation platform during the i-th data extraction. Compared with the preset response delay risk coefficient threshold By comparing the data, we can determine whether the dynamic target object calibration position needs to be corrected during the i-th data extraction, since the data is obtained based on diversified data coupling calculation.

[0048] (4) In this invention, the response delay risk coefficient during the i-th data extraction of the simulation platform is combined with the risk coefficient of the response delay. The data, obtained by combining a deep learning model with a large amount of historical data, can be used to determine the positional error distance of the a-th dynamic target object during the i-th data extraction on the simulation platform. Accurate calculations are performed to correct the real-time position data of dynamic targets collected by the lidar sensor, thereby ensuring the accuracy of the dynamic target labeling data. Attached Figure Description

[0049] The invention will now be further described with reference to the accompanying drawings.

[0050] Figure 1 This is a flowchart of a method for generating an autonomous driving simulation annotation dataset according to the present invention;

[0051] Figure 2This is a flowchart illustrating how the extracted static traffic sign information is stored in the static traffic sign dataset path specified in the configuration file, as described in this invention. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] Please see Figure 1 As shown, in one embodiment, this application provides a method for generating an autonomous driving simulation annotation dataset, the method comprising:

[0054] S1: Dynamically load the executable file of the dataset generation tool through the plug-in interface of the autonomous driving simulation platform, and import it through high-precision maps to build a complex road network model containing multiple lanes, traffic signs, and traffic lights.

[0055] S2: Based on the modeling data of autonomous driving scenarios, add multiple simulated sensors to the simulated vehicle, including cameras, LiDAR, and GPS;

[0056] S3: Complete the configuration of the *.json file for dataset parameter information to support dynamic response analysis, providing hardware parameters, software dynamic response parameters, and simulation platform operating parameters;

[0057] S4: Start the simulation and, during the simulation run, use the dataset generation tool to extract the required dynamic target and static traffic sign information from the data generated by each sensor, the detected target information, and the road network file of the simulation platform.

[0058] S5: By combining hardware parameters, software dynamic response parameters and simulation platform operating parameters, the real-time position data of dynamic targets collected by the lidar sensor during the simulation process is corrected.

[0059] S6: Combine the real-time location data of the corrected dynamic target objects and save the data to the dynamic target object dataset path specified in the configuration file at each time step;

[0060] S7: After the simulation is completed, save the extracted static traffic sign information in the static traffic sign dataset path specified in the configuration file;

[0061] Through the above technical solution, this example provides a method for generating an autonomous driving simulation annotation dataset. First, the executable file of the dataset generation tool is dynamically loaded through the plug-in interface of the autonomous driving simulation platform. A complex road network model containing multiple lanes, traffic signs, and traffic lights is constructed by importing high-precision maps. Based on the modeling data of the autonomous driving scenario, multiple simulation sensors, including cameras, LiDAR, and GPS, are added to the simulation vehicle. Then, the *.json file of the dataset parameter information is configured to support dynamic response analysis, providing hardware parameters, software dynamic response parameters, and simulation platform operating parameters. The simulation is then started. During the simulation, the dataset generation tool extracts the required dynamic target object and static traffic sign information from the data generated by each sensor, the detected target object information, and the simulation platform road network file. Subsequently, the real-time location data of the dynamic target objects collected by the LiDAR sensor during the simulation is corrected by combining the hardware parameters, software dynamic response parameters, and simulation platform operating parameters. The corrected real-time location data of the dynamic target objects is then saved to the dynamic target object dataset path specified in the configuration file at each time step. Finally, after the simulation ends, the extracted static traffic sign information is saved to the static traffic sign dataset path specified in the configuration file.

[0062] Through the above technical solution, this example constructs a simulation dataset using simulation software. The simulation software can generate various simulation scenarios with different road conditions, weather conditions, traffic flow, and other situations that are difficult to encounter or too dangerous in the real world. This not only reduces the cost and difficulty of collecting real vehicle data, but also improves the accuracy of data annotation and ensures higher data annotation efficiency by combining hardware parameters, software dynamic response parameters, and simulation platform operating parameters to correct the real-time position data of dynamic targets collected by the lidar sensor during the simulation.

[0063] The data provided in S3 includes:

[0064] Hardware parameters, software dynamic response parameters, and simulation platform operating parameters, wherein the hardware parameters include hardware latency, the software dynamic response parameters include software latency, and the simulation platform operating parameters include filtering algorithm processing time, bandwidth, and data transmission volume;

[0065] Through the above technical solution, this example provides the data provided in S3, including hardware parameters, software dynamic response parameters, and simulation platform operating parameters. Among them, hardware parameters include hardware latency, software dynamic response parameters include software latency, and simulation platform operating parameters include filtering algorithm processing time, bandwidth, and data transmission volume. Based on diversified and multi-dimensional data support, it is possible to analyze the latency status during the simulation platform operation process, thereby further determining whether there is lag in the data collected by the LiDAR, and ensuring the accuracy of the labeling data of dynamic targets.

[0066] The correction process in S5 includes:

[0067] Through formula Calculate the corrected bandwidth size for the i-th data extraction from the simulation platform. ;

[0068] Where i represents any data extraction from the simulation platform. This represents the total number of data extractions performed by the simulation platform during the i-th data extraction. Let be the bandwidth of the simulation platform during the i-th data extraction. For all The average value, This is a proportionality coefficient, set based on empirical fitting.

[0069] Through the above technical solution, this example provides the corrected bandwidth size for the i-th data extraction by the simulation platform. It can be done through the formula The calculation yields the result, where the formula is... The bandwidth fluctuation value of the simulation platform from the first data extraction to the i-th data extraction can be calculated. This data reflects the bandwidth fluctuation of the simulation platform from the first data extraction to the i-th data extraction. In this case, since the bandwidth may fluctuate dynamically due to other traffic, the data acquisition interval is uneven. Uneven sampling will introduce false frequency components, resulting in an overestimation of the amplitude and thus an overestimation of the acquired data. Therefore, introducing this data to correct the bandwidth value of the simulation platform at the i-th data extraction can improve the accuracy and reliability of the bandwidth value data, thereby providing accurate data support for subsequent analysis of the latency status during the operation of the simulation platform.

[0070] The correction process in S5 also includes:

[0071] Through formula Calculate the transmission delay during the i-th data extraction from the simulation platform. ;

[0072] in, This represents the amount of data transmitted during the i-th data extraction by the simulation platform.

[0073] Using the above technical solution, this example provides the transmission delay during the i-th data extraction in the simulation platform. It can be done through the formula The calculation yielded this formula, which is existing technology and will not be elaborated upon here.

[0074] The correction process in S5 also includes:

[0075] Through formula Calculate the response delay risk coefficient for the i-th data extraction from the simulation platform. ;

[0076] in, The hardware latency during the i-th data extraction by the simulation platform. The software delay during the i-th data extraction by the simulation platform. The preset software delay, To define a function, if Then let Otherwise, let , The above standard value can be selected and set based on the allowable error from empirical data, serving as the delay standard value. , and The weighting coefficients are set based on empirical fitting. The preset bandwidth size, This represents the processing time of the filtering algorithm during the i-th data extraction in the simulation platform. For all The average value;

[0077] Through the above technical solution, this example provides the response delay risk coefficient for the i-th data extraction of the simulation platform. It can be done through the formula The calculation yields the result, where the formula is... The dispersion coefficients of the filtering algorithm processing time from the first data extraction to the i-th data extraction can be calculated. Clearly, the higher the software latency, hardware latency, and transmission latency of the simulation platform during the i-th data extraction, the smaller the bandwidth of the simulation platform during the i-th data extraction. Furthermore, the larger the dispersion coefficients of the filtering algorithm processing time from the first data extraction to the i-th data extraction, the greater the response latency risk coefficient of the simulation platform during the i-th data extraction. The higher the latency, the greater the latency, indicating a delay in the data acquired by the lidar during that data extraction. Conversely, the lower the software latency, hardware latency, and transmission latency during the i-th data extraction, the larger the bandwidth of the simulation platform during that data extraction, and the smaller the dispersion coefficient of the filtering algorithm processing time from the first data extraction to the i-th data extraction, the lower the response latency risk coefficient of the simulation platform during the i-th data extraction. The lower the value, the less lag there is in the data collected by the lidar during that data extraction.

[0078] Specifically, the software latency, hardware latency, and transmission latency during the i-th data extraction of the simulation platform directly reflect the latency status of the simulation platform. The lower the bandwidth during the i-th data extraction, the less it can meet the requirements of high-precision simulation, indirectly causing data retransmission or compensation calculation, increasing latency. Finally, the larger the dispersion coefficient of the filtering algorithm processing time from the first data extraction to the i-th data extraction, the greater the latency in multi-target simulation scenarios. In such scenarios, the filtering algorithm needs to allocate computing resources to each target channel, leading to system latency. Therefore, based on diverse data, the latency status of the simulation platform during each data extraction can be analyzed, thereby providing accurate data support for subsequent correction of the real-time position data of dynamic targets collected by the lidar sensor.

[0079] The correction process in S5 also includes:

[0080] By obtaining the response delay risk coefficient of the simulation platform during the i-th data extraction. Compared with the preset response delay risk coefficient threshold Perform a comparison;

[0081] like It was determined that there was a serious response delay during the i-th data extraction of the simulation platform, and the dynamic target calibration position during the i-th data extraction needed to be corrected.

[0082] like It is determined that there is no response delay during the i-th data extraction of the simulation platform, and there is no need to correct the dynamic target object calibration position during the i-th data extraction;

[0083] Through the above technical solution, this example obtains the response delay risk coefficient of the simulation platform during the i-th data extraction. Compared with the preset response delay risk coefficient threshold By comparing the data, we can determine whether the dynamic target object calibration position needs to be corrected during the i-th data extraction, since the data is obtained based on diversified data coupling calculation.

[0084] The correction process in S5 also includes:

[0085] Through formula Calculate the position error distance of the a-th dynamic target object during the i-th data extraction on the simulation platform. ;

[0086] Where 'a' is any dynamic target object. The distance between the a-th dynamic target and the simulated vehicle during the i-th data extraction by the simulation platform. The adjustment coefficient lookup table function has a range of values ​​that are related to... The numerical values ​​correspond one-to-one. It should be noted that the values ​​of the adjustment coefficient lookup table function can be based on empirical data. The impact of the range of numerical values ​​on the position error was obtained based on testing.

[0087] Using the above technical solution, this example provides the positional error distance of the a-th dynamic target object during the i-th data extraction of the simulation platform. It can be done through the formula The calculated response delay risk coefficient is obtained by combining it with the data extraction time of the i-th simulation platform. The data, obtained by combining a deep learning model with a large amount of historical data, can be used to determine the positional error distance of the a-th dynamic target object during the i-th data extraction on the simulation platform. Accurate calculations are performed to correct the real-time position data of dynamic targets collected by the lidar sensor, thereby ensuring the accuracy of the dynamic target labeling data.

[0088] The dynamic loading process in S1 includes:

[0089] During the simulation process, the dataset generation tool uses the SDK provided by the simulation platform. The executable file of the dataset generation tool is installed in the corresponding folder and runs together with the official model and tools of the simulation platform as a module.

[0090] Through the above technical solution, this example provides the dynamic loading process in S1. During the simulation, the dataset generation tool uses the SDK provided by the simulation platform, installs the executable file of the dataset generation tool in the corresponding folder, and runs it as a module along with the official models and tools of the simulation platform. With this setting, the SDK provides standardized APIs and communication protocols, ensuring seamless integration between the dataset generation tool and the underlying architecture of the simulation platform. Furthermore, the official SDK is usually updated synchronously with the simulation platform version, avoiding crashes or data errors caused by toolchain incompatibility. As a module, it runs with the official models and tools of the simulation platform and can be dynamically loaded or unloaded like a plugin without modifying the core code of the simulation platform. The modularity also allows the dataset generation tool to be developed, tested, and deployed independently, reducing dependence on the main process of the simulation platform.

[0091] The extracted data in S4 includes:

[0092] The data generated by each sensor, the information of detected targets, the road logic information in the simulation platform's road network file, and the position and attitude data of the simulated vehicle in the world coordinate system;

[0093] Through the above technical solution, this example provides the extracted data in S4, including data generated by various sensors, information on detected targets, road logic information in the simulation platform's road network file, and position and attitude data of the simulated vehicle in the world coordinate system. These data cover various scenarios and conditions. By calibrating these data, a large amount of labeled data can be provided for training subsequent deep learning-based autonomous driving algorithms to ensure that the algorithms can cope with various complex situations in practical applications.

[0094] The process of saving data to the dynamic target object dataset path specified in the configuration file in S6 includes:

[0095] According to the frequency in the configuration file, the dynamic target tag information and timestamps are saved in chronological order from top to bottom to a tag file in *.txt format;

[0096] Through the above technical solution, this example provides a process for saving data to the dynamic target object dataset path specified in the configuration file. As an example, if the configuration file contains a LiDAR, a LiDAR folder is created. Based on the angle information contained in the LiDAR point cloud, the point cloud data of the rotating LiDAR is divided into single-frame point clouds in 360° circles. The acquired point cloud data is saved as a single-frame point cloud file in *.pcd format, named with a timestamp, and saved to the LiDAR folder.

[0097] If the configuration file includes cameras, create an image folder, and then create separate folders for different cameras within the image folder, based on the number of cameras included. Save the camera data as *.png format images according to the frequency specified in the configuration file, naming each image with a timestamp, and save them to the camera folder.

[0098] If the configuration file includes GPS sensors, then according to the frequency specified in the configuration file, the attitude and timestamp of the parking space are saved in chronological order from top to bottom to a *.txt file. Furthermore, the bounding box information of different categories of dynamic traffic participants, such as cars, buses, trucks, and pedestrians, is obtained from the dynamic object interface of the simulation software. This information includes the category, the coordinates of the bounding box center point, length, width, height, and orientation angle. Finally, according to the frequency specified in the configuration file, the dynamic object label information and timestamps are saved in chronological order from top to bottom to a *.txt label file. This setup provides a large amount of dynamic object annotation data for training subsequent deep learning-based autonomous driving algorithms, ensuring that the algorithms can handle various complex situations in practical applications.

[0099] Please see Figure 2 As shown, the saving process in S7 includes:

[0100] S71: Obtain static element information from the interface of static facility targets based on simulation software;

[0101] S72: Based on the static feature attribute file, extract the static feature information contained in the file and redefine the coordinates of the static features;

[0102] S73: Based on different feature categories, static feature information is processed into different *.json format files;

[0103] Through the above technical solution, this example provides the saving process in S7. First, based on the static facility target interface of the simulation software, static element information is obtained. Then, according to the static element attribute file, the static element information contained in the file is extracted, and the coordinates of the static elements are redefined. According to different element categories, the static element information is processed into different *.json format files. Through this setting, a large amount of static target annotation data can be provided for training of subsequent deep learning-based autonomous driving algorithms, so that the annotation covers various scenarios and conditions, ensuring that the algorithm can cope with various complex situations in practical applications.

[0104] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for generating an autonomous driving simulation annotation dataset, characterized in that, The method includes: S1: Dynamically load the executable file of the dataset generation tool through the plug-in interface of the autonomous driving simulation platform, and import it through high-precision maps to build a complex road network model containing multiple lanes, traffic signs, and traffic lights. S2: Based on the modeling data of autonomous driving scenarios, add multiple simulated sensors to the simulated vehicle, including cameras, LiDAR, and GPS; S3: Complete the configuration of the *.json file for dataset parameter information to support dynamic response analysis, providing hardware parameters, software dynamic response parameters, and simulation platform operating parameters; S4: Start the simulation and, during the simulation run, use the dataset generation tool to extract the required dynamic target and static traffic sign information from the data generated by each sensor, the detected target information, and the road network file of the simulation platform. S5: By combining hardware parameters, software dynamic response parameters and simulation platform operating parameters, the real-time position data of dynamic targets collected by the lidar sensor during the simulation process is corrected. S6: Combine the real-time location data of the corrected dynamic target objects and save the data to the dynamic target object dataset path specified in the configuration file at each time step; S7: After the simulation is completed, the extracted static traffic sign information will be saved in the static traffic sign dataset path specified in the configuration file.

2. The method for generating an autonomous driving simulation annotation dataset according to claim 1, characterized in that, The data provided in S3 includes: The hardware parameters, software dynamic response parameters, and simulation platform operating parameters include hardware latency, software dynamic response parameters including software latency, and simulation platform operating parameters including filtering algorithm processing time, bandwidth, and data transmission volume.

3. The method for generating an autonomous driving simulation annotation dataset according to claim 1, characterized in that, The correction process in S5 includes: Through formula Calculate the corrected bandwidth size for the i-th data extraction from the simulation platform. ; Where i represents any data extraction from the simulation platform. This represents the total number of data extractions performed by the simulation platform during the i-th data extraction. Let be the bandwidth of the simulation platform during the i-th data extraction. For all The average value, This is a proportionality coefficient, set based on empirical fitting.

4. The method for generating an autonomous driving simulation annotation dataset according to claim 3, characterized in that, The correction process in S5 also includes: Through formula Calculate the transmission delay during the i-th data extraction from the simulation platform. ; in, This represents the amount of data transmitted during the i-th data extraction by the simulation platform.

5. The method for generating an autonomous driving simulation annotation dataset according to claim 4, characterized in that, The correction process in S5 also includes: Through formula Calculate the response delay risk coefficient for the i-th data extraction from the simulation platform. ; in, The hardware latency during the i-th data extraction by the simulation platform. The software delay during the i-th data extraction by the simulation platform. The preset software delay, To define a function, if Then let Otherwise, let , This is the standard value for delay. , and The weighting coefficients are set based on empirical fitting. The preset bandwidth size, This represents the processing time of the filtering algorithm during the i-th data extraction in the simulation platform. For all The average value.

6. The method for generating an autonomous driving simulation annotation dataset according to claim 5, characterized in that, The correction process in S5 also includes: By obtaining the response delay risk coefficient of the simulation platform during the i-th data extraction. Compared with the preset response delay risk coefficient threshold Perform a comparison; like It was determined that there was a serious response delay during the i-th data extraction of the simulation platform, and the dynamic target calibration position during the i-th data extraction needed to be corrected. like It is determined that there is no response delay during the i-th data extraction of the simulation platform, and there is no need to correct the dynamic target calibration position during the i-th data extraction.

7. The method for generating an autonomous driving simulation annotation dataset according to claim 6, characterized in that, The correction process in S5 also includes: Through formula Calculate the position error distance of the a-th dynamic target object during the i-th data extraction on the simulation platform. ; Where 'a' is any dynamic target object. The distance between the a-th dynamic target and the simulated vehicle during the i-th data extraction by the simulation platform. The adjustment coefficient lookup table function has a range of values ​​that are related to... The numerical values ​​correspond one-to-one.

8. The method for generating an autonomous driving simulation annotation dataset according to claim 1, characterized in that, The dynamic loading process in S1 includes: During the simulation, the dataset generation tool uses the SDK provided by the simulation platform. The executable file of the dataset generation tool is installed in the corresponding folder and runs together with the official model and tools of the simulation platform as a module.

9. The method for generating an autonomous driving simulation annotation dataset according to claim 1, characterized in that, The extracted data in S4 includes: The data generated by each sensor, the information of the detected target objects, the road logic information in the road network file of the simulation platform, and the position and attitude data of the simulated vehicle in the world coordinate system.

10. The method for generating an autonomous driving simulation annotation dataset according to claim 1, characterized in that, The process of saving data to the dynamic target object dataset path specified in the configuration file in S6 includes: According to the frequency specified in the configuration file, the dynamic target tag information and timestamps are saved in chronological order from top to bottom to a tag file in *.txt format.

11. The method for generating an autonomous driving simulation annotation dataset according to claim 1, characterized in that, The saving process in S7 includes: S71: Obtain static element information from the interface of static facility targets based on simulation software; S72: Based on the static feature attribute file, extract the static feature information contained in the file and redefine the coordinates of the static features; S73: Based on different feature categories, static feature information is processed into different *.json format files.

Images