Automatic driving model training data set construction method and module and application thereof
By processing multi-source data in a unified data layer and filtering and matching according to task goals, the poor model performance caused by a single data source in the autonomous driving system is solved, and efficient training data set construction and model performance improvement are achieved.
Patent Information
- Application Number
- CN202411186991.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-27
- Publication Date
- 2024-11-29
AI Technical Summary
The training data sets of existing autonomous driving systems are usually single source and difficult to adapt to different environments and data formats, resulting in poor performance of models in different environments and scenarios and low training efficiency.
Through the unified data layer, heterogeneous and heterologous data are converted into standard training data in a unified format, and the training data set is formed based on the model's task objectives, making full use of heterologous data to quickly Extend the training dataset.
It significantly improves the training efficiency and effectiveness of the autonomous driving model, overcomes the problems of difficulty in integrating heterologous data, slow expansion of training data, and insufficient generalization capabilities of model, and provides an efficient training and performance optimization foundation for the autonomous driving model.
Smart Images

Figure CN119046686A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of autonomous driving, and specifically relates to a method and module for constructing an autonomous driving model training data set and its application. Background Art
[0002] With the rapid development of autonomous driving technology, autonomous driving systems rely on a large amount of high-quality data for model training. These data mainly come from a variety of sensors, map data, and other public driving data sets. In order to achieve high-precision prediction and decision-making of autonomous driving models, the diversity and scale of training data sets are crucial.
[0003] In existing technologies, the training data set for autonomous driving systems usually comes from a single source, mainly relying on data collected from specific brands or specific environments. Although these data can provide sufficient training samples in some scenarios, when the system needs to operate in different environments, different geographical locations, or face different types of obstacles, the performance of the model is often unsatisfactory.
[0004] Specifically, existing feature extraction architectures are often designed for a single data source and are difficult to adapt to data formats and structures from different sources. When faced with heterogeneous data, these architectures usually require complex data conversion and preprocessing, which is time-consuming and labor-intensive, and prone to introducing errors. Even after completing these complex processing steps, it is difficult to ensure the consistency and quality of the training data set generated in the end. In addition, the existing architecture lacks flexibility and cannot quickly adjust the data ratio according to different task objectives, resulting in inefficient model training and an inability to fully utilize the available heterogeneous data to quickly expand the training data set.
[0005] Therefore, it is necessary to provide a new solution to the above technical problems. Summary of the invention
[0006] The purpose of the present invention is to provide a method, module and application of constructing an autonomous driving model training data set, which can process multi-source native data, convert these heterogeneous and heterogeneous data into standard training data in a unified format, and then use the advantages of heterogeneous data to achieve rapid expansion of the training data set, thereby significantly improving the training efficiency and effect of the autonomous driving model.
[0007] To achieve the above purpose, the technical solution provided by the present invention is as follows:
[0008] In a first aspect, the present invention provides a method for constructing a training data set for an autonomous driving model, comprising: acquiring native data from multiple data sources; processing the native data, wherein the processing includes data labeling and format conversion, so as to convert the native data into standard training data in a unified format to form a standard training database; based on the task objectives of the model, screening standard training data that meets the task objectives from the standard training database to form a training data set.
[0009] In one or more embodiments, the native data includes data acquired in real time by the unmanned vehicle and / or data from a public data set; the data acquired in real time by the unmanned vehicle includes: scene data corresponding to when the model prediction effect is lower than expected during the operation of the unmanned vehicle and scene data encountered for the first time by the unmanned vehicle.
[0010] In one or more embodiments, the native data includes perception data, positioning data and map data; processing the native data includes: aligning the time frames of the perception data and the positioning data to keep the perception data and the positioning data consistent on the time axis.
[0011] In one or more embodiments, the time frames of the perception data and the positioning data are aligned, including: taking the time frame of the perception data as a reference, querying the positioning data time frame closest to the perception data time frame; if the time interval between the closest positioning data time frame and the perception data time frame is less than a preset threshold, retaining the positioning data corresponding to the positioning data time frame, otherwise discarding the positioning data corresponding to the positioning data time frame.
[0012] In one or more embodiments, standard training data that meets the task objectives is screened from the standard training database to form a training data set, including: based on the task objectives of the model, data matching is performed on the screened standard training data, and different weights are assigned according to the labeled categories of the data to form a training data set.
[0013] In one or more embodiments, standard training data that meets the task objectives is selected from the standard training database, including: filtering static target obstacles and low-speed moving target obstacle data, wherein the low-speed moving target obstacles include target obstacles with a speed less than 0.01 m / s and target obstacles that move a distance less than 2 m in the next 8 seconds; calculating the shortest number of historical frames required to generate an accurate trajectory based on the speed of the target obstacle, and filtering out target obstacle data with a number of historical frames less than a preset minimum value; and filtering out target obstacle data with an angle greater than 30° among motor vehicle target obstacles and target obstacle data with an angle greater than 90° among bicycle target obstacles based on the angle between the speed direction and the displacement direction of the target obstacle.
[0014] In one or more embodiments, standard training data that meets the task objectives is selected from the standard training database to form a training data set, including: segmenting the lane centerline and lane boundary class data according to a preset interval parameter, performing linear interpolation on each segmented line segment, and sorting the segmented sub-map elements according to the weighted distance between the segmented sub-map elements and the target obstacle; adding points at the end of the coordinate sequence of the crosswalk to form a complete ring, segmenting the ring elements twice, and performing linear interpolation on each segmented line segment.
[0015] In one or more embodiments, the calculation formula for the weighted distance between the sub-map element and the target obstacle includes:
[0016] target_to_other_angle=arctan2(other_y-target_y,other_x-target_x);
[0017] other_to_target_angle=arctan2(target_y-other_y,target_x-other_x);
[0018] target_velocity_angle=arctan2(target_vy,target_vx);
[0019] other_to_target_angle=arctan2(other_vy,other_vx);
[0020] relative_speed_target=(target_speed*(π-|normalize_angle(target_to_other_angle-target_velocity_angle)|)) / π;
[0021] relative_speed_other=(other_speed*(π-|normalize_angle(other_velocity_angle-other_to_target_angle)|)) / π;
[0022] weight_distance=normal_distance / (c_coef+relative_speed_other+relative_speed_target);
[0023] Among them, target_to_other_angle represents the direction angle from the target obstacle to the submap element, [target_x, target_y] represents the coordinates of the target obstacle at the current moment, [other_x, other_y] represents the coordinates of a point in the submap element, [target_vx, target_vy] represents the velocity vector of the target obstacle at the current moment, [other_vx, other_vy] represents the velocity vector of a point in the submap element, other_to_target_angle represents the direction angle from the submap element to the target obstacle, relative_speed_target represents the speed of the target obstacle relative to the submap element, target_speed represents the speed of the target obstacle, target_velocity_angle represents the speed direction of the target obstacle, relative_speed_other represents the speed of the submap element relative to the target obstacle, other_speed represents the speed of the submap element, other_velocity_angle represents the speed direction of the submap element, weight_distance represents the weighted distance between the submap element and the target obstacle, normal_distance represents the Euclidean distance between the submap element and the target obstacle, and c_coef is a constant.
[0024] In one or more embodiments, standard training data that meets the task objectives is selected from the standard training database to form a training data set, including: setting a future frame feature sequence with a preset number of frames, the first frame of the sequence is the frame at time t+1, and the current frame of the target obstacle is the frame at time t; rotating the features of the target obstacle so that the feature direction is aligned with the movement direction of the obstacle, and converting the global coordinates of the target obstacle into local coordinates; and performing linear interpolation on the state of the future frame to reduce the prediction uncertainty caused by time fluctuation.
[0025] In the second aspect, the present invention provides an autonomous driving model training data set construction module, which includes: a unified data layer, used to obtain native data from multiple data sources and process the native data, the processing including data labeling and format conversion, so as to convert the native data into standard training data in a unified format to form a standard training database; a feature data layer, used to screen standard training data that meets the task objectives from the standard training database based on the model's task objectives to form a training data set.
[0026] In a third aspect, the present invention provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method for constructing an autonomous driving model training data set as described above is implemented.
[0027] In a fourth aspect, the present invention provides a computer-readable medium, which carries computer execution instructions. When the computer execution instructions are executed by a processor, they are used to implement the method for constructing an autonomous driving model training data set as described above.
[0028] Compared with the prior art, the method, module and application of constructing an autonomous driving model training data set provided by the present invention convert heterogeneous and heterogeneous data into standard training data in a unified format by processing multi-source native data, and then construct a standard training database. According to the specific task objectives of the model to be trained, the data that best meets the task requirements is screened out from the standard training database, and a training data set is formed through a reasonable data ratio. The advantages of heterogeneous data are fully utilized to achieve rapid expansion of the training data set, thereby significantly improving the training efficiency and effect of the autonomous driving model. BRIEF DESCRIPTION OF THE DRAWINGS
[0029] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.
[0030] Figure 1 A schematic diagram of an application scenario of a method for constructing a training data set in one embodiment of the present invention;
[0031] Figure 2 A flowchart of a method for constructing a training data set in one embodiment of the present invention;
[0032] Figure 3 A structural block diagram of a training data set construction module in one embodiment of the present invention;
[0033] Figure 4 FIG. 4 is a structural block diagram of an electronic device in an embodiment of the present invention. DETAILED DESCRIPTION
[0034] In order to enable those skilled in the art to better understand the technical solutions in the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.
[0035] With the rapid development of autonomous driving technology, the quality and diversity of model training data have become one of the key factors affecting the performance of autonomous driving systems. In existing technologies, the training data of autonomous driving models mainly relies on data sets from a single or limited source. Although this method can work effectively in specific scenarios, its generalization ability and adaptability are often insufficient when faced with complex and changeable actual road environments. Specifically, the construction of data sets in existing technologies is often limited to data collected from a single brand or a single environment, resulting in poor performance of the model when encountering data from different sources or different environments, and the inability to fully utilize data resources from multiple sources.
[0036] The inventors further analyzed and found that the existing feature extraction architecture design also has certain limitations. It is difficult to effectively integrate heterogeneous data. Especially when processing heterogeneous data, it is inefficient and easy to introduce errors, which ultimately makes the quality and scale of the training data set difficult to meet the efficient training requirements of the autonomous driving model.
[0037] In response to these problems, the present invention proposes an innovative method for constructing a training dataset for an autonomous driving model. The core idea is to process multi-source native data through a unified data layer, convert these heterogeneous and heterogeneous data into standard training data in a unified format, and then construct a standard training database. This database not only covers diverse data from different sources, but also provides a high-quality, unified-format data foundation for subsequent model training.
[0038] On this basis, the feature data layer selects the data that best meets the task requirements from the standard training database according to the specific task objectives of the model to be trained, and forms a training data set through a reasonable data ratio. In this process, the system can not only flexibly adapt to the requirements of different task objectives, but also make full use of the advantages of heterogeneous data to quickly expand the training data set, thereby significantly improving the training efficiency and effect of the autonomous driving model.
[0039] The technical implementation idea of the present invention is to overcome the problems existing in the prior art such as the difficulty in integrating heterogeneous data, slow expansion of training data, and insufficient model generalization ability through standardized data processing and intelligent data screening, thereby providing a basis for efficient training and performance optimization of autonomous driving models.
[0040] Please refer to Figure 1 FIG. 1 is a schematic diagram of an application scenario of a method for constructing a training data set for an autonomous driving model in an embodiment of the present invention. Figure 1 In the application scenario shown, an autonomous driving model training dataset construction module 101 that integrates the autonomous driving model training dataset construction method of the present invention, a cloud server 102, multiple unmanned vehicles of different brands 103 and a public dataset platform 104 are included. The autonomous driving model training dataset construction module 101 includes a unified data layer (Unified layer) 101A and a feature data layer (Feature layer) 101B.
[0041] The native data of the unmanned vehicle 103 and the public data set platform 104 can be uploaded to the cloud server 102. The autonomous driving model training data set construction module 101 can download the native data in the cloud server 102, and process the downloaded native data through the unified data layer 101A, converting the native data into standard training data in a unified format to form a standard training database. The feature data layer 101B can filter the standard training data that meets the task objectives from the standard training database based on the model's task objectives to form a training data set.
[0042] Specifically, the autonomous driving model training data set construction module 101 is the core part of the entire system. The module 101 interacts with external data sources through the cloud server 102, including unmanned vehicles 103 of different brands and a public data set platform 104. The unmanned vehicle 103 can travel on actual roads and collect environmental data around the vehicle in real time through various sensors (such as lidar, camera, radar, etc.). These data cover road scenes, obstacle information, vehicle status, traffic signs, etc. In addition, the public data set platform 104 provides data from other research institutions, enterprises or public projects, which can come from simulated environments, unmanned vehicles of other brands or historical data sets.
[0043] All these raw data are first uploaded to the cloud server 102 through the unmanned vehicle 103 and the public dataset platform 104. At this stage, the data can be heterogeneous, and they can have different formats, different resolutions, and different sampling frequencies. In order to resolve the differences of these heterogeneous data and integrate them into a unified format, the unified data layer 101A in the autonomous driving model training dataset construction module 101 processes these raw data.
[0044] The primary task of the unified data layer 101A is to process the native data downloaded from the cloud server 102. The processing process includes two main aspects: data annotation and format conversion. Data annotation refers to the detailed labeling of each piece of native data. The labels may include the source of the data (for example, from a certain brand of unmanned vehicle or a public data set), the data collection scene (such as urban roads, highways, rainy days or sunny days, etc.), the timestamp of the data, and other metadata related to training. These annotations provide key information about the content and quality of the data, which can provide a basis for the subsequent screening and use of data.
[0045] The unified data layer 101A converts the format of the labeled data. Since there may be many differences in the format of native data, such as different data formats of different sensors or inconsistent resolutions of different data sets, the unified data layer 101A needs to convert these data into a standard format within the system. This process ensures that all data is uniformly processed before entering the standard training database, eliminating structural differences between data, thereby improving data compatibility and utilization efficiency.
[0046] The processed and converted data is stored in the standard training database. This database is the core resource library for autonomous driving model training, covering high-quality data from different sources that have been standardized. Since all data has been labeled and formatted, the standard training database not only provides a unified data structure, but also provides detailed metadata, making subsequent data screening and use more efficient and accurate.
[0047] The feature data layer 101B is a key module for extracting and constructing a specific training data set from a standard training database. The feature data layer 101B can filter out training data that meets the requirements from the standard training database according to the task objectives of the model to be trained.
[0048] For example, if the goal of the model is to improve the accuracy of obstacle detection, the feature data layer 101B will give priority to high-quality data that contains rich obstacle information, and this data may come from a variety of scenarios and environments, such as city streets, highways, complex intersections, etc. In order to further optimize the training effect, the feature data layer 101B can also match the screened data and give different types of data different weights. For example, data in certain scenarios may be more scarce and important, and the system can increase the proportion of this data in the training set to ensure that the model performs more reliably in key scenarios.
[0049] When the training data set is built, the system uses it in the training process of the autonomous driving model. During the training process, the model gradually improves its adaptability to different driving scenarios through continuous learning and optimization. Since the data set covers driving data from a variety of different environments and sources, the trained model can better cope with complex situations on the actual road and has stronger generalization ability and stability.
[0050] Please refer to Figure 2 As shown, a flowchart of a method for constructing an autonomous driving model training data set in one embodiment of the present invention is shown. The method for constructing an autonomous driving model training data set specifically comprises the following steps:
[0051] S201: Acquire native data from multiple data sources.
[0052] It should be noted that when implementing the step S201, these native data can be obtained in a variety of ways. Unmanned vehicles themselves are one of the important data sources. These unmanned vehicles are usually equipped with a variety of sensors, such as cameras, laser radars, millimeter wave radars, etc., which can collect information about the surrounding environment in real time while the vehicle is driving.
[0053] For example, cameras can capture images and videos of the road, lidar can generate high-precision 3D point cloud data, and millimeter-wave radar can detect the distance and speed of obstacles ahead. These sensors work together to generate rich raw data covering different driving scenarios and conditions, such as sunny, rainy, daytime, night, and driving in urban roads, highways, rural roads, etc.
[0054] Public dataset platforms are another important data source. Many research institutions and companies will publicly release the driving datasets they collect or simulate to promote the development of autonomous driving technology. For example, autonomous driving datasets such as Waymo, NuScenes, and KITTI have been widely used in academic research and commercial development. These public datasets usually include driving data from different regions and different periods, providing great data diversity. By combining these public datasets with actual collected data, the data source can be further enriched to ensure the diversity and comprehensiveness of the training dataset.
[0055] In step S101, native data can be downloaded from multiple data sources (such as unmanned vehicles and public dataset platforms) through specific network protocols and data transmission mechanisms. These data can be stored on a cloud server.
[0056] Native data of various types and formats can be obtained from multiple data sources, including different sensor data, image data, location information, and other environmental data. The purpose of obtaining multi-source data is to ensure that the system can cover the widest possible range of driving scenarios and conditions, thereby providing rich samples for subsequent autonomous driving model training. If you only rely on data from a single source, the trained model may show certain limitations in practical applications and it will be difficult to cope with driving needs in different environments. By obtaining data from multiple data sources, the system can cover more driving scenarios, such as different weather conditions, different time periods, different road types, etc., which will help improve the generalization and robustness of the model.
[0057] S202: Processing the native data, wherein the processing includes data labeling and format conversion, so as to convert the native data into standard training data in a unified format to form a standard training database.
[0058] It should be noted that in step S202, the native data obtained from multiple data sources are standardized, including data annotation and format conversion. The main purpose of these processing steps is to convert the native data from various sources and in different formats into standard training data in a unified format, and store these standardized data in a standard training database to provide a basis for data consistency for subsequent model training.
[0059] Since native data may come from different sensors, vehicle brands or public data sets, their content and structure may vary greatly. Therefore, in order to effectively identify and utilize these data in subsequent processing, each piece of data needs to be labeled. The content of data annotation can include the source of the data (such as a driverless car from a certain brand or a public data set), the environment and scene of data collection (such as urban roads, rural roads, highways, rainy days, sunny days, etc.), timestamps, geographic location information, and other training-related metadata.
[0060] These annotations provide detailed background information for the data, allowing the system to quickly locate the most appropriate data set according to task requirements when filtering and using data. For example, annotations can help the system quickly filter out relevant data when training an obstacle detection model for rainy scenes without having to traverse the entire database.
[0061] After the annotation is completed, the format of the native data is converted. Since data from different sources may have inconsistent formats, such as image data may come from cameras with different resolutions, or the data format of the lidar may be different from that of the millimeter-wave radar, these differences will bring difficulties to subsequent data processing. In order to eliminate these differences, all native data can be unified into a standard format through format conversion. The specific conversion process includes scaling images of different resolutions to the same resolution, converting different types of sensor data into a unified structured format, or time-aligning data of different time frequencies. Through these conversions, all data are ensured to have a consistent format and structure before entering the standard training database, which greatly simplifies the subsequent data processing and analysis process.
[0062] Data labeling and format conversion can convert raw data into standardized training data, which is stored in a standard training database. The establishment of a standard training database provides a standardized basic data pool for autonomous driving model training. This database includes labeled data with a unified format and structure to ensure that the required information can be read and processed efficiently and accurately when using this data for model training. The standard training database can effectively eliminate the differences between data from different sources, allowing data from different brands of driverless cars and public data sets to be used in a unified manner, thereby enriching the diversity and coverage of training data.
[0063] In a specific example, suppose the system needs to process image data and lidar data from multiple unmanned vehicles. These data may come from different vehicle brands, with different camera resolutions and different lidar data formats. In step S202, the system first adds annotation information to each data point, such as which brand of vehicle collected the data, the time and location of data collection, current weather conditions, etc. Next, the system will uniformly scale image data of different resolutions to a standard resolution, and convert the lidar data format into a universal structured format. Through these processes, the system converts data of different formats into consistent standard training data and stores it in a standard training database.
[0064] In an exemplary embodiment, the native data includes perception data, positioning data and map data, and processing the native data includes: aligning the time frames of the perception data and the positioning data so that the perception data and the positioning data maintain consistency on the time axis.
[0065] Perception data is usually collected by sensors on the vehicle (such as cameras, radars, lidars, etc.), providing detailed information about the vehicle's surroundings; positioning data is collected through GPS, IMU (inertial measurement unit) and other devices, providing the vehicle's precise position and motion status in the global coordinate system; map data provides detailed road information, such as lane lines, traffic signs, traffic light locations, etc. Since the frame rate of positioning data is generally higher than that of perception data, in order to ensure that these data can be used consistently, the time frames of perception data and positioning data need to be aligned.
[0066] The process of time frame alignment aims to synchronize perception data and localization data to the same time reference. Perception data and localization data usually have different sampling frequencies. For example, a camera may capture images at a frequency of 10 frames per second, while the GPS may update the vehicle's position 30 times per second. Since the two types of data have different sampling frequencies, if time frame alignment is not performed, their correspondence on the time axis will be deviated, which may lead to wrong decisions when using these data. For example, without time alignment, the system may mistakenly match the image captured by the camera with the wrong GPS data, resulting in a misjudgment of the actual location of the vehicle.
[0067] In order to achieve time frame alignment, the following specific implementation method can be adopted: when processing data, the data with high sampling frequency is downsampled based on the lower sampling frequency. For example, assuming that the camera captures 5 frames per second and the GPS updates the location data 10 times per second, the system can reduce the 10 frames per second location data to 5 frames per second through interpolation or interpolation technology, and ensure that each frame image accurately corresponds to a timestamp in the GPS data. In this way, the perception data and positioning data can be accurately matched to the same time point, ensuring that the data used at the same time point corresponds to each other.
[0068] Furthermore, a more accurate time alignment method can be used, that is, the time frame of the sensing data is used as a reference to perform time matching operations. When processing data, the timestamp of the sensing data is given priority, and the timestamp of the positioning data is compared with it.
[0069] Then, query the time frame of the positioning data closest to the time frame of the perception data. Positioning data is usually provided by devices such as GPS and IMU, and the frame rate of these data may be higher than that of the perception data. For example, GPS may update location data 10 times per second, while the camera may capture 5 frames per second. In order to find the positioning data closest to the perception data on the timeline, it is necessary to compare the timestamps of each frame of perception data and positioning data to find the frame with the smallest time difference.
[0070] After finding the positioning data time frame closest to the perception data time frame, the time interval between the two time frames will be checked. If the time interval is less than a preset threshold (for example, 0.12 seconds), the system will retain the positioning data corresponding to the positioning data time frame and pair it with the perception data. This means that the system believes that the two time frames are close enough to be considered as data occurring at the same point in time. If the time interval is greater than the preset threshold, the system will discard the data corresponding to the positioning data time frame to avoid using inaccurate positioning data.
[0071] S203: Based on the task objective of the model, standard training data that meets the task objective is selected from the standard training database to form a training data set.
[0072] It is understandable that in the standard training database, standardized data that has been processed, annotated and format converted has been stored. These data come from a wide range of sources, covering different environments, scenarios, sensor types, etc., and therefore have great diversity and wide applicability. However, different autonomous driving tasks have different data requirements. For example, the obstacle detection model needs to focus on relevant data including obstacle location, size, motion trajectory, etc., while the path planning model may pay more attention to road boundaries, traffic signs and road conditions. Therefore, in step S203, it is necessary to filter out the most relevant training data from the database according to the specific model task objectives.
[0073] Mission objectives can be specified by developers or analyzed by automated systems based on training requirements. The system will extract data from a standard training database based on the requirements of the mission objectives using predefined filters. These filters can include various metadata in the data annotation, such as scene type, weather conditions, sensor type, time period, etc. For example, when training an obstacle detection model, the system may select data that is annotated with a lot of dynamic obstacle information, while when training a night driving model, the system may prefer night driving data sets.
[0074] By selecting data in a targeted manner, we can ensure that the sample data received by the model during training is highly consistent with the actual application scenario. This consistency not only improves the generalization ability of the model, but also effectively reduces training time and resource consumption.
[0075] In an exemplary embodiment, standard training data that meets the task objectives is screened from the standard training database to form a training data set, including: matching the screened standard training data based on the task objectives of the model, assigning different weights according to the labeled categories of the data, and forming a training data set.
[0076] In order to form an efficient and targeted training data set, it is necessary to screen out standard training data that meets the task objectives from the standard training database and process these data in a proportion. This process includes classifying the screened data and assigning different weights according to the labeled categories of the data to ensure that the final training data set can meet the training needs of the model to the greatest extent.
[0077] For example, if the goal is to train an autonomous driving model that can drive in adverse weather conditions, data collected in rainy, snowy, or foggy conditions will be preferred. These data contain specific environmental features, such as slippery roads, obstructed vision, and changing traffic signals, which help the model learn how to make correct driving decisions in adverse conditions.
[0078] After screening out the standard training data that meets the task objectives, the system will further process these data, namely data matching. The core of data matching is to assign different weights according to the labeled categories of the data to reflect the importance of these data in model training. The labeled categories may include data collection scenarios (such as urban roads, highways, rural roads), weather conditions (such as sunny days, rainy days, foggy days), time periods (such as daytime, nighttime), and specific event types (such as emergency braking, obstacle detection, lane change operations, etc.). By assigning weights to these categories, the system can adjust the proportion of different data in the training set, thereby ensuring that the model can focus on learning those contents that are most critical to the task objectives.
[0079] Suppose you want to train an autonomous driving model specifically for complex urban traffic scenarios. Urban traffic is characterized by dense vehicles, a large number of pedestrians, and numerous traffic lights. Therefore, the system needs to focus on these features in the training data set. Specifically, the system can select a large amount of data labeled with the "urban road" category and give a higher weight to the "peak hour" data, because during peak hours, traffic volume is large, vehicle speeds are slow, and there are many emergencies, which are all the key learning content for the model. At the same time, the system may reduce the weight of some "highway" or "rural road" data, because although these data help the generalization of the model, they are less relevant to the current task objectives.
[0080] Through this data allocation method, the training data set will be more in line with the actual needs of the model. Data with different weights occupy different proportions in the training set, so that the model can more effectively focus on learning the features that are most critical to the task objectives during the training process, so that the model can not only adapt to specific scenarios more quickly, but also significantly improve its performance in these scenarios.
[0081] In an exemplary embodiment, standard training data that meets the task objectives is selected from the standard training database, including: filtering static target obstacles and slow-moving target obstacle data, wherein the slow-moving target obstacles include target obstacles with a speed less than 0.01 m / s and target obstacles that move a distance less than 2 m in the next 8 seconds; calculating the shortest number of historical frames required to generate an accurate trajectory based on the speed of the target obstacle, and filtering out target obstacle data with a number of historical frames less than a preset minimum value; filtering out target obstacle data with an angle greater than 30° among motor vehicle target obstacles and target obstacle data with an angle greater than 90° among bicycle target obstacles based on the angle between the speed direction and the displacement direction of the target obstacle.
[0082] In order to ensure that the training data set can effectively improve the performance of the model, the present invention proposes a series of refined screening criteria, mainly for filtering and selecting the data of target obstacles. This process includes filtering static or slow-moving target obstacles, filtering based on speed calculation history frames, and filtering according to the angle between the speed direction and the displacement direction. There is no filtering restriction for pedestrian obstacles because the movement trajectory of pedestrians may have a large degree of randomness. The purpose of these steps is to ensure that the target obstacle data in the training data set is the most valuable for model training, thereby improving the accuracy and reliability of the model in practical applications.
[0083] Specifically, first, filter out static target obstacles and slow-moving target obstacle data. Static target obstacles usually refer to objects whose relative positions remain basically unchanged, such as vehicles and buildings parked on the roadside. For autonomous driving models, although these static objects also need to be perceived, their contribution to the training of trajectory prediction models is limited because their trajectories are fixed. Therefore, when filtering data, the system will give priority to filtering out these static targets. Slow-moving target obstacles include objects with a speed of less than 0.01m / s and objects that move less than 2 meters in the next 8 seconds. The motion characteristics of this type of target obstacle are weak, and the trajectory changes are not obvious. They have little effect on the dynamic response capability of the training prediction model, so they will also be filtered out.
[0084] Then, based on the speed of the target obstacle, the shortest number of historical frames required to generate an accurate trajectory is calculated, and the target obstacle data with less historical frames than the preset minimum value is filtered out. The number of historical frames determines how much historical data the model can use to infer future motion trajectories. The faster the speed of the target obstacle, the more historical data is usually required to accurately predict its future position. Therefore, it is necessary to calculate the shortest number of historical frames required based on the speed of each target to ensure that the filtered data has sufficient time series information. If the number of historical frames of a target obstacle is less than the preset minimum value (for example, only 2 frames), it is considered that its historical data is insufficient to support accurate trajectory prediction, and it is filtered out.
[0085] Finally, the data is further screened based on the angle between the speed direction and the displacement direction of the target obstacle. The angle refers to the difference between the direction of motion of the target obstacle and its actual displacement direction. For autonomous driving models, target obstacles with smaller angles (such as motor vehicle obstacles with angles less than 30°) usually have more stable and predictable motion trajectories, while targets with larger angles may move irregularly and are difficult to accurately predict. Therefore, the system will filter out data with angles greater than 30° in motor vehicle target obstacles to ensure that the training data set mainly contains target obstacles with relatively stable motion trajectories. Similarly, for bicycle target obstacles, the system will filter out target data with angles greater than 90°, because the movement directions of these bicycles may be extremely unstable, making it difficult to provide meaningful trajectory information for the model.
[0086] In an exemplary embodiment, standard training data that meets the task objectives is selected from the standard training database to form a training data set, including: segmenting the lane centerline and lane boundary class data according to a preset interval parameter, performing linear interpolation on each segmented line segment, and sorting the segmented sub-map elements according to the weighted distance between the segmented sub-map elements and the target obstacle; adding points at the end of the coordinate sequence of the crosswalk to form a complete ring, segmenting the ring elements twice, and performing linear interpolation on each segmented line segment.
[0087] It should be noted that in the autonomous driving system, map data such as lane center lines, lane boundaries, and crosswalks are important bases for path planning and driving decisions. These data determine key tasks such as how the vehicle stays in the lane, how to avoid obstacles, and how to pass through crosswalks. Therefore, when constructing an autonomous driving model training data set, in order to ensure the accuracy and consistency of the training data, the present invention proposes a series of specific processing steps, including segmentation and interpolation of lane center line and lane boundary data, and point filling, segmentation, and interpolation of crosswalk data.
[0088] Lane centerlines and lane boundaries are important reference lines used by autonomous driving systems for path planning, and they are usually extracted directly from map data. However, these line segments in the original map data may be too long or too short due to acquisition accuracy or map resolution, resulting in an inability to accurately describe the true shape of the road in practical applications. In order to overcome this problem, the present invention first divides these line segments according to preset interval parameters. Specifically, the original long line segment is divided into several short sub-line segments according to a preset distance (for example, every 64 meters). The purpose of this is to decompose the long line segment into shorter, easier-to-process segments, thereby improving the accuracy of the map data.
[0089] Then, linear interpolation is performed on each segmented line segment. Linear interpolation is a commonly used mathematical method used to generate more points between known discrete points to make the data smoother and more continuous. In this process, several new points can be inserted on each segmented sub-segment (for example, 16 points are interpolated per segment). These interpolation points can further refine and refine the shape of the line segment so that the lane centerline and lane boundary can more accurately reflect the actual shape of the road. This processing step ensures the continuity and consistency of the map data, especially in complex sections such as bends and forks, where the interpolated line segments can provide more accurate path information.
[0090] In order to make the relationship between the segmented sub-map elements and the target obstacles clearer, they can be sorted according to the weighted distance between them. Weighted distance refers to a comprehensive distance measurement calculated after considering factors such as the direction and speed of the obstacle's movement. This measurement method can better reflect the actual relationship between obstacles and map elements than the simple Euclidean distance. For example, the system will prioritize the centerline and boundary of the lane that is closer to the obstacle, because these parts have a greater impact on the model's obstacle avoidance and path planning.
[0091] Since the crosswalk is a closed-loop map element, when processing the crosswalk data, points can be added at the end of the crosswalk coordinate sequence to form a complete ring. A crosswalk is usually a closed figure, but in the original map data, due to collection errors or incomplete data, the starting and ending points of the crosswalk may not be closed. Therefore, the system needs to add one or more points at the end of the coordinate sequence so that the sequence is connected from beginning to end to form a complete ring. This step ensures the integrity of the crosswalk data and helps the system accurately identify and process these key sections during training.
[0092] Then, the crosswalk is divided twice, and linear interpolation is performed on each segment after the division. The purpose of the second division is to further refine the ring of the crosswalk, make the length of each segment more uniform, and avoid accuracy problems caused by some segments being too long. Linear interpolation is used to generate more interpolation points on these segmented segments to further improve the precision of the data. For example, on a curved crosswalk, the interpolated data can better describe its actual shape, thus providing a more accurate reference for model training.
[0093] By segmenting, interpolating, and filling in lane centerlines, lane boundaries, and crosswalk data, the system can significantly improve the accuracy and consistency of these key map elements. This process not only makes the generated training dataset more realistic and accurate, but also provides more reliable input data for the autonomous driving model, thereby improving the model's performance in actual road environments.
[0094] Specifically, the calculation formula for the weighted distance between the sub-map element and the target obstacle includes:
[0095] target_to_other_angle=arctan2(other_y-target_y,other_x-target_x);
[0096] other_to_target_angle=arctan2(target_y-other_y,target_x-other_x);
[0097] target_velocity_angle=arctan2(target_vy,target_vx);
[0098] other_to_target_angle=arctan2(other_vy,other_vx);
[0099] relative_speed_target=(target_speed*(π-|normalize_angle(target_to_other_angle-target_velocity_angle)|)) / π;
[0100] relative_speed_other=(other_speed*(π-|normalize_angle(other_velocity_angle-other_to_target_angle)|)) / π;
[0101] weight_distance=normal_distance / (c_coef+relative_speed_other+relative_speed_target);
[0102] Among them, target_to_other_angle represents the direction angle from the target obstacle to the submap element, [target_x, target_y] represents the coordinates of the target obstacle at the current moment, [other_x, other_y] represents the coordinates of a point in the submap element, [target_vx, target_vy] represents the velocity vector of the target obstacle at the current moment, [other_vx, other_vy] represents the velocity vector of a point in the submap element, other_to_target_angle represents the direction angle from the submap element to the target obstacle, relative_speed_target represents the speed of the target obstacle relative to the submap element, target_speed represents the speed of the target obstacle, target_velocity_angle represents the speed direction of the target obstacle, relative_speed_other represents the speed of the submap element relative to the target obstacle, other_speed represents the speed of the submap element, other_velocity_angle represents the speed direction of the submap element, weight_distance represents the weighted distance between the submap element and the target obstacle, normal_distance represents the Euclidean distance between the submap element and the target obstacle, and c_coef is a constant. The arctan2 function is an inverse tangent function that returns the angle between two points, usually expressed in radians.
[0103] In the calculation formula of relative_speed_target, the angle between the speed direction of the target obstacle and the direction of the target obstacle relative to the sub-map element is first calculated, and the normalize_angle function is used to normalize the angle (usually the angle is limited to the range of [-π,π]). Then, this angle is subtracted from π to ensure that the closer to the direction of movement, the greater the weight. Finally, this weight is multiplied by the speed of the target obstacle and divided by π for normalization. The main function of this formula is to quantify the speed at which the target obstacle approaches the sub-map element. The larger the relative speed, the faster the obstacle approaches the sub-map element, which may increase the risk of collision, so it should have a greater impact in the weighted distance.
[0104] The same is true for relative_speed_other, which is used to evaluate the impact of the movement of sub-map elements (such as lane boundaries) on target obstacles. In practical applications, map elements are usually stationary, but in dynamic map scenarios or when the sub-map elements themselves are assumed to have motion characteristics, this formula provides a means to consider the impact of motion on path planning.
[0105] In the calculation formula of weight_distance, normal_distance is the Euclidean distance between the target obstacle and the sub-map element, c_coef is a constant used to prevent the denominator from being zero and to adjust the weighting effect. The relative_speed_other and relative_speed_target in the denominator are the weights of the relative speeds of the target obstacle and the sub-map element.
[0106] By dividing the Euclidean distance by the sum of the weight and constant of the relative speed, the calculated weighted distance can reflect the "actual" distance at different relative speeds. The greater the speed, the greater the weight, the larger the denominator, and the smaller the final weighted distance. This means that when the relative speed is large, the actual distance between the two objects is shortened, which is consistent with the intuition that more attention should be paid to the distance between the two when approaching at high speed in reality. The weighted distance can more accurately reflect the degree of threat posed by the target obstacle to path planning, especially when approaching at high speed or when the direction is unstable. The system can make more accurate obstacle avoidance and path adjustment decisions based on the weighted distance, thereby improving driving safety.
[0107] In an exemplary embodiment, standard training data that meets the task objectives is selected from the standard training database to form a training data set, including: setting a future frame feature sequence with a preset number of frames, the first frame of the sequence is the frame at time t+1, and the current frame of the target obstacle is the frame at time t; rotating the features of the target obstacle so that the feature direction is aligned with the movement direction of the obstacle, and converting the global coordinates of the target obstacle into local coordinates; and performing linear interpolation on the state of the future frame to reduce the prediction uncertainty caused by time fluctuation.
[0108] In the model training process of the autonomous driving system, in order to improve the model's prediction accuracy of future obstacle trajectories, training data that meets the mission objectives can be screened and processed from the standard training database. This process involves setting the future frame feature sequence, rotating the feature direction, and linearly interpolating the future frame state. Through these processing steps, the system can effectively reduce the prediction uncertainty caused by time fluctuations and improve the model's prediction ability.
[0109] Specifically, first, a future frame feature sequence with a preset number of frames can be set. The setting of this sequence is based on the model's prediction requirements for the trajectory of the target obstacle in the future time period. Usually, a suitable time window is selected, such as multiple time points in the next few seconds, and the target obstacle states at these time points are included in the prediction sequence. The first frame of the sequence corresponds to the frame at time t+1, and the current frame of the target obstacle is the frame at time t. In other words, first start from the current time point, take the current frame as the basis, and then look forward to predict the state of future frames. This method can help identify and predict the trajectory changes of obstacles in advance, providing a key reference for path planning and obstacle avoidance.
[0110] Then, the features of the target obstacle are rotated so that their direction is aligned with the direction of movement of the obstacle. The core of this process is to reduce the feature deviation caused by the change in the direction of movement of the obstacle. In the actual road environment, the direction of the obstacle may change. For example, when a vehicle changes lanes, turns, or makes a U-turn, its direction of movement will change significantly compared with the previous moment. By rotating the feature direction of the target obstacle to align with its direction of movement, the movement characteristics of the obstacle can be captured more accurately, thereby improving the prediction accuracy of its future trajectory. For example, if a car was going straight at the previous moment and starts to turn left at the current moment, through the rotation processing, the system can adjust the movement characteristics of the car to the current left turn direction, so that the model can more accurately predict its future driving path.
[0111] In addition, the global coordinates of the target obstacle are converted into local coordinates. The purpose of this step is to simplify data processing and improve the computational efficiency of the prediction. In the autonomous driving scenario, the global coordinates usually refer to the position of the target obstacle in the entire map, while the local coordinates refer to the position in the local reference system relative to the current coordinates of the vehicle. By converting the global coordinates into local coordinates, the system can more centrally analyze the relative position relationship between the obstacle and the vehicle. For example, the vehicle only needs to focus on obstacles within a few hundred meters ahead, rather than all obstacles in the global map. Local coordinate conversion can help the system process relevant data faster and more accurately, and improve real-time performance.
[0112] Finally, to reduce prediction uncertainty caused by time fluctuations, the system linearly interpolates the states of future frames. Time fluctuations usually refer to time differences between different frames due to inconsistent data acquisition frequencies or unstable timestamps. This time difference may cause instability in the prediction model and even lead to incorrect predictions. Linear interpolation is an effective method to generate smooth intermediate values between known time points to fill in the missing information in the time interval. For example, if the state of the target obstacle at time t and time t+2 is known, but the exact state at time t+1 is unknown, through linear interpolation, the system can estimate the state at time t+1, thereby providing a more continuous and smooth future trajectory prediction.
[0113] Through the above processing steps, the training data set obtained is used for model training, which can significantly improve the prediction accuracy of the future trajectory of the target obstacle. Setting the future frame feature sequence and rotation feature direction ensures that the movement trend and direction change of the obstacle can be accurately captured; converting the global coordinates to local coordinates can more focus on and efficiently process the obstacle information related to the vehicle; linear interpolation processing further reduces the prediction uncertainty caused by time fluctuations, making the model more robust and accurate in practical applications.
[0114] Please refer to Figure 3 As shown, based on the same inventive concept as the aforementioned autonomous driving model training data set construction method, the present invention provides an autonomous driving model training data set construction module 300, which includes a unified data layer 301 and a feature data layer 302.
[0115] The unified data layer 301 is used to obtain native data from multiple data sources and process the native data, including data annotation and format conversion, so as to convert the native data into standard training data in a unified format to form a standard training database. The feature data layer 302 is used to filter standard training data that meets the task objectives from the standard training database based on the task objectives of the model to form a training data set.
[0116] Please refer to Figure 4 As shown, an embodiment of the present invention further provides an electronic device 400, which includes at least one processor 401, a memory 402 (such as a non-volatile memory), a memory 403, and a communication interface 404, and the at least one processor 401, the memory 402, the memory 403, and the communication interface 404 are connected together via a bus 405. At least one processor 401 is used to call at least one program instruction stored or encoded in the memory 402, so that the at least one processor 401 performs various operations and functions of the method for constructing an autonomous driving model training data set described in various embodiments of this specification.
[0117] In an embodiment of the present specification, the electronic device 400 may include, but is not limited to, a personal computer, a server computer, a workstation, a desktop computer, a laptop computer, a notebook computer, a mobile electronic device, a smart phone, a tablet computer, a cellular phone, a personal digital assistant (PDA), a handheld device, a messaging device, a wearable electronic device, a consumer electronic device, and the like.
[0118] An embodiment of the present invention also provides a computer-readable medium, which carries computer execution instructions. When the computer execution instructions are executed by a processor, they can be used to implement various operations and functions of the method for constructing an autonomous driving model training data set described in the various embodiments of this specification.
[0119] The computer-readable medium in the present invention can be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. The computer-readable storage medium can be, for example, but not limited to, a system, device or component of electricity, magnetism, light, electromagnetic, infrared, or semiconductor, or any combination of the above. More specific examples of computer-readable storage media can include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present invention, a computer-readable storage medium can be any tangible medium containing or storing a program, which can be used by an instruction execution system, device or device or used in combination with it.
[0120] In the present invention, a computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, which carries a computer-readable program code. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. A computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium, which may send, propagate, or transmit a program for use by or in conjunction with an instruction execution system, device, or device. The program code contained on the computer-readable medium may be transmitted using any appropriate medium, including but not limited to: wireless, wire, optical cable, RF, etc., or any suitable combination of the above.
[0121] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0122] The present invention is described with reference to flowcharts and / or block diagrams of methods, apparatuses, systems, and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as a combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the process in the flowchart. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0123] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.
[0124] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.
Claims
1. A method for constructing an autonomous driving model training data set, characterized in that: include: Get raw data from multiple data sources; Processing the native data, wherein the processing includes data annotation and format conversion, so as to convert the native data into standard training data in a unified format to form a standard training database; Based on the task objective of the model, standard training data that meets the task objective is screened from the standard training database to form a training data set.
2. The method for constructing an autonomous driving model training data set according to claim 1, characterized in that: The native data includes data acquired by the unmanned vehicle in real time and / or data from public data sets; The data acquired by the unmanned vehicle in real time includes: scene data corresponding to when the model prediction effect is lower than expected during the operation of the unmanned vehicle and scene data encountered by the unmanned vehicle for the first time.
3. The method for constructing an autonomous driving model training data set according to claim 1, characterized in that: Native data includes perception data, positioning data, and map data; Processing the native data includes: aligning the time frames of the perception data and the positioning data so that the perception data and the positioning data maintain consistency on the time axis.
4. The method for constructing an autonomous driving model training data set according to claim 3, characterized in that: Performing time frame alignment on the perception data and the positioning data includes: Taking the time frame of the sensing data as a reference, querying the time frame of the positioning data closest to the time frame of the sensing data; If the time interval between the most recent positioning data time frame and the sensing data time frame is less than a preset threshold, the positioning data corresponding to the positioning data time frame is retained; otherwise, the positioning data corresponding to the positioning data time frame is discarded.
5. The method for constructing an autonomous driving model training data set according to claim 1, characterized in that: The standard training data that meets the task objectives is selected from the standard training database to form a training data set, including: Based on the task objectives of the model, the selected standard training data are matched, and different weights are assigned according to the labeled categories of the data to form a training data set.
6. The method for constructing an autonomous driving model training data set according to claim 1, characterized in that: Selecting standard training data that meets the task objectives from the standard training database includes: Filter static target obstacles and slow-moving target obstacle data. The slow-moving target obstacles include target obstacles with a speed less than 0.01 m / s and target obstacles that move less than 2 m in the next 8 seconds. According to the speed of the target obstacle, the shortest number of historical frames required to generate an accurate trajectory is calculated, and the target obstacle data with a number of historical frames less than the preset minimum value is filtered out; According to the angle between the speed direction and the displacement direction of the target obstacle, the target obstacle data with an angle greater than 30° among motor vehicle target obstacles and the target obstacle data with an angle greater than 90° among bicycle target obstacles are filtered out.
7. The method for constructing an autonomous driving model training data set according to claim 1, characterized in that: The standard training data that meets the task objectives is selected from the standard training database to form a training data set, including: The lane centerline and lane boundary data are segmented according to the preset interval parameters, and linear interpolation is performed on each segmented line segment, and the segmented sub-map elements are sorted according to the weighted distance between the segmented sub-map elements and the target obstacles; Add points at the end of the coordinate sequence of the crosswalk to form a complete ring, divide the ring element twice, and perform linear interpolation on each segment.
8. The method for constructing an autonomous driving model training data set according to claim 7, characterized in that: The calculation formula of the weighted distance between the sub-map element and the target obstacle includes: target_to_other_angle=arctan2(other_y-target_y,other_x-target_x); other_to_target_angle=arctan2(target_y-other_y,target_x-other_x); target_velocity_angle=arctan2(target_vy,target_vx); other_to_target_angle=arctan2(other_vy,other_vx); relative_speed_target=(target_speed*(π-|normalize_angle(target_to_other_angle-target_velocity_angle)|)) / π; relative_speed_other=(other_speed*(π-|normalize_angle(other_velocity_angle-other_to_target_angle)|)) / π; weight_distance=normal_distance / (c_coef+relative_speed_other+relative_speed_target); Among them, target_to_other_angle represents the direction angle from the target obstacle to the submap element, [target_x, target_y] represents the coordinates of the target obstacle at the current moment, [other_x, other_y] represents the coordinates of a point in the submap element, [target_vx, target_vy] represents the velocity vector of the target obstacle at the current moment, [other_vx, other_vy] represents the velocity vector of a point in the submap element, other_to_target_angle represents the direction angle from the submap element to the target obstacle, relative_speed_target represents the speed of the target obstacle relative to the submap element, target_speed represents the speed of the target obstacle, target_velocity_angle represents the speed direction of the target obstacle, relative_speed_other represents the speed of the submap element relative to the target obstacle, other_speed represents the speed of the submap element, other_velocity_angle represents the speed direction of the submap element, weight_distance represents the weighted distance between the submap element and the target obstacle, normal_distance represents the Euclidean distance between the submap element and the target obstacle, and c_coef is a constant.
9. The method for constructing an autonomous driving model training data set according to claim 1, characterized in that: The standard training data that meets the task objectives is selected from the standard training database to form a training data set, including: Set a future frame feature sequence with a preset number of frames, where the first frame of the sequence is the frame at time t+1, and the current frame of the target obstacle is the frame at time t; Rotate the features of the target obstacle to align the feature direction with the movement direction of the obstacle, and convert the global coordinates of the target obstacle into local coordinates; The states of future frames are linearly interpolated to reduce prediction uncertainty caused by temporal fluctuations.
10. An autonomous driving model training data set construction module, characterized in that: include: A unified data layer, used to obtain native data from multiple data sources and process the native data, wherein the processing includes data annotation and format conversion, so as to convert the native data into standard training data in a unified format to form a standard training database; The feature data layer is used to filter standard training data that meets the task objectives from the standard training database based on the task objectives of the model to form a training data set.
11. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, it implements the method for constructing an autonomous driving model training data set as described in any one of claims 1 to 9.
12. A computer readable medium, characterized in that The computer-readable medium carries computer-executable instructions, which, when executed by a processor, are used to implement the method for constructing an autonomous driving model training data set as described in any one of claims 1 to 9.
Citation Information
Patent Citations
Model training method and device for automatic driving scene and electronic equipment
CN111694973A
System and method for predicting cut-in track of driverless car to surrounding vehicles
CN114067178A
Model training system, model training method and device
CN115099352A
Automatic driving scene generation method and related model training method and device
CN115438569A
Automatic driving control method and device, equipment and medium
CN117519134A
Cited By
Composite scene obstacle non-graphical trajectory prediction method
CN120440068A
Data processing method, computing device and computer readable storage medium
CN121188481A
