A method and device for processing automatic driving model data based on a vehicle-road-cloud integrated architecture, an automatic driving model self-evolution learning system, equipment and a medium

Abstract

Embodiments of the present application disclose a kind of automatic driving model data processing method, device, automatic driving model self-evolution learning system, equipment and medium under the integration architecture of vehicle-road cloud, wherein, data processing method includes: receiving road side data and car end perception data;From road side data, extract high-value scene segment data;3D static road scene is reconstructed based on car end perception data, and multiple dynamic traffic participants are identified and extracted from car end data;Multiple dynamic traffic participants, the 3D static road scene of reconstruction and high-value scene segment data are combined, and road traffic scene is generated;For each target dynamic traffic participant in road traffic scene, by switching the perspective of each target dynamic traffic participant, obtain road traffic data under multiple perspectives, and the road traffic data is used to train automatic driving model.Through the above technical scheme, the problem of insufficient model training data completeness is solved.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of autonomous driving technology, specifically to a method, apparatus, autonomous driving model self-evolutionary learning system, device, and medium for processing autonomous driving model data based on a vehicle-road-cloud integrated architecture. Background Technology

[0002] With the vigorous development of a new round of technological revolution and industrial transformation, automobiles are deeply integrated with technologies such as artificial intelligence, information and communication, and big data analytics, making intelligentization and connectivity important directions for the development of the automotive industry. The "National Guidelines for the Construction of the Internet of Vehicles Industry Standard System (Intelligent Connected Vehicles)" released in 2023 proposes that by 2030, a comprehensive intelligent connected vehicle standard system capable of supporting the coordinated development of single-vehicle intelligence and network-enabled capabilities will be fully established.

[0003] Currently, with the advancement of pilot cities for vehicle-road-cloud integration, the construction of intelligent roadside infrastructure is unfolding on a large scale, and various operation and maintenance companies have accumulated a large amount of roadside data. Roadside equipment can acquire traffic flow data from a comprehensive perspective, over a long period of time, at a high frequency, and with a wide field of view, covering different vehicle types, seasons, lighting / temperature, and holidays. However, due to a lack of high-value scenario identification and extraction tools, this data has low utilization value, resulting in significant resource waste. On the other hand, major companies and emerging car manufacturers are entering the autonomous driving model development field, hoping to seize the technological and application high ground. However, the siloed development model of each car company leads to isolated and incomplete model training data, failing to meet the requirements for training data completeness. The trained models have poor generalization ability and robustness, making it difficult to adapt to complex and ever-changing real-world traffic environments. Summary of the Invention

[0004] This invention provides a method, apparatus, self-evolutionary learning system, device, and medium for processing autonomous driving model data based on a vehicle-road-cloud integrated architecture, in order to solve the problems of low utilization of roadside data and insufficient completeness of model training data.

[0005] In a first aspect, this invention provides a method for processing autonomous driving model data based on a vehicle-road-cloud integrated architecture. The vehicle-road-cloud integrated architecture includes a data acquisition vehicle, roadside perception devices, and a cloud platform. The processing method provided in this embodiment is applied to the cloud platform and includes:

[0006] Receive roadside data uploaded by roadside sensing devices, and receive vehicle-side sensing data uploaded by data collection vehicles, wherein the roadside data includes traffic flow data;

[0007] High-value scene fragment data is extracted from roadside data. The high-value scene fragment data includes the driving trajectory information of multiple dynamic traffic participants in different scenarios. The driving trajectory information includes the location information and speed information of the trajectory points.

[0008] The 3D static road scene is reconstructed based on vehicle-side perception data, and multiple dynamic traffic participants are identified and extracted from the vehicle-side perception data.

[0009] By combining multiple dynamic traffic participants, reconstructed 3D static road scenes, and high-value scene fragment data, a road traffic scene is generated.

[0010] For each target dynamic traffic participant in a road traffic scenario, road traffic data from multiple perspectives is obtained by switching the viewpoints of each target dynamic traffic participant. This road traffic data is used to train an autonomous driving model.

[0011] Optionally, the method provided in this embodiment further includes:

[0012] Test the trained autonomous driving model and deploy the tested autonomous driving model to intelligent connected vehicles.

[0013] Optionally, high-value scene fragments can be extracted from roadside data, including:

[0014] Identify the intersections and road segments corresponding to the roadside data, and convert the high-precision map data corresponding to the intersections and road segments into local map data that meets the training data format requirements;

[0015] The driving trajectory information of each dynamic traffic participant within a set time period is extracted from the roadside data. Based on the starting point location information of each dynamic traffic participant, a search is performed in the local map to obtain the candidate driving trajectory of each dynamic traffic participant. The candidate driving trajectory is then smoothed.

[0016] Based on the smoothed candidate driving trajectories and combined with local map data, the behavioral characteristics of each dynamic traffic participant are determined. These behavioral characteristics include lane affiliation, collision characteristics, distance between trajectory points, distance between target contours, and traffic speed and volume in each lane.

[0017] Based on the behavioral characteristics of each dynamic traffic participant, the target behavior of each dynamic traffic participant is determined, and a corresponding behavior tag is added to each target behavior. The behavior tags are then stored in the scene tag library, which also includes various types of scene tags, each type of scene tag including multiple levels of sub-tags.

[0018] The target behavior labels and various scene labels that meet the model training requirements are selected from the scene label library. The selected labels are then combined, and the combined labels and their corresponding target behavior data are used as high-value scene fragment data.

[0019] Optionally, the high-precision map data corresponding to the intersection and road segment can be converted into local map data that meets the training data format requirements, including:

[0020] Convert the high-precision map data corresponding to the intersection and road segment from map format to the geographic data format ShapeFile;

[0021] Road elements are extracted from the map data after map format conversion. These road elements include lane center lines, lane lines, road lines, and intersection areas.

[0022] The extracted road elements are transformed so that the data format of the transformed road elements meets the requirements of the training data. The transformation includes converting the extracted road elements from the geocentric coordinate system WGS84 to the universal transverse Mercator coordinate system UTM, the transformation of the enumeration type, and the data point sparsification process.

[0023] Based on the road elements after data format conversion and the topological connections between each road element, local map data that meets the requirements of the training data format is obtained.

[0024] Optionally, the selected tags can be combined, including:

[0025] The selected tags are combined according to preset tag combination rules, which include any one or more of the following: logically combining different tags according to the relationship between different tags, sorting the tags according to time order, assigning different weights to different tags, and adding different priorities to different tags.

[0026] Secondly, embodiments of the present invention also provide a processing device for autonomous driving model data based on a vehicle-road-cloud integrated architecture, the processing device comprising:

[0027] The data receiving module is configured to receive roadside data uploaded by roadside sensing devices, and to receive vehicle-side sensing data uploaded by data collection vehicles, wherein the roadside data is traffic flow data;

[0028] The high-value scene fragment data extraction module is configured to extract high-value scene fragment data from roadside data. The high-value scene fragment data includes the driving trajectory information of multiple dynamic traffic participants in different scenarios. The driving trajectory includes the location information and speed information of the trajectory points.

[0029] The static road scene reconstruction module is configured to reconstruct 3D static road scenes based on vehicle-side perception data, and to identify and extract multiple dynamic traffic participants from the vehicle-side perception data.

[0030] The road traffic scene generation module is configured to combine multiple dynamic traffic participants, reconstructed 3D static road scenes, and high-value scene fragment data to generate road traffic scene data.

[0031] The training data generation module is configured to obtain road traffic data from multiple perspectives for each target dynamic traffic participant in a road traffic scenario by switching the perspectives of each target dynamic traffic participant. This road traffic data is used to train the autonomous driving model.

[0032] Optionally, the apparatus provided in this embodiment of the invention further includes:

[0033] The model deployment module is configured to test the trained autonomous driving model and deploy the tested autonomous driving model to intelligent connected vehicles.

[0034] Optional, a high-value scene fragment data extraction module, specifically including:

[0035] The local map data conversion unit is configured to determine the intersection and road segment corresponding to the roadside data, and convert the high-precision map data corresponding to the intersection and road segment into local map data that meets the training data format requirements;

[0036] The driving trajectory determination unit is configured to extract the driving trajectory information of each dynamic traffic participant within a set time period from the roadside data, and search in the local map according to the starting point location information of each dynamic traffic participant to obtain the candidate driving trajectory of each dynamic traffic participant, and smooth the candidate driving trajectory.

[0037] The feature information determination unit is configured to determine the behavioral feature information of each dynamic traffic participant based on the smoothed candidate driving trajectory and in combination with local map data. The behavioral features include lane affiliation, collision features, distance between trajectory points, distance between target contours, and traffic speed and traffic volume of each lane.

[0038] The target behavior determination unit is configured to determine the target behavior of each dynamic traffic participant based on the behavioral characteristic information of each dynamic traffic participant, add corresponding behavior tags to each target behavior, and store the behavior tags in a scene tag library. The scene tag library also includes multiple types of scene tags, and each type of scene tag includes multiple levels of sub-tags.

[0039] The label combination unit is configured to select target behavior labels and various scene labels that meet the model training requirements from the scene label library, combine the selected labels, and use the combined labels and their corresponding target behavior data as high-value scene fragment data.

[0040] Optional, the local map data transformation unit is specifically configured as follows:

[0041] Convert the high-precision map data corresponding to the intersection and road segment into the geographic data format ShapeFile;

[0042] Road elements are extracted from the map data after map format conversion. The road elements include lane center lines, lane lines, road lines, and intersection areas.

[0043] The extracted road elements are transformed so that the data format of the transformed road elements meets the requirements of the training data. The transformation includes converting the extracted road elements from the geocentric coordinate system WGS84 to the universal transverse Mercator coordinate system UTM, the transformation of the enumeration type, and the data point sparsity processing.

[0044] Based on the road elements after data format conversion and the topological connections between each road element, local map data that meets the requirements of the training data format is obtained.

[0045] Optional, the tag combination unit is specifically configured as follows:

[0046] Select target behavior labels and various scene labels that meet the model training requirements from the scene label library;

[0047] The selected tags are combined according to preset tag combination rules, wherein the preset tag combination rules include any one or more of the following: logically combining different tags according to the relationship between different tags, sorting the tags according to the time order, assigning different weights to different tags, and adding different priorities to different tags;

[0048] The combined labels and their corresponding target behavior data are used as high-value scene fragment data.

[0049] Thirdly, embodiments of the present invention also provide a self-evolving learning system for a large-scale driving model based on a vehicle-road-cloud integrated architecture, the learning system comprising:

[0050] Data collection vehicles are used to collect vehicle-side perception data and upload the vehicle-side perception data to the cloud control platform. The vehicle-side perception data includes information about the surrounding environment and the status information of traffic participants.

[0051] Roadside sensing devices are used to collect roadside data and upload it to the cloud control platform. The roadside data includes traffic flow data.

[0052] The cloud-based control platform is used to preprocess vehicle-side perception data and roadside data, and then upload the preprocessed vehicle-side perception data and roadside data to the cloud.

[0053] The cloud, used for:

[0054] Receive pre-processed vehicle-side perception data, and receive pre-processed roadside data;

[0055] High-value scene fragment data is extracted from roadside data. The high-value scene fragment data includes the driving trajectory information of multiple dynamic traffic participants under different scene labels. The driving trajectory information includes the location information and speed information of the trajectory points.

[0056] The 3D static road scene is reconstructed based on vehicle-side perception data, and multiple dynamic traffic participants are identified and extracted from the vehicle-side perception data.

[0057] Road traffic scene data is generated by combining multiple dynamic traffic participants, reconstructed 3D static road scenes, and high-value scene fragment data.

[0058] For each target dynamic traffic participant in a road traffic scenario, by switching the perspectives of each target dynamic traffic participant, road traffic data from multiple perspectives is obtained, and the autonomous driving model is trained using the road traffic data.

[0059] Test the trained autonomous driving model and deploy the tested autonomous driving model to intelligent connected vehicles.

[0060] Fourthly, embodiments of the present invention also provide a computing device, comprising:

[0061] Memory containing executable program code;

[0062] A processor coupled to the memory;

[0063] The processor calls the executable program code stored in the memory to execute the autonomous driving model data processing method based on the vehicle-road-cloud integrated architecture provided in any embodiment of the present invention.

[0064] Fifthly, embodiments of the present invention also provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for processing autonomous driving model data based on a vehicle-road-cloud integrated architecture provided in any embodiment of the present invention.

[0065] The technical solution provided in this invention identifies high-value scene fragments from roadside data. These fragments can then be used to generate road traffic data from multiple vehicle-side perspectives. This approach not only improves the utilization rate of roadside data but also effectively enhances the completeness and diversity of vehicle-side data. Training an autonomous driving model using this multi-perspective vehicle-side data significantly improves the model's generalization ability and robustness, enabling it to adapt to complex and ever-changing real-world traffic environments. Furthermore, the trained autonomous driving model has no limitations in its use and can be deployed in vehicles from various brands, overcoming the drawback of related technologies where training results from different automakers are only applicable to their own brand's vehicles and cannot be shared.

[0066] The innovative aspects of this invention include:

[0067] 1. By identifying roadside data, high-value scene fragments can be obtained. This approach not only improves the utilization rate of roadside data but also increases the diversity of training samples for autonomous driving models. Compared to using only vehicle-mounted data as training samples for autonomous driving models, the use of roadside data in this embodiment enriches the content of the training samples and effectively improves the generalization ability of the trained model, which is one of the innovations of this embodiment.

[0068] 2. By utilizing high-value scene fragments extracted from roadside data and combining them with vehicle-mounted perception data, road traffic data from the vehicle's perspective can be generated. This approach effectively improves the completeness and diversity of vehicle-mounted data. Training the autonomous driving model using vehicle-mounted data from multiple perspectives effectively enhances the model's generalization ability and robustness, enabling it to adapt to complex and ever-changing real-world traffic environments. This is one of the innovative aspects of this invention. Attached Figure Description

[0069] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0070] Figure 1a This is a schematic diagram of a vehicle-road-cloud integrated data closed loop provided in Embodiment 1 of the present invention;

[0071] Figure 1b This is a flowchart of a method for processing autonomous driving model data based on a vehicle-road-cloud integrated architecture, provided in Embodiment 1 of the present invention.

[0072] Figure 2 This is a schematic diagram of the structure of a self-evolutionary learning system for a large-scale driving model based on a vehicle-road-cloud integrated architecture, provided in Embodiment 2 of the present invention.

[0073] Figure 3 This is a structural block diagram of an autonomous driving model data processing device based on a vehicle-road-cloud integrated architecture, provided in Embodiment 3 of the present invention.

[0074] Figure 4 This is a schematic diagram of the structure of a computing device provided in Embodiment 4 of the present invention. Detailed Implementation

[0075] 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.

[0076] It should be noted that the terms "comprising" and "having," and any variations thereof, in the embodiments and drawings of this invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the steps or units listed, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0077] This invention discloses a method, apparatus, self-evolutionary learning system, device, and medium for processing autonomous driving model data based on a vehicle-road-cloud integrated architecture. These are described in detail below.

[0078] Example 1

[0079] The autonomous driving model data processing method provided in Embodiment 1 of the present invention is based on a vehicle-road-cloud integrated architecture. Figure 1a This is a schematic diagram of a vehicle-road-cloud integrated data closed loop provided in Embodiment 1 of the present invention, as shown below. Figure 1a As shown, the vehicle-road-cloud integrated architecture includes roadside sensing devices (i.e., Figure 1a Roadside data collection equipment and data collection vehicles (i.e., roadside data collection equipment) Figure 1a Intelligent driving in China and cloud computing (including smart cars) and cloud computing (including smart cars) Figure 1aThe system comprises data generation and training / deployment tools. Roadside perception devices, including roadside cameras and radar, acquire data on traffic participants, traffic events, and traffic flow status within an area of ​​interest. Data collection vehicles use onboard cameras and radar to obtain information about the surrounding environment and the status of traffic participants. Both roadside perception devices and data collection vehicles upload the collected data to a cloud control platform. The platform preprocesses, stores, and manages the received data before uploading it to the cloud. Cloud-based data generation tools generate a training dataset for the autonomous driving model based on the received roadside and vehicle-side data. Users can then train the autonomous driving model using the training / deployment tools, test and validate the model, and output a reliable large-scale model, which can be deployed to the vehicle via OTA (Over-The-Air) technology. Finally, roadside data collection devices and intelligent vehicles can collect new high-value data again and transmit it to the cloud. The cloud enables automatic optimization of the autonomous driving model, and then the optimized autonomous driving model is deployed to the vehicle via OTA tools, realizing a closed-loop data system integrating vehicles, roads, and the cloud.

[0080] Figure 1b This is a flowchart illustrating a method for processing autonomous driving model data based on a vehicle-road-cloud integrated architecture, as provided in Embodiment 1 of the present invention. This method is applied to the cloud within the vehicle-road-cloud integrated architecture and can be executed by a cloud-based data generation tool, which can be implemented through software and / or hardware. Figure 1b As shown, the method provided in this embodiment specifically includes:

[0081] S110: Receive roadside data uploaded by roadside sensing devices, and receive vehicle-side sensing data uploaded by data collection vehicles.

[0082] Roadside data includes traffic flow data for different vehicle types, seasons, lighting / temperature conditions, and holidays. This data is collected through roadside sensing devices, such as roadside cameras and radar installed on the road. These devices can acquire roadside data from a comprehensive, long-term, high-frequency, and wide-field perspective, and can utilize single-point sensing, cross-point sensing, and continuous regional sensing technologies to improve the quality of the data. Vehicle-side sensing data includes traffic environment information and traffic participant status information. This data is collected through onboard cameras and radar in the data collection vehicles.

[0083] In this embodiment, to facilitate cloud-based processing of roadside data and vehicle-mounted sensing data, the roadside sensing devices and data collection vehicles can each upload the collected data to the cloud control platform. The cloud control platform first preprocesses the roadside data and vehicle-mounted sensing data, performing operations such as noise filtering and invalid data deletion, and then uploads the preprocessed roadside data and vehicle-mounted sensing data to the cloud. The roadside data and vehicle-mounted sensing data processed in the cloud are the data preprocessed by the cloud control platform.

[0084] S120: Extract high-value scene fragment data from roadside data.

[0085] In this embodiment, the high-value scene fragment data includes the driving trajectory information of multiple dynamic traffic participants in different scenarios. These dynamic traffic participants include pedestrians, vehicles, etc., and their driving trajectory information includes the location and speed information of trajectory points. Different scenarios can be represented by scene tags.

[0086] In this embodiment, high-value scene fragments extracted from roadside data can be used to generate training samples for the autonomous driving model. This setup not only improves the utilization rate of roadside data but also increases the diversity of training samples for the autonomous driving model. Compared to using only vehicle-mounted data as training samples for the autonomous driving model, the use of roadside data in this embodiment enriches the content of the training samples and effectively improves the generalization ability of the trained model.

[0087] As an optional implementation method, extracting high-value scene fragment data from roadside data can be achieved through the following steps A to E:

[0088] A. Determine the corresponding intersections and road segments based on roadside data, and convert the high-precision map data corresponding to the intersections and road segments into local map data that meets the training data format requirements.

[0089] In this embodiment, the corresponding intersection and road segment can be determined based on the location of the roadside sensing device, and high-precision map data corresponding to that intersection and road segment can be extracted. High-precision map data is typically used in high-precision navigation for autonomous driving. However, during the training phase of the autonomous driving model, it needs to be converted into local map data that meets the training data format requirements. These training data format requirements include map format requirements and coordinate system transformations for road elements, which can be achieved through the following steps:

[0090] a1. Convert the high-precision map data corresponding to the intersection and road segment into the geographic data format (ShapeFile) required for model training, such as OpenDrive (an open file format) and NDS (Navigation Data Standard).

[0091] a2. Extract road elements from the map data after map format conversion.

[0092] The road elements include lane center lines, lane lines, road lines, and intersection areas.

[0093] a3. Transform the extracted road elements so that the data format of the transformed road elements meets the requirements of the training data.

[0094] The data format conversion of road elements includes: converting the extracted road elements from the geocentric coordinate system (WGS84 coordinate system) to UTM (Universal Transverse Mercator coordinate system), converting the enumeration type, and data point sparsification processing.

[0095] a4. Based on the road elements after data format conversion and the topological connection relationships between each road element, obtain local map data that meets the requirements of the training data format.

[0096] Specifically, the topological connections of each road element in the local map after data conversion can be determined based on the topological connections of each road element in the high-precision map.

[0097] In this embodiment, by adopting the above technical solution, high-precision map data of a certain intersection segment can be converted into local map data that meets the requirements of training data format.

[0098] B. Extract the driving trajectory information of each dynamic traffic participant within a set time period from the roadside data, and search in the local map according to the starting point location information of each dynamic traffic participant to obtain the candidate driving trajectory of each dynamic traffic participant, and smooth the candidate driving trajectory.

[0099] The time period can be set according to actual needs. In this embodiment, a sliding window approach can be used, with a processing cycle of several time domain lengths (e.g., 10-60 seconds) to extract the driving trajectory information of all dynamic traffic participants within this time period. For each dynamic traffic participant, a driving path search can be performed based on the participant's starting point position and the local map to obtain all possible driving paths, i.e., candidate driving paths. Then, a trajectory optimization algorithm, such as the FemPosSmooth algorithm (an algorithm for smoothing reference lines), can be used to optimize the driving trajectory of the dynamic traffic participant to mitigate position jumps in the driving trajectory. Then, based on the processed trajectory information, a filtering algorithm (e.g., the Kalman filter algorithm) can be used to optimize the speed information in the driving trajectory of the dynamic traffic participant, finally obtaining the processed trajectory and speed information.

[0100] C. Based on the smoothed candidate driving trajectories and combined with local map data, determine the behavioral characteristics of each dynamic traffic participant.

[0101] Among them, behavioral characteristics include lane affiliation, collision characteristics, distance between trajectory points, distance between target contours, and traffic speed and traffic volume in each lane. This embodiment does not specifically limit the types and number of behavioral characteristics of dynamic participants.

[0102] Specifically, regarding the behavioral characteristic of lane affiliation, the current lane and the target lane that the current dynamic traffic participant wants to change to can be determined based on the driving trajectory information of the current dynamic traffic participant and in combination with the local map.

[0103] For collision features, collision detection can be performed based on various collision detection algorithms to predict potential collisions or already occurred between current dynamic traffic participants and other dynamic traffic participants.

[0104] For the behavioral feature of trajectory distance, the Euclidean distance between different trajectory points at any time or consecutive time can be calculated, as well as the Euclidean distance between the contours of dynamic traffic participants.

[0105] For the behavioral characteristic of traffic flow speed, the average speed distribution and traffic volume of all lanes can be calculated.

[0106] D. Based on the behavioral characteristics of each dynamic traffic participant, determine the target behavior of each dynamic traffic participant, add corresponding behavior tags to each target behavior, and store each behavior tag in the scene tag library.

[0107] The behaviors of dynamic traffic participants include lane changes, congestion, other vehicles cutting into the current lane, and obstacle avoidance. This embodiment does not specifically limit the types of behaviors of dynamic traffic participants.

[0108] Specifically, for lane changing behavior, the behavior characteristics of lane ownership can be used to determine whether dynamic traffic participants want to change lanes, and indicators such as the number of lane changes, lane change time, and lane change lateral speed / acceleration can be calculated. Then, collision characteristics and trajectory distance can be used to determine dangerous lane changing situations.

[0109] For the behavior of congestion and slow traffic, the congestion and slow traffic scenario can be determined based on behavioral characteristics such as lane affiliation and traffic flow speed.

[0110] For the behavior of other vehicles cutting into the current lane, it can be determined whether there are other vehicles cutting into the current lane of a dynamic traffic participant based on lane affiliation, collision characteristics and trajectory distance, and the TTC (Time To Collision, the estimated time when two vehicles may collide) and the shortest distance can be calculated.

[0111] For obstacle detour behavior, behavioral characteristics such as lane affiliation and traffic speed can be used, and the obstacle detour scenario can be determined based on the detected obstacle type.

[0112] For the typical behaviors of dynamic traffic participants, in the actual model training process, it is necessary to combine the needs of the users and carry out targeted parameter calibration and feature value selection in order to better meet the specific needs of the users.

[0113] After determining the target behaviors of dynamic traffic participants, corresponding behavior labels can be added to each target behavior data segment, and these labels can be stored in a scene label library for subsequent training of autonomous driving models. The scene label library stores various types of labels, including lighting condition labels, weather condition labels, time period labels, visibility labels, road type labels, traffic flow status labels, and target behavior labels. Each category is further subdivided into multiple subcategories, such as the intensity level of lighting conditions, weather variations (sunny or rainy), data from different time periods, visibility distance classifications, road type, and specific actions of the target behavior.

[0114] E. Select target behavior labels and various scene labels that meet the model training requirements from the scene label library, combine the selected labels, and use the combined labels and their corresponding target behavior data as high-value scene fragment data. Each type of scene label includes multiple levels of sub-labels.

[0115] In this embodiment, high-value scenarios are defined using tags, which can effectively extract high-value scenarios for autonomous driving from real-world data. To ensure the comprehensiveness and effectiveness of high-value scenarios, the tagging operations in this embodiment mainly include two parts: tag filtering and tag combination.

[0116] During the tag selection phase, scene tags are filtered based on a series of predefined criteria. These predefined criteria include several major categories such as lighting conditions, weather conditions, time of day, visibility, road type, traffic flow status, and target behavior. Each criterion is further subdivided into multiple subcategories, such as the intensity level of lighting conditions, weather changes (sunny or rainy), data from different time periods, visibility distance classification, road type, and specific actions of the target behavior.

[0117] During the label combination stage, the selected labels can be combined according to pre-set label combination rules to ensure that the combined labels meet the model training requirements. The pre-set label combination rules may include any one or more of the following:

[0118] (1) Logical combination, such as using logical AND, OR, NOT conditions to determine the relationship between tags;

[0119] (2) Sequence combination, such as sorting the labels according to the time series, such as sorting the labels according to the order of changing lanes first, then cutting into a certain lane, and then changing lanes again after cutting into a certain lane;

[0120] (3) Weight allocation, that is, assigning different weights to different labels to reflect their influence in scene recognition;

[0121] (4) Priority combination, that is, different priorities are given according to the importance of the label.

[0122] The following example, which involves selecting high-value scenarios with low perceived risk during driving, illustrates the process of tag combination:

[0123] ① Use logical combinations (such as logical "AND") and conditional combinations, including labels such as "lighting conditions (low light)", "weather conditions (rainy day)" and "time of day (night)" to perform preliminary filtering of the data;

[0124] ② Use sequence combination to match the tags such as "target behavior (CutIn)", "target behavior (lane change)", and "target behavior (CutIn)" in chronological order to perform secondary filtering on the data;

[0125] ③ Use weight allocation to assign weights to labels such as "road type (intersection area)", "visibility (low visibility)" and "traffic flow status (high traffic volume)" from high to low, and filter the data again according to the set threshold of weight and value, such as extracting data with weight and value higher than the set threshold.

[0126] ④ Use priority combinations to sort the data according to the priority of the label "Most Quantity (Target Behavior (CutIn))" which is higher than the priority of the label "Most Quantity (Target Behavior (Lane Change))".

[0127] Following the above label combination process, after combining all the labels, the resulting data is high-value scene data, which is dynamic traffic participant trajectory data containing various scene labels and behavior labels. These data fragments are written into the training sample database and can be used by subsequent model training software.

[0128] It should be noted that the high-value scenario data extraction provided in this embodiment employs tag filtering and tag combination methods to meet users' needs for high-value scenario extraction. This embodiment does not specifically limit the tag filtering and combination rules. In actual engineering applications, users can customize high-value scenarios according to their actual usage needs and can perform targeted tag filtering and combination processing based on actual definitions.

[0129] S130. Reconstruct the 3D static road scene based on vehicle-side perception data, and identify and extract multiple dynamic traffic participants from the vehicle-side perception data.

[0130] The vehicle-side perception data includes information about the surrounding environment and the status of traffic participants. Based on this data, a 3D static road scene in a traffic scenario can be reconstructed. There are various methods for 3D scene reconstruction. For example, 3DGS (3D Gaussian Splatting) technology can be used to explicitly represent a lattice in 3D space, effectively organizing the transmitted information in a Gaussian distribution to achieve real-time image rendering. Alternatively, a laser beam can be used to illuminate the object's surface, and accurate 3D data can be obtained by measuring the laser reflection time or angle. Another method is to use deep learning algorithms to train a large amount of image data to learn the 3D shape and appearance information of objects. This embodiment does not specifically limit the 3D scene reconstruction method. For the reconstructed 3D static road scene, the static traffic participants included, such as buildings, traffic signs, streetlights, and green belts, can be accurately labeled using target 3D annotation methods.

[0131] In addition, dynamic traffic participants can be identified from the vehicle-side perception data. In this embodiment, information on dynamic traffic participants, including vehicles and pedestrians, can be extracted from the vehicle-side perception data for use in the generation of subsequent road traffic scenarios.

[0132] S140. Combine multiple dynamic traffic participants, reconstructed 3D static road scenes, and high-value scene fragment data to generate a road traffic scene.

[0133] In this embodiment, multiple dynamic traffic participants extracted from vehicle-side perception data can be matched with the driving trajectory information of multiple dynamic traffic participants in high-value scene fragment data to obtain the driving trajectory information of multiple target dynamic traffic participants. Then, the driving trajectory information of multiple target dynamic traffic participants can be matched with the reconstructed 3D static road scene, and the driving trajectory information of the target dynamic traffic participants can be embedded into the reconstructed 3D static road scene in real time according to the timestamp. During this process, the consistency of these two parts of data in time and space must be ensured.

[0134] Specifically, the reconstructed 3D static road scene data can be imported into the selected 3D modeling software or traffic simulation platform. During this process, the scene scale, coordinate system, etc., can be adjusted as needed to ensure that the 3D static road scene is consistent with the actual traffic conditions. Then, the driving trajectory information of each dynamic participant can be mapped onto the 3D static road scene according to the timestamp. In this process, it is necessary to ensure that the time in the 3D scene is consistent with the timestamp of the dynamic traffic participants' driving trajectory. After obtaining the road traffic scene, it can be displayed through professional visualization tools or platforms.

[0135] S150. For each target dynamic traffic participant in the road traffic scenario, road traffic data from multiple perspectives is obtained by switching the perspectives of each target dynamic traffic participant.

[0136] In this embodiment, by switching the perspectives of various target dynamic traffic participants, road traffic data from multiple perspectives can be obtained. This road traffic data can be used to train the autonomous driving model. After training, the trained autonomous driving model can be tested. The testing process can be achieved through a model deployment tool. This tool allows for rapid deployment and verification, ensuring the model's performance on the vehicle-side hardware. Furthermore, the deployment tool software supports the integration of deployment operations and code reuse across different hardware platforms, simplifying the deployment process.

[0137] Furthermore, the tested autonomous driving models can be deployed to intelligent connected vehicles of various brands via OTA tools, and the autonomous driving models of different brands of vehicles can be upgraded via OAT tools.

[0138] In related technologies, each automaker typically collects road data only through its own road survey vehicles, and this data is not shared with other automakers. This results in limited and incomplete training data for autonomous driving models. Furthermore, each automaker only trains its own self-developed autonomous driving models, and the training results are not shared. Autonomous driving models trained using such data have poor generalization ability and robustness, making them difficult to adapt to complex and changing real-world traffic environments. In this embodiment,

[0139] By identifying roadside data, high-value scene fragments can be obtained. These fragments can then be used to generate road traffic data from the vehicle's perspective. This approach not only improves the utilization rate of roadside data but also effectively enhances the completeness and diversity of vehicle-side data. Training autonomous driving models using vehicle-side data from multiple perspectives significantly improves the generalization ability and robustness of the trained models, enabling them to adapt to complex and ever-changing real-world traffic environments. Furthermore, the trained autonomous driving models have no limitations in their use and can be deployed in vehicles from various brands, overcoming the drawback of previous technologies where training results from different automakers were only applicable to their own brand's vehicles and could not be shared.

[0140] Example 2

[0141] Figure 2 This is a schematic diagram of the structure of a self-evolutionary learning system for a large-scale driving model based on a vehicle-road-cloud integrated architecture, as provided in Embodiment 2 of the present invention. Figure 2 As shown, the system provided in this embodiment includes a data collection vehicle 210, a roadside sensing device 220, a cloud control platform 230, and a cloud platform 240, wherein...

[0142] The data collection vehicle 210 is used to collect vehicle-side perception data and upload the vehicle-side perception data to the cloud control platform 230. The vehicle-side perception data includes surrounding environment information and traffic participant status information.

[0143] The roadside sensing device 220 is used to collect roadside data and upload the roadside data to the cloud control platform 230. The roadside data includes traffic flow data.

[0144] The cloud control platform 230 is used to preprocess vehicle-side perception data and roadside data, and upload the preprocessed vehicle-side perception data and roadside data to the cloud 240.

[0145] The cloud-based 240 is configured with data generation tools, model training tools, and model deployment tools. The data generation tools are used for:

[0146] Receive pre-processed vehicle-side perception data, and receive pre-processed roadside data;

[0147] High-value scene fragment data is extracted from roadside data. The high-value scene fragment data includes the driving trajectory information of multiple dynamic traffic participants under different scene labels. The driving trajectory includes the location information and speed information of the trajectory points.

[0148] The 3D static road scene is reconstructed based on vehicle-side perception data, and the multiple dynamic traffic participants are extracted from the vehicle-side perception data.

[0149] Road traffic scene data is generated by combining multiple dynamic traffic participants, reconstructed 3D static road scenes, and high-value scene fragment data.

[0150] For each target dynamic traffic participant in a road traffic scenario, road traffic data from multiple perspectives is obtained by switching the viewpoints of each target dynamic traffic participant.

[0151] After generating road traffic data from multiple vehicle-side perspectives, the model training tool uses the road traffic data and the initial user dataset as input to train the autonomous driving model, improving its generalization ability. The trained autonomous driving model is then rapidly deployed and validated using a model deployment tool to ensure its performance on the vehicle-side hardware. The deployed and validated autonomous driving model is then deployed to the vehicle via OTA (Over-The-Air) updates, which can also be used to upgrade the vehicle-side model version, improving the performance of the autonomous vehicle. When roadside perception devices and data collection vehicles upload new data to the cloud, the cloud can generate new training samples following the same process. These new training samples are then used to train the autonomous driving model, continuously optimizing its performance.

[0152] In this embodiment, the process of identifying high-value scene fragments from roadside data, and the process of generating road traffic data from multiple vehicle perspectives using high-value scene fragments and vehicle-side perception data, can be referred to the description in the above embodiment, and will not be repeated here.

[0153] Example 3

[0154] Figure 3 This is a structural block diagram of an autonomous driving model data processing device based on a vehicle-road-cloud integrated architecture, as provided in Embodiment 3 of the present invention. Figure 3 As shown, the processing device provided in this embodiment includes: a data receiving module 310, a high-value scene fragment data extraction module 320, a static road scene reconstruction module 330, a road traffic scene generation module 340, and a training data generation module 350, wherein...

[0155] The data receiving module 310 is configured to receive roadside data uploaded by roadside sensing devices and vehicle-side sensing data uploaded by data collection vehicles, wherein the roadside data is traffic flow data.

[0156] The high-value scene fragment data extraction module 320 is configured to extract high-value scene fragment data from roadside data. The high-value scene fragment data includes the driving trajectory information of multiple dynamic traffic participants in different scenarios. The driving trajectory includes the location information and speed information of the trajectory points.

[0157] The static road scene reconstruction module 330 is configured to reconstruct a 3D static road scene based on vehicle-side perception data and extract multiple dynamic traffic participants from the vehicle-side perception data.

[0158] The road traffic scene generation module 340 is configured to combine multiple dynamic traffic participants, reconstructed static road scenes, and high-value scene fragment data to generate road traffic scene data.

[0159] The training data generation module 350 is configured to obtain road traffic data from multiple perspectives for each target dynamic traffic participant in the road traffic scenario by switching the perspectives of each target dynamic traffic participant. This road traffic data is used to train the autonomous driving model.

[0160] Optional, the high-value scene fragment data extraction module 320 specifically includes:

[0161] The local map data conversion unit is configured to determine the intersection and road segment corresponding to the roadside data, and convert the high-precision map data corresponding to the intersection and road segment into local map data that meets the training data format requirements;

[0162] The driving trajectory determination unit is configured to extract the driving trajectory information of each dynamic traffic participant within a set time period from the roadside data, and search in the local map according to the starting point location information of each dynamic traffic participant to obtain the candidate driving trajectory of each dynamic traffic participant, and smooth the candidate driving trajectory.

[0163] The feature information determination unit is configured to determine the behavioral feature information of each dynamic traffic participant based on the smoothed candidate driving trajectory and in combination with local map data. The behavioral features include lane affiliation, collision features, distance between trajectory points, distance between target contours, and traffic speed and traffic volume of each lane.

[0164] The target behavior determination unit is configured to determine the target behavior of each dynamic traffic participant based on the behavioral characteristic information of each dynamic traffic participant, add corresponding behavior tags to each target behavior, and store the behavior tags in a scene tag library. The scene tag library also includes multiple types of scene tags, and each type of scene tag includes multiple levels of sub-tags.

[0165] The label combination unit is configured to select target behavior labels and various scene labels that meet the model training requirements from the scene label library, combine the selected labels, and use the combined labels and their corresponding target behavior data as high-value scene fragment data.

[0166] Optional, the local map data transformation unit is specifically configured as follows:

[0167] Convert the high-precision map data corresponding to the intersection and road segment into the geographic data format ShapeFile;

[0168] Road elements are extracted from the map data after map format conversion. The road elements include lane center lines, lane lines, road lines, and intersection areas.

[0169] The extracted road elements are transformed so that the data format of the transformed road elements meets the requirements of the training data. The transformation includes converting the extracted road elements from the geocentric coordinate system WGS84 to the universal transverse Mercator coordinate system UTM, the transformation of the enumeration type, and the data point sparsity processing.

[0170] Based on the road elements after data format conversion and the topological connections between each road element, local map data that meets the requirements of the training data format is obtained.

[0171] Optional, the tag combination unit is specifically configured as follows:

[0172] Select target behavior labels and various scene labels that meet the model training requirements from the scene label library;

[0173] The selected tags are combined according to preset tag combination rules, wherein the preset tag combination rules include any one or more of the following: logically combining different tags according to the relationship between different tags, sorting the tags according to the time order, assigning different weights to different tags, and adding different priorities to different tags;

[0174] The combined labels and their corresponding target behavior data are used as high-value scene fragment data.

[0175] Example 4

[0176] Please see Figure 4 , Figure 4 This is a schematic diagram of a computing device according to Embodiment 4 of the present invention. The computing device is a cloud server. Figure 4 As shown, the computing device may include:

[0177] Memory 701 storing executable program code;

[0178] Processor 702 coupled to memory 701;

[0179] The processor 702 calls the executable program code stored in the memory 701 to execute the autonomous driving model data processing method based on the vehicle-road-cloud integrated architecture provided in any embodiment of the present invention.

[0180] This invention discloses a computer-readable storage medium storing a computer program that enables a computer to execute the method for processing autonomous driving model data based on a vehicle-road-cloud integrated architecture provided in any embodiment of this invention.

[0181] In various embodiments of the present invention, it should be understood that the sequence number of each process does not necessarily imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0182] In the embodiments provided by this invention, it should be understood that "B corresponding to A" means that B is associated with A, and B can be determined based on A. However, it should also be understood that determining B based on A does not mean determining B solely based on A; B can also be determined based on A and / or other information.

[0183] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0184] If the integrated units described above are implemented as software functional units and sold or used as independent products, they can be stored in a computer-accessible memory. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several requests to cause a computer device (which can be a personal computer, server, or network device, specifically a processor in the computer device) to execute some or all of the steps of the methods described in the various embodiments of the present invention.

[0185] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-Erasable Programmable Read-Only Memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium capable of carrying or storing data.

[0186] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of one embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing the present invention.

[0187] Those skilled in the art will understand that the modules in the apparatus of the embodiments can be distributed in the apparatus of the embodiments as described in the embodiments, or they can be located in one or more devices different from this embodiment with corresponding changes. The modules of the above embodiments can be combined into one module, or they can be further divided into multiple sub-modules.

[0188] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for processing autonomous driving model data based on a vehicle-road-cloud integrated architecture, wherein, The vehicle-road-cloud integrated architecture includes data collection vehicles, roadside sensing devices, and a cloud platform. The processing method is applied to the cloud platform, characterized in that the processing method includes: The system receives roadside data uploaded by the roadside sensing device and vehicle-side sensing data uploaded by the data collection vehicle, wherein the roadside data includes traffic flow data. High-value scene fragment data is extracted from the roadside data, wherein the high-value scene fragment data includes the driving trajectory information of multiple dynamic traffic participants in different scenarios, and the driving trajectory information includes the location information and speed information of the trajectory points; The 3D static road scene is reconstructed based on the vehicle-side perception data, and the multiple dynamic traffic participants are identified and extracted from the vehicle-side perception data. The road traffic scene is generated by combining the multiple dynamic traffic participants, the reconstructed 3D static road scene, and the high-value scene fragment data. For each target dynamic traffic participant in the road traffic scenario, road traffic data from multiple perspectives is obtained by switching the viewpoints of each target dynamic traffic participant. The road traffic data is used to train the autonomous driving model. The extraction of high-value scene fragment data from the roadside data includes: Identify the intersections and road segments corresponding to the roadside data, and convert the high-precision map data corresponding to the intersections and road segments into local map data that meets the training data format requirements; The driving trajectory information of each dynamic traffic participant within a set time period is extracted from the roadside data, and the candidate driving trajectory of each dynamic traffic participant is obtained by searching the local map based on the starting point location information of each dynamic traffic participant, and the candidate driving trajectory is smoothed. Based on the smoothed candidate driving trajectories and combined with the local map data, the behavioral characteristics of each dynamic traffic participant are determined. The behavioral characteristics include lane affiliation, collision characteristics, distance between trajectory points, distance between target contours, and traffic speed and volume in each lane. Based on the behavioral characteristic information of each dynamic traffic participant, the target behavior of each dynamic traffic participant is determined, and a corresponding behavior tag is added to each target behavior. The behavior tag is then stored in a scene tag library, which also includes multiple types of scene tags, each type of scene tag including multiple levels of sub-tags. The target behavior labels and various scene labels that meet the model training requirements are selected from the scene label library, and the selected labels are combined. The combined labels and their corresponding target behavior data are used as high-value scene fragment data.

2. The method according to claim 1, characterized in that, The method further includes: Test the trained autonomous driving model and deploy the tested autonomous driving model to intelligent connected vehicles.

3. The method according to claim 1, characterized in that, The step of converting the high-precision map data corresponding to the intersection and road segment into local map data that conforms to the training data format requirements includes: Convert the high-precision map data corresponding to the intersection and road segment into the geographic data format ShapeFile; Road elements are extracted from the map data after map format conversion. The road elements include lane center lines, lane lines, road lines, and intersection areas. The extracted road elements are transformed so that the data format of the transformed road elements meets the requirements of the training data. The transformation includes converting the extracted road elements from the geocentric coordinate system WGS84 to the universal transverse Mercator coordinate system UTM, the transformation of the enumeration type, and the data point sparsity processing. Based on the road elements after data format conversion and the topological connections between each road element, local map data that meets the requirements of the training data format is obtained.

4. The method according to claim 1, characterized in that, The process of combining the selected tags includes: The selected tags are combined according to preset tag combination rules, which include any one or more of the following: logically combining different tags according to the relationship between different tags, sorting the tags according to time order, assigning different weights to different tags, and adding different priorities to different tags.

5. A device for processing autonomous driving model data based on a vehicle-road-cloud integrated architecture, characterized in that, The processing device includes: The data receiving module is configured to receive roadside data uploaded by roadside sensing devices and vehicle-side sensing data uploaded by data collection vehicles, wherein the roadside data is traffic flow data. The high-value scene fragment data extraction module is configured to extract high-value scene fragment data from the roadside data, wherein the high-value scene fragment data includes the driving trajectory information of multiple dynamic traffic participants in different scenarios, and the driving trajectory information includes the location information and speed information of the trajectory points; The static road scene reconstruction module is configured to reconstruct a 3D static road scene based on the vehicle-side perception data, and to identify and extract the multiple dynamic traffic participants from the vehicle-side perception data. The road traffic scene generation module is configured to combine the multiple dynamic traffic participants, the reconstructed 3D static road scene, and the high-value scene fragment data to generate road traffic scene data. The training data generation module is configured to obtain road traffic data from multiple perspectives for each target dynamic traffic participant in the road traffic scenario by switching the perspectives of each target dynamic traffic participant. The road traffic data is used to train the autonomous driving model. The high-value scene fragment data extraction module specifically includes: The local map data conversion unit is configured to determine the intersection and road segment corresponding to the roadside data, and convert the high-precision map data corresponding to the intersection and road segment into local map data that meets the training data format requirements; The driving trajectory determination unit is configured to extract driving trajectory information of each dynamic traffic participant within a set time period from the roadside data, and search in the local map according to the starting point location information of each dynamic traffic participant to obtain candidate driving trajectories of each dynamic traffic participant, and smooth the candidate driving trajectories. The feature information determination unit is configured to determine the behavioral feature information of each dynamic traffic participant based on the smoothed candidate driving trajectory and in combination with the local map data. The behavioral features include lane affiliation, collision features, distance between trajectory points, distance between target contours, and traffic speed and traffic volume of each lane. The target behavior determination unit is configured to determine the target behavior of each dynamic traffic participant based on the behavioral characteristic information of each dynamic traffic participant, add corresponding behavior tags to each target behavior, and store the behavior tags in a scene tag library. The scene tag library also includes multiple types of scene tags, and each type of scene tag includes multiple levels of sub-tags. The tag combination unit is configured to select target behavior tags and various scene tags that meet the model training requirements from the scene tag library, combine the selected tags, and use the combined tags and their corresponding target behavior data as high-value scene fragment data.

6. A self-evolutionary learning system for autonomous driving models based on a vehicle-road-cloud integrated architecture, characterized in that, The learning system includes: The data collection vehicle is used to collect vehicle-side perception data and upload the vehicle-side perception data to the cloud control platform. The vehicle-side perception data includes surrounding environment information and traffic participant status information. Roadside sensing devices are used to collect roadside data and upload the roadside data to the cloud control platform. The roadside data includes traffic flow data. The cloud control platform is used to preprocess the vehicle-side perception data and the roadside data, and upload the preprocessed vehicle-side perception data and the preprocessed roadside data to the cloud. The cloud is used for: Receive the preprocessed vehicle-side perception data, and receive the preprocessed roadside data; High-value scene fragment data is extracted from the roadside data, wherein the high-value scene fragment data includes the driving trajectory information of multiple dynamic traffic participants in different scenarios, and the driving trajectory information includes the location information and speed information of the trajectory points; The 3D static road scene is reconstructed based on the vehicle-side perception data, and the multiple dynamic traffic participants are identified and extracted from the vehicle-side perception data. The multiple dynamic traffic participants, the reconstructed 3D static road scene, and the high-value scene fragment data are combined to generate road traffic scene data; For each target dynamic traffic participant in the road traffic scenario, road traffic data from multiple perspectives is obtained by switching the viewpoints of each target dynamic traffic participant, and the autonomous driving model is trained based on the road traffic data. Test the trained autonomous driving model and deploy the tested autonomous driving model to intelligent connected vehicles; The extraction of high-value scene fragment data from the roadside data includes: Identify the intersections and road segments corresponding to the roadside data, and convert the high-precision map data corresponding to the intersections and road segments into local map data that meets the training data format requirements; The driving trajectory information of each dynamic traffic participant within a set time period is extracted from the roadside data, and the candidate driving trajectory of each dynamic traffic participant is obtained by searching the local map based on the starting point location information of each dynamic traffic participant, and the candidate driving trajectory is smoothed. Based on the smoothed candidate driving trajectories and combined with the local map data, the behavioral characteristics of each dynamic traffic participant are determined. The behavioral characteristics include lane affiliation, collision characteristics, distance between trajectory points, distance between target contours, and traffic speed and volume in each lane. Based on the behavioral characteristic information of each dynamic traffic participant, the target behavior of each dynamic traffic participant is determined, and a corresponding behavior tag is added to each target behavior. The behavior tag is then stored in a scene tag library, which also includes multiple types of scene tags, each type of scene tag including multiple levels of sub-tags. The target behavior labels and various scene labels that meet the model training requirements are selected from the scene label library, and the selected labels are combined. The combined labels and their corresponding target behavior data are used as high-value scene fragment data.

7. A computing device, characterized in that, The computing device includes: One or more processors; Storage device for storing one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement the method for processing autonomous driving model data based on the vehicle-road-cloud integrated architecture as described in any one of claims 1-4.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the method for processing autonomous driving model data based on the vehicle-road-cloud integrated architecture as described in any one of claims 1-4.

Images