Automatic driving soft recharge method and system and readable storage medium

Abstract

The invention provides an automatic driving soft recharge method and system and a readable storage medium, and the method comprises the following steps: collecting real vehicle road test data, uploading the real vehicle road test data to a cloud object storage service for classified storage, and forming a cloud data resource pool; preprocessing the real vehicle road test data in the cloud data resource pool, extracting scene features and automatically labeling the scene features, generating structured scene data and storing the structured scene data in a generalization scene library; s3, selecting target scene data from the generalization scene library, scheduling and operating a to-be-tested automatic driving algorithm, recharging the target scene data into the to-be-tested automatic driving algorithm for testing, and automatically generating a test report containing problem data identifiers and performance indexes; judging whether the performance index in the test report meets a preset test passing condition or not; and if yes, outputting a final test report, and completing the soft recharge test.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving technology, and in particular to an autonomous driving soft reflow method, system, and readable storage medium. Background Technology

[0002] The development of autonomous driving algorithms requires extensive real-world data validation. Traditional soft-backtesting primarily utilizes real-vehicle road test Logsim data and simulation data. Real-vehicle road test Logsim data collection suffers from high costs, long cycles, and incomplete scenario coverage. Simulation data uses virtual scenarios generated by simulation engines, which differ significantly from real-vehicle data, leading to inconsistencies between test results and actual vehicle performance. Regarding data storage and retrieval, real-vehicle Logsim data (sensor, bus, and vehicle status data) is often stored on local hard drives or scattered in the cloud without effective classification and management, resulting in fragmented data storage and low efficiency in data retrieval and reuse. Data transmission between the cloud and local storage typically has high latency, and without effective network optimization, Logsim data is large, often measured in GB. The ability to quickly and stably download and upload data packets directly impacts testing efficiency. In terms of algorithm service deployment, traditional methods require manual installation of dependencies. Different versions of dependencies vary significantly, leading to inconsistencies between the development and testing environments and unreliable test results. Summary of the Invention

[0003] This invention provides a method, system, and readable storage medium for soft re-implantation of autonomous driving, which can at least solve the technical problems of insufficient realism, insufficient scenario coverage, low re-implantation efficiency, and cumbersome deployment of algorithm services in existing soft re-implantation tests.

[0004] The technical solution of the present invention is as follows: On the one hand, a method for soft re-feedback of autonomous driving is provided, including the following steps: S1. Collect real vehicle road test data and upload the real vehicle road test data to the cloud object storage service for classification and storage to form a cloud data resource pool; S2. Preprocess the real vehicle road test data in the cloud data resource pool, extract scene features and automatically label them, generate structured scene data and store it in the generalized scene library. S3. Select target scenario data from the generalized scenario library, schedule and run the autonomous driving algorithm to be tested, feed the target scenario data back into the autonomous driving algorithm to be tested for testing, and automatically generate a test report containing problem data identifiers and performance indicators. S4. Determine whether the performance indicators in the test report meet the preset test pass conditions; If the conditions are met, the final test report will be output, and the soft reflow test will be completed. If the conditions are not met, the problem data is analyzed based on the problem data identifier in the test report to locate the defects of the autonomous driving algorithm under test and optimize it. Subsequently, based on the optimized autonomous driving algorithm, a new round of soft backfeed test is started using the target scene data or a dataset containing the scene corresponding to the problem data identifier, and the process returns to step S3 to schedule and run the optimized algorithm and perform backfeed test.

[0005] In one optional implementation, step S1, the acquisition of real vehicle road test data includes: acquiring raw perception data through on-board sensors and acquiring vehicle status data through the vehicle bus interface, wherein the on-board sensors include at least one of a camera, a lidar, and a millimeter-wave radar.

[0006] In one optional implementation, step S2, the preprocessing of the road test data includes at least one of the following operations: outlier removal, missing value completion, timestamp alignment, and data validity verification.

[0007] In one optional implementation, before step S3, step S0, algorithm containerization packaging and management, is also included: the autonomous driving algorithm to be tested and its runtime dependent environment are packaged into a container image and uploaded to a container image repository for version management. In step S3, the scheduling and running of the autonomous driving algorithm to be tested specifically involves: pulling the corresponding container image from the container image repository and scheduling and running the corresponding containerized algorithm instance through the container orchestration platform.

[0008] In one optional implementation, the container orchestration platform is a Kubernetes cluster; in step S3, the containerized algorithm instance is created and managed in the form of Pod through the Kubernetes cluster, and computing resources are dynamically scheduled according to the test load.

[0009] In one optional implementation, step S2, extracting scene features and automatically labeling them, includes using a multimodal Transformer model for automatic labeling; the labels for the scene feature annotations include at least one of environmental labels, traffic participant labels, road structure labels, and functional labels.

[0010] In an optional implementation, step S2 further includes: based on the features of the scene data in the generalized scene library, using a generative adversarial network (GAN) to generalize the scene and generate derived scene data to expand the generalized scene library.

[0011] In an optional implementation, step S3, the automatic generation of a test report containing problem data identifiers, includes: Record the output results of the autonomous driving algorithm under test on the target scene data; The output results are compared with the preset expected results to mark the problematic data, and one or more preset performance indicators are calculated based on the output results.

[0012] In one optional implementation, in step S4, the test pass condition is: the performance index reaches a preset threshold, and / or the number or proportion of the problematic data is lower than a preset threshold.

[0013] In one alternative implementation, step S4 includes: By analyzing the original scene data and algorithm output results corresponding to the problematic data identifiers using data visualization tools, algorithm defects can be identified. Adjust algorithm parameters, modify model structure, or update model weights based on the positioning results; and generate updated container images based on the optimized algorithm for a new round of testing.

[0014] On the other hand, an autonomous driving soft recharge system is provided, including a data acquisition module for uploading the real vehicle road test data to a cloud object storage service for classified storage, forming a cloud data resource pool; The data preprocessing module is used to preprocess the real vehicle road test data in the cloud data resource pool, extract scene features and automatically label them, generate structured scene data and store it in the generalized scene library. The soft-feedback test task execution module is used to select target scenario data from the generalized scenario library, schedule and run the autonomous driving algorithm to be tested, feed the target scenario data back into the autonomous driving algorithm to be tested for testing, and automatically generate a test report containing problem data identifiers and performance indicators. A closed-loop optimization module is used to determine whether the performance indicators in the test report meet the preset test pass conditions; If the conditions are met, the final test report will be output, and the soft reflow test will be completed. If the conditions are not met, the problem data is analyzed based on the problem data identifier in the test report to locate the defects of the autonomous driving algorithm under test and optimize it. Subsequently, based on the optimized autonomous driving algorithm, a new round of soft backfeed test is started using the target scene data or a dataset containing the scene corresponding to the problem data identifier, and the process returns to step S3 to schedule and run the optimized algorithm and perform backfeed test.

[0015] In another aspect, a readable storage medium is provided, storing the autonomous driving soft reflow method described in any of the above claims.

[0016] The method provided in this embodiment of the invention has at least the following beneficial effects: The method provided in this invention can reproduce real-vehicle driving scenarios and data through data refeeding, reducing the frequency of real-vehicle testing, lowering testing costs, and avoiding the problem of insufficient data authenticity. It simplifies the real-vehicle road test data collection process, custom-develops data collection tools, unifies data formats, and improves data parsing speed during data preprocessing and refeeding. It selects cloud object storage services to categorize and store refeeding data, optimizes network configuration to make upload and download speeds closer to bandwidth limits, reduces data transmission failures, provides basic guarantees for local / cloud collaborative storage, and improves data retrieval and reuse efficiency. It preprocesses road test data, verifies data validity, and reduces invalid data refeeding. It extracts scenario features from the refeeding data, automatically labels them, and stores them in a generalized scenario library for subsequent scenario generalization. Attached Figure Description

[0017] The accompanying drawings, which form part of this specification, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a schematic diagram of the soft reflow method for autonomous driving.

[0018] Figure 2 This is a schematic diagram illustrating the detailed process of the soft reflow method for autonomous driving.

[0019] Figure 3 This is a schematic diagram of the system framework for the soft recharge method for autonomous driving. Detailed Implementation

[0020] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0021] The following detailed description is exemplary and intended to provide further detailed explanation of the invention. Unless otherwise specified, all technical terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this invention is for describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention.

[0022] As is known from common technical knowledge, this invention can be implemented through other embodiments that do not depart from its spirit or essential characteristics. Therefore, the disclosed embodiments described above are merely illustrative in all respects and are not the only ones. All modifications within the scope of this invention or its equivalents are included in this invention.

[0023] Please see Figure 1 and Figure 2On the one hand, an autonomous driving soft re-feedback method is provided, including the following steps: S1. Collect real vehicle road test data and upload the real vehicle road test data to the cloud object storage service for classification and storage, forming a cloud data resource pool; S2. Preprocess the real vehicle road test data in the cloud data resource pool, extract scene features and automatically label them, generate structured scene data and store it in the generalized scene library; S3. Select target scenario data from the generalized scenario library, schedule and run the autonomous driving algorithm to be tested, feed the target scenario data back into the autonomous driving algorithm to be tested for testing, and automatically generate a test report containing problem data identifiers and performance indicators. S4. Determine whether the performance indicators in the test report meet the preset test pass conditions; If the conditions are met, the final test report will be output, and the soft reflow test will be completed. If the conditions are not met, the problem data is analyzed based on the problem data identifiers in the test report to locate the defects of the autonomous driving algorithm under test and optimize it. Subsequently, based on the optimized autonomous driving algorithm, a new round of soft backfeeding test is started using target scene data or a dataset containing the scene corresponding to the problem data identifiers, and the process returns to step S3 to schedule and run the optimized algorithm and perform backfeeding test.

[0024] The method provided in this embodiment of the invention has at least the following beneficial effects: The method provided in this invention can reproduce real-vehicle driving scenarios and data through data refeeding, reducing the frequency of real-vehicle testing, lowering testing costs, and avoiding the problem of insufficient data authenticity. It simplifies the real-vehicle road test data collection process, custom-develops data collection tools, unifies data formats, and improves data parsing speed during data preprocessing and refeeding. It selects cloud object storage services to categorize and store refeeding data, optimizes network configuration to make upload and download speeds closer to bandwidth limits, reduces data transmission failures, provides basic guarantees for local / cloud collaborative storage, and improves data retrieval and reuse efficiency. It preprocesses road test data, verifies data validity, and reduces invalid data refeeding. It extracts scenario features from the refeeding data, automatically labels them, and stores them in a generalized scenario library for subsequent scenario generalization.

[0025] Furthermore, this invention employs real data to construct a scenario library, avoiding the inaccuracies of simulation data; cloud-based automated testing replaces extensive real-vehicle road testing, resulting in lower costs, shorter cycles, and parallel processing capabilities. Through automatic annotation and scenario generalization, massive test cases are intelligently generated and managed, providing more comprehensive coverage. Containerization technology ensures that the algorithm operates in a completely consistent environment anywhere, guaranteeing reliable results. An automated testing, diagnosis, and verification closed loop makes the algorithm optimization process data-driven and goal-oriented, significantly improving R&D efficiency and quality.

[0026] In step S1, real vehicle road test data is collected and uploaded to the cloud object storage service for classification and storage, forming a cloud data resource pool.

[0027] In one optional implementation, step S1, collecting real vehicle road test data includes: collecting raw perception data through on-board sensors and collecting vehicle status data through the vehicle bus interface. The on-board sensors include at least one of cameras, lidar, and millimeter-wave radar.

[0028] When test vehicles equipped with onboard sensors are driving on the road, their cameras, radar, and other devices continuously record the surrounding environment (images, point clouds) and the vehicle's own status (speed, steering wheel angle, etc.), much like a dashcam. This massive amount of raw data is initially stored temporarily on the vehicle's hard drive. Upon returning to the base, these data packets are automatically uploaded to the cloud (such as a cloud storage service like Baidu Cloud) via a high-speed network. The cloud system automatically categorizes and stores the data according to its "meta-information" (such as shooting time, vehicle model, and weather), forming a vast and organized cloud data resource pool.

[0029] For example, a test vehicle drove on Shanghai's Middle Ring Road for one hour on a rainy night. All the video recordings and vehicle data generated were tagged as [Vehicle type: SUV_A, Time: 20231001, Location: Shanghai, Weather: Rainy night] and then uploaded to the corresponding folder in the cloud.

[0030] The above method solves the problems of scattered and chaotic data storage and management in traditional road test systems. Centralizing the data in the cloud lays the foundation for unified access and efficient retrieval in the future.

[0031] In step S2, the real vehicle road test data in the cloud data resource pool is preprocessed, scene features are extracted and automatically labeled, structured scene data is generated and stored in the generalized scene library.

[0032] In one optional implementation, step S2, preprocessing the road test data includes at least one of the following operations: outlier removal, missing value completion, timestamp alignment, and data validity verification.

[0033] Understandably, cloud-based data is raw and messy. This invention first "cleans" the data, for example, by repairing erroneous data frames caused by signal interference and ensuring timeline alignment. Then, a powerful AI model (multimodal Transformer) is used to automatically "view" this data and understand what happens in each video segment, such as "It's a rainy evening, there's a white car driving at a constant speed 20 meters ahead, and we're crossing an intersection." The autonomous driving soft-feedback system of this invention transforms these understanding results (scene features) into standardized labels. Finally, the "raw video clips" and corresponding "scene description labels" are packaged and stored in a generalized scene library.

[0034] For example, from one hour of driving data, the system automatically identifies 30 typical segments such as "turning left at an intersection", "pedestrian crossing", and "congested following", and assigns detailed semantic tags to each segment, making it easy to accurately search for them in the future using conditions such as "rainy day + pedestrian crossing".

[0035] As another example, in real-world road tests, a forward-facing camera captures 1920x1080 resolution RGB images at 30 frames per second (raw perception data), a roof-mounted LiDAR captures 128-line point cloud data at 10 frames per second (raw perception data), and a CAN FD bus collects messages containing information such as vehicle speed, yaw rate, steering wheel angle, and accelerator pedal opening at a frequency of 100Hz (vehicle status data). These multi-source heterogeneous data collectively constitute the input for subsequent data reinjection.

[0036] Furthermore, as an example, outlier removal can be achieved by inspecting the LiDAR point cloud and marking all points with distance values ​​greater than 200 meters or less than 0 meters (physically impossible values) as invalid and filtering them out. Missing value completion can be achieved by linear interpolation of preceding and following valid frame data to fill in missing vehicle speed data caused by occasional packet loss in the CAN bus signal. Timestamp alignment can be achieved by synchronizing the hardware timestamp of the camera exposure center, the timestamp of the start of the LiDAR scanning cycle, and the software timestamp of the CAN message reception onto a global nanosecond-level timeline based on the PTP protocol. Validity verification can be achieved by checking a continuous image frame sequence; if a complete data loss or screen distortion is found due to storage failure, all data within that time period is marked as invalid and not proceeds to subsequent processes.

[0037] The above method solves the problems of insufficient realism in test scenarios and high cost and incomplete coverage of manual annotation. Through automated processing, standardized, reusable, high-fidelity test scenarios are extracted from real data.

[0038] In step S3, target scenario data is selected from the generalized scenario library, the autonomous driving algorithm to be tested is scheduled and run, the target scenario data is fed back into the autonomous driving algorithm to be tested for testing, and a test report containing problem data identifiers and performance indicators is automatically generated.

[0039] As an example, when testing a new version of an autonomous driving algorithm, engineers select a batch of target scenario data from a generalized scenario library. The system then runs the algorithm under test on a cloud-based computing cluster. Next, the system feeds the selected scenario data (such as a 5-second video stream of a rainy intersection) into the instance of the autonomous driving algorithm under test in real time. The autonomous driving algorithm reacts to the video (e.g., identifying vehicles, planning routes). The system records the results of the autonomous driving algorithm under test throughout the process and compares them with the standard answer (preset expectations, such as the actual location of vehicles in the video), automatically grading and generating a detailed score report (test report), which indicates all incorrectly answered questions (problem data) and the total score (performance metrics).

[0040] For example, a scenario of "pedestrians crossing an intersection on a rainy night" was selected to test the perception capabilities of the autonomous driving algorithm under test. After feedback, the test report showed that the autonomous driving algorithm under test did not identify problematic data in the first 3 frames before the pedestrian appeared, and the overall pedestrian detection accuracy rate was 92% (performance metric).

[0041] The above steps address the pain points of traditional testing, such as low efficiency and inconsistent environments. They enable highly efficient testing that is automated, batch-processed, and environmentally standardized, eliminating the need for on-road testing and significantly shortening the testing cycle.

[0042] In one optional implementation, before step S3, step S0, algorithm containerization packaging and management, is also included: the autonomous driving algorithm to be tested and its running dependent environment are packaged into a container image and uploaded to the container image repository for version management. In step S3, scheduling and running the autonomous driving algorithm to be tested specifically involves: pulling the corresponding container image from the container image repository and scheduling and running the corresponding containerized algorithm instance through the container orchestration platform.

[0043] In one optional implementation, the container orchestration platform is a Kubernetes cluster; in step S3, containerized algorithm instances are created and managed in the form of Pods through the Kubernetes cluster, and computing resources are dynamically scheduled according to the test load.

[0044] Kubernetes is the most popular platform for managing software containers. It can automatically schedule computing resources and flexibly create and destroy algorithm instances.

[0045] In one optional implementation, in step S2, scene features are extracted and automatically labeled, including automatic labeling using a multimodal Transformer model; the labels for scene feature labeling include at least one of environmental labels, traffic participant labels, road structure labels, and functional labels.

[0046] This embodiment achieves a deep, semantic understanding of the scene, rather than just simple classification. For example, it can not only identify "there is a vehicle," but also understand it as "a bus that is cutting into this lane," providing a foundation for building more intelligent scene retrieval and generalization.

[0047] In an optional implementation, step S2 further includes: generalizing the scene using a generative adversarial network (GAN) based on the features of the scene data in the generalized scene library, and generating derived scene data to expand the generalized scene library.

[0048] This embodiment utilizes AI generation techniques such as GANs to create a large number of similar but reasonably varied new scenarios using existing real-world scenes as "seeds." This greatly expands the testing boundaries and uncovers rare problems. For example, based on a real-world scenario of "pedestrians crossing the street during the day," it can generalize to generate derivative scenarios such as "pedestrians crossing the street with umbrellas in a rainstorm" and "pedestrians crossing the street against the light," testing the robustness of the algorithm.

[0049] In one optional implementation, step S3, automatically generating a test report containing problem data identifiers, includes: Record the output results of the autonomous driving algorithm under test on the target scene data; The output results are compared with the preset expected results to mark the problematic data, and one or more preset performance indicators are calculated based on the output results.

[0050] As an example, in the refeedback test, the autonomous driving algorithm under test output three vehicle bounding boxes for a certain frame of image. The autonomous driving soft refeedback system compares this result with four real vehicle bounding boxes (preset expected result) in the same frame of image, using the IoU (Intersection over Union) metric for matching. Unmatched ground truth bounding boxes with an IoU less than 0.5 are marked as "missed detections" (problem data); algorithm output boxes without corresponding ground truth bounding boxes are marked as "false detections" (problem data). Finally, the average accuracy (mAP) of the entire test set is calculated as the core performance indicator.

[0051] In one optional implementation, in step S4, the test pass condition is: the performance index reaches a preset threshold, and / or the number or proportion of problematic data is lower than a preset threshold.

[0052] As an example, for the perception algorithm, the preset test pass conditions are: on the "urban road" scene subset, the mAP of vehicle detection is not less than 0.98, and the proportion of missed frames to total frames is less than 0.1%. Only when the indicators in the test report meet both of these threshold conditions is it judged as "satisfied".

[0053] The above methods enable the "test-optimization" closed loop to have a clear automated exit point, transforming R&D goals into measurable technical indicators, ensuring the directionality of algorithm iteration and the lower limit of the quality of the final deliverables.

[0054] In one alternative implementation, step S4 includes: By analyzing the original scene data and algorithm output results corresponding to the problem data identifiers using data visualization tools, algorithm defects can be located. Adjust algorithm parameters, modify model structure, or update model weights based on the positioning results; and generate updated container images based on the optimized algorithm for a new round of testing.

[0055] This embodiment describes the specific operation chain from problem discovery to completing an algorithm iteration. For example: Analysis and localization: Engineers use a web visualization tool to view a frame of data marked as "missed". The tool displays the original image of the frame and the heatmap of the algorithm's output side by side, revealing that the algorithm's response to partially occluded pedestrians in the distance is weak. Optimization implementation: Based on this, engineers decide to increase the model's training weights on difficult samples such as "occluded pedestrians" and adjust the threshold parameters of the non-maximum suppression (NMS) algorithm. Update and validation: The model is retrained using the modified code and training strategy, generating a new model weight file. Subsequently, the CI / CD pipeline automatically packages the new algorithm into a perception-fusion:v2.2 image, pushes it to the repository, and triggers a new round of testing tasks.

[0056] Please see Figure 3 On the other hand, an autonomous driving soft recharge system is provided, including a data acquisition module for uploading the real vehicle road test data to a cloud object storage service for classified storage, forming a cloud data resource pool; The data preprocessing module is used to preprocess the real vehicle road test data in the cloud data resource pool, extract scene features and automatically label them, generate structured scene data and store it in the generalized scene library. The soft-feedback test task execution module is used to select target scenario data from the generalized scenario library, schedule and run the autonomous driving algorithm to be tested, feed the target scenario data back into the autonomous driving algorithm to be tested for testing, and automatically generate a test report containing problem data identifiers and performance indicators. A closed-loop optimization module is used to determine whether the performance indicators in the test report meet the preset test pass conditions; If the conditions are met, the final test report will be output, and the soft reflow test will be completed. If the conditions are not met, the problem data is analyzed based on the problem data identifier in the test report to locate the defects of the autonomous driving algorithm under test and optimize it. Subsequently, based on the optimized autonomous driving algorithm, a new round of soft backfeed test is started using the target scene data or a dataset containing the scene corresponding to the problem data identifier, and the process returns to step S3 to schedule and run the optimized algorithm and perform backfeed test.

[0057] In one optional implementation, an algorithm image management module is also included for containerized packaging and management of algorithms: the autonomous driving algorithm to be tested and its runtime dependent environment are packaged into a container image and uploaded to a container image repository for version management.

[0058] On the other hand, a readable storage medium is provided that stores the autonomous driving soft reflow method described above.

[0059] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied 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.

[0060] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0061] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0062] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0063] 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 it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for soft re-implantation of autonomous driving technology, characterized in that, Includes the following steps: S1. Collect real vehicle road test data and upload the real vehicle road test data to the cloud object storage service for classification and storage to form a cloud data resource pool; S2. Preprocess the real vehicle road test data in the cloud data resource pool, extract scene features and automatically label them, generate structured scene data and store it in the generalized scene library. S3. Select target scenario data from the generalized scenario library, schedule and run the autonomous driving algorithm to be tested, feed the target scenario data back into the autonomous driving algorithm to be tested for testing, and automatically generate a test report containing problem data identifiers and performance indicators. S4. Determine whether the performance indicators in the test report meet the preset test pass conditions; If the conditions are met, the final test report will be output, and the soft reflow test will be completed. If the conditions are not met, the problem data is analyzed based on the problem data identifiers in the test report to locate the defects of the autonomous driving algorithm under test and optimize it. Subsequently, based on the optimized autonomous driving algorithm, using the target scene data or a dataset containing the problem data identifier corresponding to the scene, a new round of soft re-feedback testing is initiated, and the process returns to step S3, which involves scheduling and running the optimized algorithm and performing the re-feedback test.

2. The method according to claim 1, characterized in that, In step S1, the collection of real vehicle road test data includes: collecting raw perception data through on-board sensors and collecting vehicle status data through the vehicle bus interface. The on-board sensors include at least one of cameras, lidar, and millimeter-wave radar.

3. The method according to claim 1, characterized in that, In step S2, the preprocessing of the road test data includes at least one of the following operations: outlier removal, missing value completion, timestamp alignment, and data validity verification.

4. The method according to claim 1, characterized in that, Before step S3, there is also step S0, algorithm containerization packaging and management: the autonomous driving algorithm to be tested and its running dependent environment are packaged into a container image and uploaded to the container image repository for version management; In step S3, the scheduling and running of the autonomous driving algorithm to be tested specifically involves: pulling the corresponding container image from the container image repository and scheduling and running the corresponding containerized algorithm instance through the container orchestration platform.

5. The method according to claim 4, characterized in that, The container orchestration platform is a Kubernetes cluster; in step S3, the containerized algorithm instance is created and managed in the form of Pod through the Kubernetes cluster, and computing resources are dynamically scheduled according to the test load.

6. The method according to claim 1 or 5, characterized in that, Step S2 further includes: based on the features of the scene data in the generalized scene library, using a generative adversarial network (GAN) to generalize the scene and generate derived scene data to expand the generalized scene library.

7. The method according to claim 1, characterized in that, In step S3, the automatic generation of a test report containing problem data identifiers includes: Record the output results of the autonomous driving algorithm under test on the target scene data; The output results are compared with the preset expected results to mark the problematic data, and one or more preset performance indicators are calculated based on the output results.

8. The method according to claim 1, characterized in that, In step S4, the test pass condition is: the performance index reaches a preset threshold, and / or the number or proportion of the problem data is lower than a preset threshold; Step S4 includes: By analyzing the original scene data and algorithm output results corresponding to the problematic data identifiers using data visualization tools, algorithm defects can be identified. Adjust algorithm parameters, modify model structure, or update model weights based on the positioning results; and generate updated container images based on the optimized algorithm for a new round of testing.

9. An automated driving soft-recirculation system, characterized in that, include: The data acquisition module is used to upload the real vehicle road test data to the cloud object storage service for classification and storage, forming a cloud data resource pool; The data preprocessing module is used to preprocess the real vehicle road test data in the cloud data resource pool, extract scene features and automatically label them, generate structured scene data and store it in the generalized scene library. The soft-feedback test task execution module is used to select target scenario data from the generalized scenario library, schedule and run the autonomous driving algorithm to be tested, feed the target scenario data back into the autonomous driving algorithm to be tested for testing, and automatically generate a test report containing problem data identifiers and performance indicators. A closed-loop optimization module is used to determine whether the performance indicators in the test report meet the preset test pass conditions; If the conditions are met, the final test report will be output, and the soft reflow test will be completed. If the conditions are not met, the problem data is analyzed based on the problem data identifiers in the test report to locate the defects of the autonomous driving algorithm under test and optimize it. Subsequently, based on the optimized autonomous driving algorithm, using the target scene data or a dataset containing the problem data identifier corresponding to the scene, a new round of soft re-feedback testing is initiated, and the process returns to step S3, which involves scheduling and running the optimized algorithm and performing the re-feedback test.

10. A readable storage medium, characterized in that, The device stores the automatic driving soft recharge method according to any one of claims 1-8.

Images