CN-121979712-A - Automatic driving soft recharge method, system and readable storage medium

CN121979712ACN 121979712 ACN121979712 ACN 121979712ACN-121979712-A

Abstract

The invention provides an automatic driving soft recharging method, an automatic driving soft recharging system and a readable storage medium, which comprise the following steps of collecting real vehicle drive test data, uploading the real vehicle drive test data to a cloud object storage service for classified storage, and forming a cloud data resource pool; the method comprises the steps of preprocessing real vehicle drive test data in a cloud data resource pool, extracting scene characteristics, automatically marking, generating structured scene data and storing the structured scene data into a generalized scene library, S3, selecting target scene data from the generalized scene library, scheduling and running an automatic driving algorithm to be tested, recharging the target scene data into the automatic driving algorithm to be tested for testing, automatically generating a test report containing problem data identification and performance indexes, judging whether the performance indexes in the test report meet preset test passing conditions, and outputting a final test report if the performance indexes meet preset test passing conditions, so that soft recharging test is completed.

Inventors

Gao Shangyuan
JI QINGHUI
GUO YANLING

Assignees

奇瑞汽车股份有限公司

Dates

Publication Date: 20260505
Application Date: 20260113

Claims (10)

1. An automatic driving soft recharge method, which is characterized by comprising the following steps: s1, acquiring real vehicle drive test data, uploading the real vehicle drive test data to a cloud object storage service for classified storage, and forming a cloud data resource pool; S2, preprocessing the real vehicle drive test data in the cloud data resource pool, extracting scene characteristics, automatically marking, generating structured scene data and storing the structured scene data into a generalized scene library; s3, selecting target scene data from the generalized scene library, scheduling and running an automatic driving algorithm to be tested, recharging the target scene data into the automatic driving algorithm to be tested for testing, and automatically generating a test report containing a problem data identifier and a performance index; S4, judging whether the performance index in the test report meets a preset test passing condition or not; If yes, outputting a final test report to finish the soft recharging test; And then, starting a new round of soft recharging test by using the target scene data or a data set containing the scene identification corresponding to the problem data based on the optimized automatic driving algorithm, and returning to the step S3 of scheduling and running the optimized algorithm and carrying out the recharging test.
2. The method of claim 1, wherein in step S1, the acquiring real-vehicle drive test data comprises acquiring raw perception data through an onboard sensor, and acquiring vehicle state data through a vehicle bus interface, the onboard sensor comprising at least one of a camera, a laser radar, and a millimeter wave radar.
3. The method of claim 1, wherein the preprocessing of the drive test data in step S2 comprises at least one of outlier removal, missing value completion, time stamp alignment, and data validity checking of the drive test data.
4. The method according to claim 1, further comprising the steps of S0, packaging and managing the automatic driving algorithm to be tested and the running dependent environment thereof as a container mirror image, and uploading the container mirror image to a container warehouse for version management; In step S3, the automatic driving algorithm to be tested is specifically that a corresponding container mirror image is pulled from the container mirror image warehouse, and a corresponding containerized algorithm instance is scheduled and operated through a container scheduling platform.
5. The method according to claim 4, wherein the container arrangement platform is a Kubernetes cluster, and wherein in step S3, the containerized algorithm instance is created and managed in Pod form by the Kubernetes cluster, and the computing resources are dynamically scheduled according to the test load.
6. The method according to claim 1 or 5, wherein step S2 further comprises generating derivative scene data to augment the generalized scene library by generating a countermeasure network GAN for scene generalization based on features of scene data in the generalized scene library.
7. The method of claim 1, wherein in step S3, the automatically generating a test report containing an identification of problem data comprises: Recording the output result of the automatic driving algorithm to be tested on the target scene data; comparing the output result with a preset expected result to mark problem data, and calculating one or more preset performance indexes according to the output result.
8. The method according to claim 1, wherein in step S4, the test passing condition is that the performance index reaches a preset threshold value and/or the number or the duty ratio of the problem data is lower than a preset threshold value; The step S4 includes: analyzing original scene data corresponding to the problem data identification and an algorithm output result through a data visualization tool, and positioning algorithm defects; And generating an updated container mirror image for a new round of testing based on the optimized algorithm.
9. An autopilot soft recharge system, comprising: The data acquisition module is used for uploading the real vehicle drive test data to a cloud object storage service for classified storage to form a cloud data resource pool; The data preprocessing module is used for preprocessing the real vehicle drive test data in the cloud data resource pool, extracting scene characteristics and automatically marking, generating structured scene data and storing the structured scene data into the generalized scene library; The soft recharging test task execution module is used for selecting target scene data from the generalized scene library, scheduling and running an automatic driving algorithm to be tested, recharging the target scene data into the automatic driving algorithm to be tested for testing, and automatically generating a test report containing a problem data identifier and a performance index; the closed loop optimization module is used for judging whether the performance index in the test report meets the preset test passing condition or not; If yes, outputting a final test report to finish the soft recharging test; And then, starting a new round of soft recharging test by using the target scene data or a data set containing the scene identification corresponding to the problem data based on the optimized automatic driving algorithm, and returning to the step S3 of scheduling and running the optimized algorithm and carrying out the recharging test.
10. A readable storage medium, characterized in that the autopilot soft recharge method of any one of claims 1-8 is stored.

Description

Automatic driving soft recharge method, system and readable storage medium Technical Field The invention relates to the technical field of automatic driving, in particular to an automatic driving soft recharging method, an automatic driving soft recharging system and a readable storage medium. Background The development of the automatic driving algorithm needs a large amount of real scene data verification, and the traditional soft recharging mostly adopts real vehicle drive test Logsim data and simulation data. The real vehicle road test data acquisition Logsim has the problems of high cost, long period, incomplete scene coverage and the like, and the virtual scene generated by using the simulation engine in the simulation data has large difference with the real vehicle data, so that the test result is inconsistent with the real vehicle performance. In terms of data storage and reading, the real vehicle Logsim data (sensor, bus and vehicle state data) are stored in a local mobile hard disk or scattered cloud, and effective classified management is not performed, so that the data storage is scattered, and the data retrieval and multiplexing efficiency is low. The cloud and local data transmission is usually delayed higher, effective network optimization is not configured, logsim data is large in size, and whether the data packet can be downloaded and uploaded quickly and stably usually takes GB as a unit directly influences the test efficiency. In the aspect of deployment of algorithm service, the traditional method needs to manually install the dependent environment, and the dependent environments of different versions have large differences, so that the development environment is inconsistent with the testing environment, and the testing result is unreliable. Disclosure of Invention The invention provides an automatic driving soft recharging method, an automatic driving soft recharging system and a readable storage medium, which at least can solve the technical problems of insufficient authenticity, insufficient scene coverage, low recharging efficiency, complicated algorithm service deployment and the like of the existing soft recharging test. The technical scheme of the invention is as follows: In one aspect, an automatic driving soft recharge method is provided, comprising the steps of: s1, acquiring real vehicle drive test data, uploading the real vehicle drive test data to a cloud object storage service for classified storage, and forming a cloud data resource pool; S2, preprocessing the real vehicle drive test data in the cloud data resource pool, extracting scene characteristics, automatically marking, generating structured scene data and storing the structured scene data into a generalized scene library; S3, selecting target scene data from the generalized scene library, scheduling and running an automatic driving algorithm to be tested, recharging the target scene data into the automatic driving algorithm to be tested for testing, and automatically generating a test report containing problem data identification and performance indexes; S4, judging whether the performance index in the test report meets a preset test passing condition or not; If yes, outputting a final test report to finish the soft recharging test; And then, starting a new round of soft recharging test by using the target scene data or a data set containing the scene identification corresponding to the problem data based on the optimized automatic driving algorithm, and returning to the step S3 of scheduling and running the optimized algorithm and carrying out the recharging test. In an optional implementation manner, in step S1, the acquiring real vehicle drive test data includes acquiring raw sensing data through an on-board sensor, and acquiring vehicle state data through a vehicle bus interface, where the on-board sensor includes at least one of a camera, a laser radar and a millimeter wave radar. In an optional embodiment, in step S2, the preprocessing the drive test data includes at least one operation of outlier removal, missing value complementation, time stamp alignment, and data validity verification on the drive test data. In an optional implementation manner, before the step S3, the method further comprises a step S0 of algorithm containerization packaging and management, wherein the automatic driving algorithm to be tested and the operation dependent environment thereof are packaged into a container mirror image, and uploaded to a container mirror image warehouse for versioning management; In step S3, the automatic driving algorithm to be tested is specifically that a corresponding container mirror image is pulled from the container mirror image warehouse, and a corresponding containerized algorithm instance is scheduled and operated through a container scheduling platform. In an alternative embodiment, the container arranging platform is a Kubernetes cluster, and in step S3, the container algorithm instance is create