Data management system and method for autonomous driving algorithm training

Abstract

The application provides a kind of data management system and method for automatic driving algorithm training, comprising: original data storage management module: collecting original data, and using uniform metadata to uniformly manage the relevant physical information of the label information and storage address of original data; Automatic labeling module: frame extraction is carried out on video file to obtain image file; Dataset module: classify whether the data is a difficult example sample and save the labeled data as a dataset; System management module: coordinate the work flow of the entire data platform algorithm technical personnel; The data collected by the original data storage management module is uniformly managed and then enters the automatic labeling module for labeling, then the original data and labeled data in the platform are searched to form a dataset through the dataset module, and then the algorithm model is trained, and the system management module runs through the entire data management system. The application can provide strong guarantee for the training and effective application of environmental sanitation automatic driving algorithm.

Description

Technical Field

[0001] This invention relates to the field of unstructured data technology, and more specifically, to a data management system and method for training autonomous driving algorithms. Background Technology

[0002] The training of offline deep learning algorithms for automated sanitation driving relies on a large amount of unstructured data. Currently, there is no good data platform on the market to effectively manage and utilize this unstructured training data in this scenario. Model training cannot be performed directly using raw data. The management of raw data, label retrieval, difficult example retrieval, dataset generation, and other preliminary preparations all consume a significant amount of time for algorithm training technicians, severely impacting project progress and output, and urgently requiring a solution.

[0003] The invention patent with publication number CN108268600A discloses an AI-based method and apparatus for managing unstructured data, including: storing large-scale unstructured data objects through HDFS, HBase, and NoSQL; extracting data features from the unstructured data objects using AI intelligent algorithms, storing the extracted data features in an external table, and constructing an unstructured data model; using SQL to retrieve the feature table of the unstructured data objects to achieve management of unstructured data, while the model is used for similarity searches of unstructured data. However, the unstructured data filtering efficiency in this patent is low and cannot meet the algorithm's training data requirements. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a data management system and method for training autonomous driving algorithms.

[0005] According to the present invention, a data management system and method for training autonomous driving algorithms are provided, the solution of which is as follows:

[0006] Firstly, a data management system for training autonomous driving algorithms is provided, the system comprising:

[0007] Raw data storage management module: Collects raw data and uses unified metadata to manage the relevant physical information of the raw data, including its tag information and storage address;

[0008] Automatic annotation module: Extracts frames from video files to obtain image files;

[0009] Dataset module: Classifies whether the data are hard examples by setting a confidence range and saves the labeled data as a dataset;

[0010] System management module: coordinates and manages the workflow of algorithm technicians across the entire data platform;

[0011] The raw data storage management module collects and manages the data, which then enters the automatic annotation module for annotation. The dataset module then retrieves the raw and annotated data from the platform to form a dataset, which is then used for algorithm model training. The system management module runs through the entire data management system.

[0012] Preferably, the frame extraction in the automatic annotation module includes: extracting frames at specified time intervals and extracting frames during data upload;

[0013] The data upload includes video files and their corresponding GPS files. The two types of files are associated with start_time and end_time. The TIMESTAMP corresponding to the location is obtained based on the GPS location information, and then the time offset is calculated based on this time to extract the image frames from the video file.

[0014] Preferably, the specified time interval for frame extraction is applied to both overall intensive frame extraction and general frame extraction, while frame extraction during data upload is applied to specific scenarios.

[0015] Preferably, the automatic annotation module further includes an indexing module:

[0016] The images corresponding to geographical locations involved in the method of frame extraction at specified time intervals can be indexed using the following method:

[0017] 1) Obtain latitude and longitude and the corresponding time stamp;

[0018] 2) Read image filenames and write them to the time-series database;

[0019] 3) Obtain the latitude and longitude and the corresponding timestamp into memory, then query the image name written to the time series database, compare the latitude and longitude of the timestamp in the image name with the corresponding timestamp, and if the absolute value of the difference meets the preset value, then map this pair back to the time series database. The preset value is set according to the business.

[0020] 4) Based on the image name in the above mapping, request the RESTful interface of the OBS object storage service to obtain the image resource link corresponding to the image name and write it back to the time series database.

[0021] Secondly, a data management method for training autonomous driving algorithms is provided, the method comprising:

[0022] Step S1: The raw data storage management module receives the raw data and manages the raw data using a structured data representation method;

[0023] Step S2: The automatic annotation module samples the uploaded raw data according to two measurement methods, time and distance, and filters the raw data, selecting different automatic annotation models to annotate the image data in the raw data.

[0024] Step S3: In the dataset module, classify whether the original data and labeled data are difficult examples by setting the confidence range, and save the labeled original data and labeled data as datasets. Algorithm trainers can view the dataset list and related information of each dataset to complete the data retrieval business process.

[0025] In steps S1 to S3, the data platform is managed through the system management module.

[0026] Preferably, step S1 includes: the data acquisition vehicle reporting data to a designated server directory; the algorithm training personnel clicking the data upload option on the data platform portal; then filling in the name, tags, and description information for this data upload job on the data upload interface; and selecting the file to be uploaded. When the upload is triggered, the program reads the designated directory on the server and moves the data to the designated directory.

[0027] Preferably, step S1 further includes:

[0028] After the algorithm trainer submits the data upload form, the data platform receives it and reviews it. If the review fails, the data platform will reject the upload, and the user will need to re-enter the data upload job information and data file. If the form is approved, the data platform will begin the data upload job.

[0029] Preferably, step S2 includes: the algorithm trainer selecting the automatic annotation option in the data platform portal;

[0030] Automatically label data in a Session, or filter a Session by specifying filtering criteria;

[0031] After receiving the information form for automatic annotation, the data platform reviews it. If the review fails, the data platform will reject the request and the user will need to fill it out again. If the review passes, the data platform will return the list of automatic annotation tasks to the user and start the automatic annotation task.

[0032] Once the automatic labeling job is completed, the data platform will update the completion status of the automatic labeling job and retrieve all labeling results from the Redis cache service;

[0033] Users can view the annotation results on the data platform portal and decide whether to persist them and determine the hard examples. If they choose to persist the annotation results, the data platform will encapsulate the images and their annotation files in a Parquet file according to the schema set by the data platform, and write them to HDFS for persistent storage. After the persistent results are returned to the algorithm trainers, they wait for confirmation. Once the confirmation is completed, the automatic annotation job is considered finished.

[0034] Preferably, step S3 includes:

[0035] Users initiate a request by entering the data search menu and filling in the data search criteria on the data search interface. All search criteria together form a data search information form, which the data platform will review. If the review fails, the user needs to re-enter the search criteria. If the review passes, the user will enter the data search process on the data platform.

[0036] The data platform returns the dataset information obtained from the data retrieval to the algorithm trainers. The algorithm trainers can then view the list of datasets obtained from multiple data retrievals and the relevant information for each dataset to complete the data retrieval business process.

[0037] Preferably, the system management module manages the data platform by including user management, data management, menu management, tag management, and project management.

[0038] User management: The system administrator can view the users in the data platform and delete users who do not meet the usage conditions of the data platform.

[0039] Data management: Clean up existing data; system administrators clear data from the data platform according to data numbers.

[0040] Menu management: The system administrator manages aspects of the system's function menus, including adding, deleting, modifying, and querying them;

[0041] Tag Management: Manage tag information, perform add, delete, and query operations on tags. Each tag has its own category and specific tag value.

[0042] Project Management: Allows you to view and delete projects in the data platform. Each project corresponds to a project in the data model. Deleting a project will delete all information and data associated with that project.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] 1. The data management system and method constructed by this invention will aggregate and manage a large amount of effective data, and provide services such as data storage management, data filtering, automatic annotation, data retrieval, and system management, which will greatly improve the efficiency of algorithm engineering operations;

[0045] 2. When algorithm trainers want to use hard examples to fine-tune the trained model or directly use hard examples to train the model, they need to spend a lot of time and effort to find ideal training data. The hard example data retrieval module provided by this invention can significantly improve the performance of training models with hard examples and shorten the development cycle.

[0046] 3. Different datasets have different data annotation standards and varying raw data collection conditions, making it difficult to integrate and use data from different datasets. Ordinary algorithm trainers lack professional tools for the unified storage, management, and use of this autonomous driving data, resulting in low data utilization efficiency. This invention's system reverses this situation by providing professional tools for the unified storage, management, and use of this autonomous driving data.

[0047] Other beneficial effects of the present invention will be explained in detail through the introduction of specific technical features and technical solutions in specific embodiments. Those skilled in the art should be able to understand the beneficial technical effects brought about by these technical features and technical solutions through the introduction of these technical features and technical solutions. Attached Figure Description

[0048] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0049] Figure 1 This is a basic flowchart of autonomous driving.

[0050] Figure 2 System module diagram;

[0051] Figure 3 Flowchart of the raw data storage management module;

[0052] Figure 4 Flowchart for the automatic annotation module;

[0053] Figure 5 This is a flowchart for the dataset module. Detailed Implementation

[0054] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0055] This invention provides a data management system for training autonomous driving algorithms, referring to... Figure 1 and Figure 2 As shown, the system specifically includes the following:

[0056] Raw data storage management module: Collects raw data and uses unified metadata to manage the relevant physical information of the raw data, including its tag information and storage address.

[0057] Automatic annotation module: Extracts frames from video files to obtain image files.

[0058] The automatic annotation module includes frame extraction at specified time intervals and during data upload. The data upload includes video files and their corresponding GPS files, which are represented by start_time and end_time. The TIMESTAMP corresponding to the location is obtained based on the GPS location information, and the time offset is calculated based on this time to extract the image frames from the video file.

[0059] Frame skipping at specified time intervals is applied to both overall intensive frame skipping and general frame skipping, while frame skipping during data upload is applied to specific scenarios.

[0060] The automatic annotation module also includes an indexing module: the images corresponding to geographical locations involved in the method of extracting frames at specified time intervals can be indexed as follows:

[0061] 1) Obtain latitude and longitude and the corresponding time stamp;

[0062] 2) Read image filenames and write them to the time-series database;

[0063] 3) Obtain the latitude and longitude and the corresponding timestamp into memory, then query the image name written to the time series database, compare the latitude and longitude of the timestamp in the image name with the corresponding timestamp, and if the absolute value of the difference meets the preset value, then map this pair back to the time series database. The preset value is set according to the business.

[0064] 4) Based on the image name in the above mapping, request the RESTful interface of the OBS object storage service to obtain the image resource link corresponding to the image name and write it back to the time series database.

[0065] Dataset module: Classifies whether the data is a difficult example by setting a confidence range and saves the labeled data as a dataset.

[0066] System Management Module: Coordinates and manages the workflow of algorithm technicians across the entire data platform.

[0067] The raw data storage management module collects and manages the data, which is then automatically labeled by the labeling module. The dataset module can then retrieve the raw and labeled data from the platform to form a dataset for algorithm model training. The system management module runs through the entire data management system.

[0068] This invention also provides a data management method for training autonomous driving algorithms, the method specifically including:

[0069] Step S1: The raw data storage management module receives the raw data and manages the raw data using a structured data representation.

[0070] Step S1 includes: the data acquisition vehicle reports data to the designated server directory; the algorithm training personnel click the data upload option on the data platform portal, then fill in the name, tags, and description information for this data upload job on the data upload interface, and select the file to be uploaded; when the upload is triggered, the program reads the designated directory on the server and moves the data to the designated directory. Step S1 also includes:

[0071] After the algorithm trainer submits the data upload form, the data platform receives it and reviews it. If the review fails, the data platform will reject the upload, and the user will need to re-enter the data upload job information and data file. If the form is approved, the data platform will begin the data upload job.

[0072] Step S2: The automatic annotation module samples the uploaded raw data according to both time and distance measurement methods, filters the raw data, and selects different automatic annotation models to annotate the image data (image data is a part of the raw data).

[0073] Step S2 includes: the algorithm trainer selects the automatic labeling option in the data platform portal; performs an automatic labeling job on the data in a Session, or filters the Session by specifying filtering conditions.

[0074] After receiving the information form for automatic annotation, the data platform reviews it. If the review fails, the data platform will reject the request and the user will need to fill it out again. If the review passes, the data platform will return the list of automatic annotation tasks to the user and start the automatic annotation task.

[0075] Once the automatic labeling job is completed, the data platform will update the completion status of the automatic labeling job and retrieve all labeling results from the Redis cache service;

[0076] Users can view the annotation results on the data platform portal and decide whether to persist them and determine the hard examples. If they choose to persist the annotation results, the data platform will encapsulate the images and their annotation files in a Parquet file according to the schema set by the data platform, and write them to HDFS for persistent storage. After the persistent results are returned to the algorithm trainers, they wait for confirmation. Once the confirmation is completed, the automatic annotation job is considered finished.

[0077] Step S3: In the dataset module, classify whether the original data and labeled data are difficult examples by setting the confidence range, and save the labeled original data and labeled data as datasets. Algorithm trainers can view the dataset list and related information of each dataset to complete the data retrieval business process.

[0078] Step S3 includes: the user enters the data retrieval menu to initiate a request, fills in the data retrieval conditions on the data retrieval interface, and all the retrieval conditions together form a data retrieval information form. The data platform will review the information form. If the review fails, the user needs to re-enter the retrieval conditions. If the review passes, the user will enter the data retrieval process of the data platform.

[0079] The data platform returns the dataset information obtained from the data retrieval to the algorithm trainers. The algorithm trainers can then view the list of datasets obtained from multiple data retrievals and the relevant information for each dataset to complete the data retrieval business process.

[0080] In steps S1 to S3, the data platform is managed through the system management module.

[0081] Specifically, the system management module manages the data platform, including user management, data management, menu management, tag management, and project management.

[0082] User management: The system administrator can view the users in the data platform and delete users who do not meet the usage conditions of the data platform.

[0083] Data management: Clean up existing data; system administrators clear data from the data platform according to data numbers.

[0084] Menu management: The system administrator manages aspects of the system's function menus, including adding, deleting, modifying, and querying them;

[0085] Tag Management: Manage tag information, perform add, delete, and query operations on tags. Each tag has its own category and specific tag value.

[0086] Project Management: Allows you to view and delete projects in the data platform. Each project corresponds to a project in the data model. Deleting a project will delete all information and data associated with that project.

[0087] The present invention will now be described in more detail.

[0088] This invention provides a data management system for training autonomous driving algorithms, referring to... Figure 2 As shown, the system specifically includes:

[0089] In the raw data storage management module: based on the information collected by the collection vehicle, the physical information such as the tag information and storage address of the raw data is managed in a unified manner using unified metadata.

[0090] Automatic annotation module: includes automatic annotation job module, customized data sampling module (fuzzy filtering, customized frame extraction, data label filtering, etc.), and filtering and uploading module.

[0091] Automatic image annotation is used to obtain object categories and bounding box information from images. Therefore, it requires frame extraction from video files to obtain image files. There are two frame extraction strategies: one is to extract frames at specified time intervals, and the other is to include both the video file and its corresponding GPS file during data upload. These two files are mapped using `start_time` and `end_time`. The `TIMESTAMP` corresponding to the GPS location is obtained, and the time offset is calculated based on this time to extract the image frames from the video file. The first frame extraction method is generally used for overall dense frame extraction and general frame extraction, while the second method is generally used for specific scenarios.

[0092] Indexing Module: The images corresponding to geographical locations involved in the first frame extraction method can be indexed as follows: obtain latitude and longitude and the corresponding time-series timestamp; read the image file name and write it to the time-series database; obtain latitude and longitude and the corresponding timestamp in memory, then query the image name written to the time-series database, compare the latitude and longitude timestamp in the image name with the corresponding timestamp, and if the absolute value of the difference meets the preset value, then map this pair back to the time-series database. The preset value is set according to the business; finally, based on the image name in the above mapping, request the RESTful interface of the OBS object storage service to obtain the image resource link corresponding to the image name and write it back to the time-series database.

[0093] Dataset module: includes data retrieval module, difficult example data retrieval module, and image data generation module.

[0094] The system management module is responsible for user management, tag management, and project management for the entire system. It is used to coordinate and manage the workflow of algorithm technicians across the entire data platform.

[0095] In the raw data storage management module: After the data acquisition vehicle uploads raw data through the system's I / O module, this raw data is managed using a structured data representation method, facilitating the implementation of the data platform system's functions. The specific steps are as follows:

[0096] Reference Figure 3 As shown, the data acquisition vehicle reports data to the designated server directory. Algorithm trainers click the data upload option on the data platform portal, then fill in the name, tags, and description information for this data upload job on the data upload interface, and select the files to be uploaded. A single data upload job can contain multiple files of different formats. When the upload is triggered, the program reads the designated directory on the server and moves the data to the designated directory. After the algorithm trainers submit the data, the data platform receives the data upload information form and reviews it. The review includes whether the name and description conform to the specifications, whether the selected tag information conflicts, and whether the uploaded data files meet the format requirements, etc.

[0097] If the review fails, the data platform will reject the application, and the user will need to re-enter the data upload job information and data files. If the information form passes the review, the data platform will begin the data upload job. The data platform first creates a Session table corresponding to the original data for this data upload job (this table records information related to the session entity, including its project number, tag, name, description, start time, and end time, etc.), and creates Segments based on the characteristics of the uploaded data. Since a single upload job can contain multiple files in different formats, the files are categorized by file format to obtain their subject information; uploaded data may include video data, GPS text data, and LiDAR point cloud data, etc. The data platform needs to standardize the file formats, with video files uniformly in MKV format. The processed data is assigned to appropriate data containers and persisted in the OBS object storage service. The data platform returns the data upload job list to the user, thus completing a data upload job.

[0098] The automatic annotation module samples the uploaded raw data using both time and distance metrics. Data is filtered by setting label information, time range, and other conditions, and different automatic annotation models are selected to annotate the image data. The specific steps are as follows:

[0099] Reference Figure 4As shown, algorithm trainers first select the automatic annotation option in the data platform portal. Automatic annotation can be performed on data from a single session, or multiple sessions that meet specific filtering criteria can be automatically annotated. This example illustrates automatic annotation for multiple sessions. Algorithm trainers select filtering criteria for sessions, including label information and time range. They also need to choose a pre-defined automatic annotation model to automatically annotate objects in the image. The choice of model depends on the object categories of interest to the algorithm trainer in the image.

[0100] To generate the image data to be labeled, video files in OBS need to be sampled. There are two sampling methods: time-based sampling and distance-based sampling. Algorithm trainers fill in the sampling frequency under the selected sampling method to complete the information required for the automatic labeling task. After receiving the automatic labeling information form, the data platform reviews it, including the correctness of the filter criteria and the range of sampling frequency values. If the review fails, the data platform rejects the request, requiring the user to resubmit; if the review passes, the data platform returns the user's automatic labeling task list and begins the automatic labeling task. The data platform first filters the sessions under the current project according to the filter criteria and obtains the session data streams that meet the filter criteria from OBS. The data platform samples the obtained data streams according to the sampling method and sampling frequency set by the user to obtain the image data to be labeled. Optionally, image data can be obtained from the index module of the automatic labeling module.

[0101] Next, based on the user-selected automatic annotation model, the image data is automatically annotated. Images are input into the model in batches according to a queue, and the annotation results are output. The annotation results are continuously written to a Redis cache service for caching. Users can refresh the job status to view the annotation execution status in real time. After the automatic annotation job is completed, the data platform will update the completion status of the automatic annotation job and retrieve all annotation results from the Redis cache service. Users can view the annotation results in the data platform portal and decide whether to persist them and determine the hard examples. If the annotation results are persisted, the data platform will encapsulate the images and their annotation files in a Parquet file according to the schema set by the data platform and write it to HDFS for persistent storage. The persisted results are returned to the algorithm training personnel for confirmation. Once confirmation is completed, the automatic annotation job is considered finished.

[0102] Reference Figure 5As shown, the dataset module classifies data as difficult examples by setting a confidence range and saves the labeled data as a dataset. Algorithm trainers can input search conditions (label conditions, model conditions, bounding box conditions, confidence range, and maximum number of images to retrieve, etc.) according to their data needs to retrieve raw and labeled data from the platform and form a dataset. Algorithm trainers can perform incremental operations on the image data as needed. The specific steps are as follows:

[0103] Users initiate a data retrieval request by entering the data retrieval menu and filling in data retrieval conditions on the data retrieval interface, including label information, annotation model, object bounding box range, confidence range, and maximum number of images. All retrieval conditions together form a data retrieval information form, which the data platform will review. If the review fails, the user needs to re-enter the retrieval conditions; if the review passes, the user enters the data retrieval process on the data platform. The data platform first constructs a real-time distributed SQL query statement based on the data retrieval conditions. This distributed SQL query statement is constructed through a series of string operations, connects to HDFS which stores images and their annotation results, and queries the stored data. Next, the query results are compiled into a dataset, allocated to a container, and persisted in HDFS. The data platform returns the retrieved dataset information to the algorithm trainers, who can view the list of datasets obtained from multiple data retrievals and related information for each dataset, thus completing the data retrieval business process.

[0104] In the system management module, the system administrator can manage the data platform, specifically including user management, data management, menu management, tag management, and project management. The system administrator can view the users on the data platform and delete users who do not meet the platform's usage requirements. Tag management involves managing tag information, allowing for adding, deleting, and querying tags. Each tag has its category and specific tag value; for example, if the category is weather, tag values ​​could be Sunny, Cloudy, Rainy, etc. Project management involves viewing and deleting projects on the data platform. Each project corresponds to a project in the data model; deleting a project deletes all information and data associated with that project. Data management primarily involves cleaning up existing data; the system administrator can clear data from the data platform based on data IDs such as dataset_id. Menu management allows the system administrator to manage system function menus by adding, deleting, modifying, and querying them.

[0105] This invention provides a data management system and method for training autonomous driving algorithms, which unifies the management of unstructured data required for autonomous driving system training and organically integrates it with automated labeling and sampling processes. The unstructured data labeling system for autonomous driving can perform frame sampling and matching based on filtering conditions (latitude and longitude and distance frames collected by sanitation workers, image blur filtering, labels for sanitation work objects, etc.). Training models with difficult examples can significantly improve model performance. When algorithm trainers want to use difficult examples to fine-tune the trained model or directly train the model using difficult examples, they do not need to spend a lot of time and effort searching for ideal training data.

[0106] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0107] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A data management system oriented to automatic driving algorithm training, characterized in that, Comprise: The original data storage management module: collect original data, and uniformly manage the relevant physical information of the original data including the label information and storage address with unified metadata; Automatic labeling module: frame the video file to get image file; Dataset module: classify the data as difficult example sample by setting confidence range, and save the labeled data as dataset; System management module: coordinate the work flow of the algorithm technical personnel of the whole data platform; The data collected by the original data storage management module is uniformly managed and then enters the automatic labeling module for labeling, and then the original data and labeled data in the platform are searched to form a dataset through the dataset module, and finally the algorithm model is trained, and the system management module runs through the whole data management system; The automatic labeling module samples the uploaded original data according to time and distance measurement methods respectively, and filters the original data, and selects different automatic labeling models to label the image data in the original data. 2.The data management system for automatic driving algorithm training of claim 1, wherein, The frame extraction in the automatic labeling module includes: frame extraction at a specified time interval and frame extraction when data is uploaded; Wherein, the data uploaded includes video files and corresponding GPS files, the two files are corresponded by using start_time and end_time, the TIMESTAMP corresponding to the position is obtained according to the position information of GPS, and then the picture frame in the video file is taken out according to the time offset calculated according to the time. 3.The data management system for automatic driving algorithm training of claim 2, wherein, The frame extraction at a specified time interval is applied to overall dense frame extraction and general frame extraction. 4.The data management system for automatic driving algorithm training of claim 2, wherein, The automatic labeling module further comprises an index module: The geographical position corresponding picture involved in the frame extraction method at a specified time interval can be indexed by the following method: 1) Obtain the latitude and longitude and the corresponding time sequence timestamp; 2) Read the picture file name and write it into the time sequence database; 3) Obtain the latitude and longitude and the corresponding timestamp to the memory, and then query the picture name written in the time sequence database, compare the timestamp latitude and longitude in the picture name with the corresponding timestamp, and if the absolute value of the difference meets the preset value, the mapping is written back to the time sequence database, and the preset value is set according to the business; 4) According to the picture name in the above mapping, request the restful interface of the obs object storage server to obtain the picture resource link corresponding to the picture name, and write back to the time sequence database.

5. A data management method for automatic driving algorithm training, based on the data management system for automatic driving algorithm training according to any one of claims 1-4, characterized in that, Comprise: Step S1: the original data storage management module receives original data, and manages the original data by means of structured data representation; Step S2: the automatic labeling module samples the uploaded original data according to time and distance measurement methods respectively, and filters the original data, and selects different automatic labeling models to label the image data in the original data; Step S3: in the dataset module, whether the original data and labeled data is a difficult example sample is classified by setting confidence range, and the labeled original data and labeled data is saved as a dataset, and the algorithm training personnel checks the dataset list and the related information of each dataset to complete the data retrieval business process; The data platform is managed by the system management module in steps S1-S3.

6. The data management method for automatic driving algorithm training according to claim 5, characterized in that, The step S1 comprises: the data collection vehicle reports data to a designated server directory, an algorithm training personnel clicks a data uploading option on a data platform portal, then fills in a name, a label and a description information for this data uploading job on a data uploading interface, and selects a file to be uploaded, when uploading is triggered, a program reads the above-mentioned designated server directory, and moves the data to the designated directory.

7. The data management method for automatic driving algorithm training according to claim 6, characterized in that, The step S1 further comprises: After the algorithm training personnel submits, the data platform receives a data uploading information sheet, and audits it, if the audit fails, the data platform will reject it, and the user needs to fill in the data uploading job information and the data file again; if the information sheet passes the audit, the data platform will start the data uploading job. 8.The data management method for automatic driving algorithm training according to claim 5, wherein, The step S2 comprises: the algorithm training personnel selects an automatic labeling option on the data platform portal; Automatic labeling job is performed on the data in a Session, or the Session is filtered through a specified filtering condition; After the data platform receives the automatic labeling information sheet, it audits it, if the audit fails, the data platform will reject the request, and the user needs to fill it in again; if the audit passes, the data platform will return the automatic labeling job list information to the user, and start the automatic labeling job; After the automatic labeling job is completed, the data platform will update the automatic labeling job completion status, and take out all the labeling results from the Redis cache service; The user views the labeling results on the data platform portal, and decides whether to persist the labeling results and determine the difficult example data; if the user chooses to persist the labeling results, the data platform will encapsulate the pictures and their labeling files in a Parquet file according to the Schema set by the data platform, and write them into the HDFS for persistent storage; after the persistent result is returned to the algorithm training personnel, it waits for confirmation, and the confirmation completes the automatic labeling job. 9.The data management method for automatic driving algorithm training according to claim 5, wherein, The step S3 comprises: The user enters a data retrieval menu to initiate a request, fills in data retrieval conditions on a data retrieval interface, all the retrieval conditions together constitute a data retrieval information sheet, and the data platform will audit the information sheet; if the audit fails, the user needs to fill in the retrieval conditions again, if the audit passes, the data retrieval process in the data platform is entered; The data platform returns the data set information obtained by data retrieval to the algorithm training personnel, the algorithm training personnel views the data set list obtained by multiple data retrievals and the related information of each data set, and completes the data retrieval business process. 10.The data management method for automatic driving algorithm training according to claim 5, wherein, The system management module manages the data platform, which comprises user management, data management, menu management, label management and project management; The user management: a system administrator views the users in the data platform, and deletes a user who does not meet the use conditions of the data platform; The data management: the system administrator cleans up the data in the data platform according to the data number; The menu management: the system administrator manages the system function menu in aspects of adding, deleting, modifying and inquiring; The project management: the system administrator manages the project in aspects of adding, deleting, modifying and inquiring. Tag management: manage tag information, add, delete and search tag, each tag information has its own category and specific tag value; Project management: view and delete the project in the data platform, the project corresponds to the project in the data model, when deleting a project, delete all information and data belonging to the project.

Images