Intelligent vehicle lamp AI perception data set version management method and acquisition method thereof

Abstract

The invention relates to the technical field of data set management, in particular to an AI perception data set version management method of an intelligent vehicle lamp and an acquisition method thereof, and the management method comprises the following steps: S1, determining a label category of a picture; s2, uploading a data set; s3, uploading the picture to a memory; s4, splitting the original labeling information according to the label category, and generating labeling information of each picture under different labels; s5, writing the labeling information of different labels into the corresponding picture label information sub-databases; s6, querying related pictures as an initial data set according to business requirements, and generating an information recording file; s7, adding the data set into version control management; and S8, when the content in the data set is adjusted, increased and deleted, the latest information record file needs to be regenerated, and then the step S7 is returned. All data sets are managed through the development platform and can be shared in time and used by others, and the change historical records can be effectively managed.

Description

Technical Field

[0001] The present invention relates to the technical field of data set management, and in particular to a version management method and an acquisition method of an AI perception data set for smart vehicle lights. Background Art

[0002] Smart headlights require cameras to constantly sense the surrounding environment and implement some intelligent headlight controls (e.g. adaptive high beam, intelligent cornering lighting, dynamic light pattern adjustment). AI models have increasingly higher requirements for the data size and quality of headlight perception, and these data increasingly need to be managed in a unified manner. To this end, we need to provide a version management method for AI perception data sets for smart headlights and a method for obtaining them.

[0003] At present, Git is usually used to manage different versions of dataset files. Git is an open source distributed version control system that can directly store files or data and record each submission information. At the same time, hash encryption and distributed storage ensure the security and integrity of data. Therefore, it is often used to manage code bases and small amounts of data. When Git is used for version control, the entire project can be traced back to a previous state. However, this method has the following shortcomings:

[0004] 1. When using Git for version control, the size of the Git repository is limited. Especially when you need to store a large number of data files (for example, large files such as images and videos), the granularity of version control may not be fine enough to track specific changes. Each submission will add a new version of the data set to the history record, which will cause the history record to become very large and take up a lot of storage space. At the same time, long-term storage of multiple versions of a large number of data sets will increase storage costs;

[0005] 2. When the content of a dataset changes, especially when multiple people modify the same version of a dataset, merge conflicts are likely to occur. Forced modifications will cause confusion in the data version, making it impossible for others to share and locate the data in a timely manner, thus affecting subsequent operations.

[0006] 3. For large-scale data sets, especially for version management of unstructured data such as images, each submission and recovery takes a long time, which is not conducive to project development (that is, since Git is based on files, this method is only suitable for small-scale data sets, but for large-scale data sets, due to the huge amount of data, each submission and push will consume a lot of resources and storage space, and each pull and backtracking will take a lot of time, which is not efficient for version control of some large data sets);

[0007] 4. Simply managing data sets, but not managing the data in them, will make it impossible to fully tap the value of the data. Summary of the invention

[0008] The technical problem to be solved by the present invention is: in order to solve the technical problem that the existing Git management of large-scale data set files takes a long time, the present invention provides an AI perception data set version management method for smart car lights and a method for obtaining the same. By improving the data set file management method, the management time of large-scale data sets can be shortened.

[0009] The technical solution adopted by the present invention to solve the technical problem is: a method for managing the version of an AI perception data set of an intelligent vehicle lamp, wherein the data set version management system comprises: a memory, a development platform and a database, wherein the memory is connected to the development platform, and the development platform is connected to the database;

[0010] The management approach includes the following steps:

[0011] S1. Determine the label category of the image based on the training task of the deep learning model corresponding to the business needs;

[0012] S2, upload the dataset and split it into images and original annotation information;

[0013] S3, the picture is uploaded to the storage device, and the storage address and its own information are written into the basic information sub-database;

[0014] S4, splitting the original annotation information according to the label category, and generating the annotation information of each image under different labels;

[0015] S5, writing the annotation information of different labels into the corresponding image label information sub-database;

[0016] S6. Query relevant pictures as the initial data set according to business needs, and generate information record files;

[0017] S7, adding the data set to version control management;

[0018] S8. When the content in the data set is adjusted, added, or deleted, the latest information record file needs to be regenerated, and then return to step S7.

[0019] Therefore, for the image dataset, by splitting it into two parts, pictures and picture label information, and saving them in the storage and database respectively, the data information is decoupled from the dataset. The pictures can be obtained through database query, which improves the utilization rate of the data and facilitates the mining of more value of the data. Since the dataset does not directly save the picture file and its original annotation information file, but records the relevant information of the document in the basic information sub-database of the corresponding picture and the relevant information of the document in the corresponding picture label information sub-database, a small amount of storage cost can be spent when performing version control management. All datasets are managed through the development platform, can be shared and used by others in a timely manner, and the change history records can also be effectively managed.

[0020] Furthermore, the memory is used to store original files and can back up data; the development platform is used to manage and query data; the database uses mongoDB to record basic information, tag information and data set information of files, and the database automatically assigns a unique identification id to each record information, and has a data backup function; the database includes: a basic information sub-database, a picture tag information sub-database and a data set information sub-database.

[0021] Furthermore, in step S1, the label categories include: target category, scene and lighting condition; in step S2, first, the data set needs to be manually uploaded to the development platform, and then the development platform will split the data set into pictures and original annotation information; wherein: the original annotation information includes: key targets or information in different pictures, and can be the annotation information of different targets of multiple different label categories in each picture; the split annotation information is the annotation information of the same category targets in each picture under a single label category.

[0022] Furthermore, in step S3, the development platform will first upload the image to the memory and record the storage address of the image, and then write the storage address and its own information of the image into the basic information sub-database; the basic information sub-database is a sub-database for recording basic information of files; wherein: the basic information sub-database includes multiple collections, each collection contains multiple documents, and each document records the basic information of a file.

[0023] Furthermore, in step S4, the annotation information of each image is in a dictionary format, and the annotation information of the image under a single tag is recorded through a key-value pair, and one image corresponds to one dictionary.

[0024] Further, in step S5, the image tag information sub-database is used to record the tag information of the image; wherein: the image tag information sub-database includes multiple collections, each collection contains multiple documents, and each document records the labeling information of all targets under the same label category in a picture; when the labeling information is changed and a new version needs to be added, the version number can be added in ascending order, and the latest labeling information is filled in the key-value pair corresponding to the latest version number.

[0025] Furthermore, in step S6, the development platform can filter out images that meet the conditions according to the query conditions, the query conditions are the information recorded in the database, and quickly find the corresponding document in the collection corresponding to the database through the unique identification ID of the document; the format of the information recording file is json format, and json records the relevant content of this data set. In this way, relevant data processing operations can be implemented.

[0026] Furthermore, the step S7 comprises the following steps:

[0027] S7-1. Use the DVC tool to track and record the information record file corresponding to the data set, and generate the corresponding .dvc file;

[0028] S7-2. Use the git tool to track and submit the .dvc file through the git add command, and record the commit id of this submission;

[0029] S7-3, query the data set information sub-database according to the name of the data set, obtain the highest historical version number of the data set, and add a new version number;

[0030] S7-4. Write the name of the dataset, new version number, commit id, update date, uploader and other information into the dataset information sub-database.

[0031] Furthermore, in step S7, the data set information sub-database is used to record the data set information sub-database; wherein: the data set information sub-database includes multiple sets, each set contains multiple documents, and each document records the DVC information of a version of the data set.

[0032] A method for obtaining a version management method of an AI perception data set for smart headlights comprises the following steps:

[0033] Sa. Select the corresponding dataset name and its version number on the development platform;

[0034] Sb. The development platform will query the dataset of this version in the dataset information sub-database and obtain the corresponding commit id;

[0035] Sc. Use the git tool and use the git checkout command to switch to the .dvc file corresponding to the commit id;

[0036] Sd, use the dvc checkout command to switch to the corresponding .dvc file to track the information in the record file data set information;

[0037] Se. According to the data set information in the information record file, obtain the document information of the basic information sub-database and the document information of the picture tag information sub-database corresponding to the picture;

[0038] Sf. Obtain the storage address of the image according to the document information of the basic information sub-database, and obtain the marking information of the corresponding label according to the document information of the image label information sub-database.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] 1. The present invention aims at image data sets, by splitting them into two parts, namely pictures and picture label information, and saving them in a memory and a database respectively, thereby decoupling data information from data sets. Pictures can be obtained through database query, which improves data utilization and facilitates mining more data value.

[0041] 2. Since the data set of the present invention does not directly save the image file and its original annotation information file, but records the relevant information of the document in the basic information sub-database corresponding to the image and the relevant information of the document in the corresponding image tag information sub-database, a small amount of storage cost can be spent when performing version control management.

[0042] 3. All data sets of the present invention are managed through the development platform, can be shared and used by others in a timely manner, and the change history records can also be effectively managed. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0044] Figure 1 A control block diagram of the data set version management system of the present invention;

[0045] Figure 2 It is a flow chart of the AI ​​perception data set version management method of the smart headlight of the present invention;

[0046] Figure 3 is a flow chart of step S7 of the present invention;

[0047] Figure 4 The present invention is a flowchart of a method for obtaining the version management method of the AI ​​perception data set of the smart headlights.

[0048] In the figure: 1. Dataset version management system; 2. Storage; 3. Development platform; 4. Database. DETAILED DESCRIPTION

[0049] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0050] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise", "axial", "radial", "circumferential" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention. In addition, features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, unless otherwise specified, "multiple" means two or more.

[0051] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0052] like Figures 1 to 3 As shown, a method for managing the version of an AI perception data set of a smart headlight, a data set version management system 1 includes: a memory 2, a development platform 3 and a database 4, the memory 2 is connected to the development platform 3, and the development platform 3 is connected to the database 4;

[0053] The management approach includes the following steps:

[0054] S1. Determine the label category of the image based on the training task of the deep learning model corresponding to the business needs;

[0055] S2, upload the dataset and split it into images and original annotation information;

[0056] S3, upload the picture to storage 2, and write the storage address and its own information into the basic information sub-database;

[0057] S4, splitting the original annotation information according to the label category, and generating the annotation information of each image under different labels;

[0058] S5, writing the annotation information of different labels into the corresponding image label information sub-database;

[0059] S6. Query relevant pictures as the initial data set according to business needs, and generate information record files;

[0060] S7, adding the data set to version control management;

[0061] S8. When the content in the data set is adjusted, added, or deleted, the latest information record file needs to be regenerated, and then return to step S7. Therefore, for the image data set, by splitting it into two parts, the picture and the picture label information, and saving them in the memory 2 and the database 4 respectively, the decoupling of the data information and the data set is achieved, and the picture can be obtained by querying the database 4, which improves the utilization rate of the data and facilitates the mining of more data value (the prior art takes a long time to submit and restore each version of a large-scale data set for unstructured data such as pictures; the data in the prior art data set lacks unified management, and as the amount of data increases, some key and high-quality data is not fully mined, and the data life cycle is short); because the data set does not directly save the picture file and its original annotation information file, but records the corresponding The relevant information of the document in the basic information sub-database of the picture (which can further obtain the storage address of the picture) and the relevant information of the document in the corresponding picture tag information sub-database (which can further obtain the annotation information under the corresponding tag of the picture) can spend a small amount of storage cost when performing version control management (the same file in different data sets in the prior art will be stored multiple times, increasing unnecessary storage costs); all data sets are managed through the development platform 3, can be shared and used by others in a timely manner, and the change history can also be effectively managed (in the prior art, different people modifying the data set will cause data version confusion, and the data cannot be shared and located by others in a timely manner, which will affect subsequent operations).

[0062] Specifically, the management method is: perform specific processing on the data set required by the business (for example, splitting, storing, and writing to the database 4), decoupling the image files and label information in the data set from the data set, so that the images and their label information can not only be synthesized into the corresponding data set, but also can be used in other subsequent aspects (for example, data retrieval, data mining).

[0063] In this embodiment, the memory 2 is used to store the original file and can back up the data; the development platform 3 is used to manage and query the data; the database 4 uses mongoDB to record the basic information, label information and data set information of the file. The database 4 automatically assigns a unique identification id to each recorded information and has a data backup function; the database 4 includes: a basic information sub-database, a picture label information sub-database and a data set information sub-database. Specifically, the development platform 3 includes but is not limited to the query, data processing and data set query and download functions of the database 4. The data processing includes but is not limited to the separation of picture labels, file upload, and database 4 write operations; the development platform 3 is connected to the memory 2 and the database 4 through the intranet, and can be accessed and managed by multiple people at the same time.

[0064] Specifically, the mongoDB database is a non-relational database 4, which is particularly suitable for storing unstructured or semi-structured data, and does not require pre-defined table results. It can dynamically and flexibly add field information, and is therefore very suitable for recording relevant information of images.

[0065] In this embodiment, in step S1, the label categories include: target categories (e.g., cars, people), scenes (e.g., urban areas, highways), and lighting conditions (e.g., weak light, strong light); in step S2, first, the data set needs to be manually uploaded to the development platform 3, and then the development platform 3 will split the data set into images and original annotation information; wherein: the original annotation information includes: key targets or information in different images, and can be annotation information for different targets of multiple different label categories in each image; the split annotation information is the annotation information of the same category of targets in each image under a single label category. Specifically, the required label category can be determined according to the current business needs, and the label information recorded in the image can be obtained by manual or automatic annotation.

[0066] In this embodiment, in step S3, the development platform 3 first uploads the image to the memory 2, and records the storage address of the image, and then writes the storage address and self information of the image (for example: collection source, date, purpose) into the basic information sub-database; the basic information sub-database is a sub-database for recording the basic information of the file; wherein: the basic information sub-database includes multiple collections (for example: picture collection, video collection), and there are multiple documents in the collection, and each document records the basic information of a file. Specifically, the storage address can create a hierarchical path of storage according to the self information of the image, so as to summarize and organize the files; the format of the document is: {"_id": "document's unique identifier id", "file_name": "file name", "source": "file collection source", "data": "upload date", "purpose": "file purpose", "address": "file storage address", "uploader": "uploader"}; for the image file, it can be saved to the image collection, and each document in the collection records the basic information of one image.

[0067] In this embodiment, in step S4, the annotation information of each image is in dictionary format, and the annotation information of the image under a single label is recorded through a key-value pair, and one image corresponds to one dictionary. Specifically, the format of the dictionary content is: {"file_name": "image name", "width": "image pixel width", "height": "image pixel height", "label": ["Information of all targets under this label in this image"]}, where: "label" is the label category, and the information of each target can be {"id": "target id", "struct_type": "annotation type (box, point, polygon, etc.)", "data": "target location", "attributes": ["attribute 01", "attribute 02"]}.

[0068] In this embodiment, in step S5, the image tag information sub-database is used to record the sub-database of the tag information of the image; wherein: the image tag information sub-database includes multiple sets (each set corresponds to a tag category), and there are multiple documents in the set, and each document records the annotation information of all targets under the same tag category in a picture; when the annotation information is changed and a new version needs to be added, the version number can be added in ascending order, and the latest annotation information is filled in the key-value pair corresponding to the latest version number. Specifically, the format of the document is: {"_id": "document's unique identifier id", "file_name": "image name", "version_01": "image version_01 annotation information", "version_02": "file version_02 annotation information"}, and the annotation information of each version of the image is the annotation information of the image under a single tag.

[0069] In this embodiment, in step S6, the development platform 3 can filter out images that meet the conditions according to the query conditions, the query conditions are the information recorded in the database 4, and quickly find the corresponding document in the collection of the corresponding database 4 through the unique identification id of the document; the format of the information recording file is json format, and json records the relevant content of this data set. In this way, relevant data processing operations can be implemented.

[0070] Specifically, the relevant content is: {"_id":"the unique identification id of the document","dataset_name":"dataset name","type":"the type of files in the dataset","number":"the number of files","uploader":"uploader","data":"upload date","project":"the task to which it belongs","data":["the relevant information corresponding to each image in this dataset"]}, wherein: the relevant information corresponding to each image in "data" is: {"base":"the relevant information of the document in the basic information sub-database corresponding to the image","label":"the relevant information of the document in the image label information sub-database corresponding to the image"}, the relevant information of the document in the basic information sub-database is: {"db":"database 4 name","collection":"collection name","_id":"the unique identification id of the document"}, the relevant information of the document in the image label information sub-database is: {"db":"database 4 name","collection":"collection name","_id":"the unique identification id of the document","version":"the version number of the annotation information"}.

[0071] In this embodiment, step S7 includes the following steps:

[0072] S7-1. Use the DVC tool to track and record the information record file corresponding to the data set, and generate the corresponding .dvc file;

[0073] S7-2. Use the git tool to track and submit the .dvc file through the git add command, and record the commit id of this submission;

[0074] S7-3, query the data set information sub-database according to the name of the data set, obtain the highest historical version number of the data set, and add a new version number;

[0075] S7-4, write the name, new version number, commit id, update date, uploader and other information of the data set into the data set information sub-database;

[0076] In step S7, the data set information sub-database is used to record the data set information sub-database;

[0077] The dataset information sub-database includes multiple collections (each collection corresponds to a dataset with a name), and each collection contains multiple documents, each of which records the DVC information of a version of the dataset. Specifically, DVC (Data Version Control) is a data version management tool for deep learning, which can generate metadata to describe data and models through the dvc add command, and can restore the corresponding data based on the metadata.

[0078] Specifically, the format of dvc information is: {“_id”: “the unique identification id of the document”, “dataset_name”: “dataset name”, “version”: “version number”, “commit_id”: “commit id when git is submitted”, “data”: “update date”, “uploader”: “uploader”}.

[0079] Specifically, the control method is: produce a data set information record file json (containing the relevant content of this data set, such as: data set name, upload date, database 4 information of each image), where: the database 4 information of the image includes: the storage address of the image and related label information, and then use dvc to track and record the data set information record file json, and use the commit id submitted by git, and finally write the data set name, version number, commit id, update date and other information into the data set information sub-database; if the data in the data set changes, regenerate the information record file json, add a new version number, and repeat the above operation.

[0080] Specifically, the .dvc file is a file used by the DVC data version management tool to track data sets, and can be understood as the key to open the target data set. Although the specific content recorded in the .dvc file may vary in different versions or configurations of DVC, in general, the .dvc file mainly records metadata related to the data set, including but not limited to: the unique identifier of the data set (for example, the hash value, MD5 value or other checksum of the data to ensure the integrity and consistency of the data), the storage path of the data set (that is, the location information of the data set in the local or remote storage, which is convenient for DVC to perform data management and access), dataset dependencies (if a dataset is obtained by some processing of other datasets or files, the .dvc file may record these dependencies so that the source of the dataset can be regenerated or tracked when needed), and dataset version information (recording the version number of the dataset or the commit hash value for version control and backtracking); in addition, DVC is integrated with other tools (such as git) through .dvc files to achieve more comprehensive version management and collaboration (for example, git is responsible for managing the version of .dvc files, while DVC is responsible for tracing and managing the actual dataset through .dvc files).

[0081] like Figure 4 As shown, a method for obtaining a version management method of an AI perception dataset for smart headlights includes the following steps:

[0082] Sa. Select the corresponding dataset name and its version number on the development platform 3;

[0083] Sb, the development platform 3 will query the dataset of this version in the dataset information sub-database and obtain the corresponding commit id;

[0084] Sc. Use the git tool and use the git checkout command to switch to the .dvc file corresponding to the commit id;

[0085] Sd, use the dvc checkout command to switch to the corresponding .dvc file to track the information in the record file data set information;

[0086] Se. According to the data set information in the information record file, obtain the document information of the basic information sub-database and the document information of the picture tag information sub-database corresponding to the picture;

[0087] Sf. Obtain the storage address of the image according to the document information of the basic information sub-database, and obtain the labeling information of the corresponding label according to the document information of the image label information sub-database. Specifically, the development platform 3 can combine and package the storage address of the image and the labeling information of the corresponding label according to the storage address of the image and the labeling information of the corresponding label, and then use them for the training of the deep learning model.

[0088] Specifically, the acquisition method is: when a certain version of a data set is needed, query the database 4 to obtain the corresponding commitid, and then the development platform 3 automatically backtracks according to the commit id to obtain the information record file json of the corresponding data set, and then the complete information of the corresponding data set can be obtained according to the json.

[0089] In summary, the present invention aims at an image dataset, by splitting it into two parts, namely, pictures and picture label information, and saving them in the memory 2 and the database 4 respectively, thereby realizing the decoupling of data information from the dataset. The pictures can be obtained by querying the database 4, thereby improving the utilization rate of the data and facilitating the mining of more value of the data. Since the dataset does not directly save the picture file and its original annotation information file, but records the relevant information of the document in the basic information sub-database of the corresponding picture and the relevant information of the document in the corresponding picture label information sub-database, a small amount of storage cost can be spent when performing version control management. All datasets are managed through the development platform 3, can be shared and used by others in a timely manner, and the change history records can also be effectively managed.

[0090] The above description is based on the ideal embodiment of the present invention. Through the above description, relevant personnel can make various changes and modifications without departing from the technical concept of the present invention. The technical scope of the present invention is not limited to the content in the specification, and its technical scope must be determined according to the scope of the claims.

Claims

1. A method for managing the version of AI perception datasets for smart headlights, characterized in that: The dataset version management system (1) includes: A memory (2), a development platform (3) and a database (4), wherein the memory (2) is connected to the development platform (3), and the development platform (3) is connected to the database (4); The management approach includes the following steps: S1. Determine the label category of the image based on the training task of the deep learning model corresponding to the business needs; S2, upload the dataset and split it into images and original annotation information; S3, uploading the picture to the memory (2), and writing the storage address and its own information into the basic information sub-database; S4, splitting the original annotation information according to the label category, and generating the annotation information of each image under different labels; S5, writing the annotation information of different labels into the corresponding image label information sub-database; S6. Query relevant pictures as the initial data set according to business needs, and generate information record files; S7, adding the data set to version control management; S8. When the content in the data set is adjusted, added, or deleted, the latest information record file needs to be regenerated, and then return to step S7.

2. The AI ​​perception data set version management method for smart car lights according to claim 1 is characterized in that: The memory (2) is used to store original files and can back up data; The development platform (3) is used to manage and query data; The database (4) uses mongoDB to record the basic information, tag information and data set information of the file. The database (4) automatically assigns a unique identification ID to each piece of recorded information and has a data backup function. The database (4) comprises: Basic information sub-database, image label information sub-database and dataset information sub-database.

3. The AI ​​perception data set version management method for smart car lights according to claim 1, characterized in that: In step S1, the tag categories include: Target category, scene, and lighting conditions; In step S2, first, the data set needs to be manually uploaded to the development platform (3), and then the development platform (3) will split the data set into images and original annotation information; The original annotation information includes key targets or information in different images, and can be annotation information of different targets of multiple different label categories in each image; The split annotation information is the annotation information of the same category of targets in each image under a single label category.

4. The AI ​​perception data set version management method for smart car lights according to claim 1, characterized in that: In step S3, the development platform (3) first uploads the image to the memory (2), records the storage address of the image, and then writes the storage address of the image and its own information into the basic information sub-database; The basic information sub-database is a sub-database used to record the basic information of the file; Among them: the basic information sub-database includes multiple collections, each collection contains multiple documents, and each document records the basic information of a file.

5. The method for managing the AI ​​perception dataset version of smart headlights according to claim 1, characterized in that: In step S4, the annotation information of each image is in a dictionary format, and the annotation information of the image under a single tag is recorded through key-value pairs, and one image corresponds to one dictionary.

6. The AI ​​perception data set version management method for smart car lights according to claim 1, characterized in that: In step S5, the picture tag information sub-database is used to record the tag information of the picture; The image tag information sub-database includes multiple collections, each collection contains multiple documents, and each document records the labeling information of all objects under the same tag category in a picture; When the annotation information changes and a new version needs to be added, the version numbers can be added in ascending order, and the latest annotation information is filled in the key-value pair corresponding to the latest version number.

7. The method for managing the version of the AI ​​perception dataset of the smart vehicle light according to claim 1, characterized in that: In step S6, the development platform (3) can filter out images that meet the conditions according to the query conditions, the query conditions being the information recorded in the database (4), and quickly find the corresponding document in the collection corresponding to the database (4) through the unique identification ID of the document; The format of the information record file is json, which records the relevant content of this data set.

8. The method for managing the AI ​​perception dataset version of smart vehicle lights according to claim 1, characterized in that: The step S7 comprises the following steps: S7-1. Use the DVC tool to track and record the information record file corresponding to the data set, and generate the corresponding .dvc file; S7-2. Use the git tool to track and submit the .dvc file through the git add command, and record the commit id of this submission; S7-3, query the data set information sub-database according to the name of the data set, obtain the highest historical version number of the data set, and add a new version number; S7-4. Write the name of the dataset, new version number, commit id, update date, uploader and other information into the dataset information sub-database.

9. The method for managing the AI ​​perception dataset version of smart vehicle lights according to claim 1, characterized in that: In step S7, the data set information sub-database is used to record the data set information sub-database; The dataset information sub-database includes multiple collections, each collection contains multiple documents, and each document records the DVC information of a version of the dataset.

10. A method for acquiring the AI ​​perception data set version management method for smart vehicle lights according to any one of claims 1 to 9, characterized in that: The following steps are involved: Sa. Select the corresponding dataset name and its version number on the development platform (3); Sb, the development platform (3) will query the dataset of this version in the dataset information sub-database and obtain the corresponding commit id; Sc. Use the git tool and use the git checkout command to switch to the .dvc file corresponding to the commit id; Sd, use the dvc checkout command to switch to the corresponding .dvc file to track the information in the record file data set information; Se. According to the data set information in the information record file, obtain the document information of the basic information sub-database and the document information of the picture tag information sub-database corresponding to the picture; Sf. Obtain the storage address of the image according to the document information of the basic information sub-database, and obtain the marking information of the corresponding label according to the document information of the image label information sub-database.

Images