An intelligent driving scene data labeling and auditing method and system, and a storage device

Abstract

The application provides a kind of intelligent driving scene data labeling auditing method, system and storage device, method includes: according to data storage path, read label file;Parsing label file, the labeling data is one-to-one corresponding with image data, generates image label data sequence based on time sequence;According to image label data sequence and auditing demand, construct first enhanced prompt word, and generate data auditing plan using large model;Based on the data auditing plan, generate the directed acyclic graph of calling tool chain;Based on the directed acyclic graph, dynamically call each tool to audit the labeling data and image data, generate data auditing report and output.The intelligent driving scene data labeling auditing method of the application realizes automation from data reading, parsing, planning, executing auditing and report output whole process, greatly reduces manual operation link, improves data auditing efficiency, and improves data quality and reliability.

Description

Technical Field

[0001] This application relates to the field of intelligent driving data processing technology, and in particular to a method, system and storage device for labeling and reviewing intelligent driving scenario data. Background Technology

[0002] As a typical application of the deep integration of artificial intelligence and the transportation field, intelligent driving relies on the accurate support of massive amounts of data for its operation. The accuracy and quality of this data directly affect the safety and robustness of the intelligent driving system. Currently, the data acquisition process for intelligent driving includes specifying the data requirement scope, data collection, data annotation, and data annotation review. In the data annotation review stage, manual review or deep learning model review is commonly used. Manual review is not only inefficient and unable to handle massive amounts of data, but also prone to human error, resulting in some errors not being detected in a timely manner. Deep learning model review, based on fixed rules and limited training data, lacks a deep understanding of data semantics and the ability to flexibly handle complex scenarios, making it difficult to accurately identify annotation errors and guarantee data quality. Therefore, there is an urgent need for an efficient and high-quality data annotation review method. Summary of the Invention

[0003] This application aims to solve the aforementioned technical problems by providing a method, system, and storage device for data annotation and review in intelligent driving scenarios.

[0004] A method for data annotation and review in intelligent driving scenarios, the method comprising: reading a tag file according to a data storage path, wherein the tag file contains annotation data; parsing the tag file to correspond the annotation data one-to-one with image data, and generating an image tag data sequence based on time sequence; constructing a first enhanced prompt word according to the image tag data sequence and review requirements, and generating a data review plan using a large model; generating a directed acyclic graph of the toolchain based on the data review plan; and dynamically calling various tools to review the annotation data and image data based on the directed acyclic graph, generating and outputting a data review report.

[0005] This solution automates the entire process from data reading, parsing, planning, execution review, and report output, significantly reducing manual operations, improving data review efficiency, and minimizing manpower investment. It generates image label data sequences based on time series, ensuring the review process considers the time dimension, improving data quality and reliability. Leveraging the powerful semantic understanding and reasoning capabilities of a large-scale model, it generates data review plans that adapt to different review standards (such as the labeling specifications of different car manufacturers), avoiding the rigidity of rule engines. Based on a directed acyclic graph (DAG) dynamic toolchain design, it can flexibly combine various review tools to adapt to changing needs. Furthermore, the knowledge enhancement characteristics of the large-scale model enable it to continuously learn new review rules and scenario knowledge, improving its ability to handle complex scenarios and new error types, exhibiting good scalability and long-term applicability.

[0006] Further, the label file is parsed to correspond the annotation data one-to-one with the image data, and an image label data sequence is generated based on time sequence, including: the label file also includes a timestamp corresponding to the annotation data; the validity of the annotation data and the image data is verified; based on the timestamp, the image data and the corresponding annotation data of the same route are arranged in order to generate the image label data sequence.

[0007] By verifying the validity of labeled and image data, abnormal image and labeled data can be identified, improving data quality. Verification of the validity of labeled and image data includes, but is not limited to, identifying low-quality abnormal image data caused by sensor noise or harsh environments, as well as labeled data with incorrect timestamp formats or obvious inconsistencies. In this solution, by verifying the timestamp information of the labeled data, after identifying abnormal data, the corresponding labeled and image data are removed, ensuring the accuracy and validity of the data sequence. By establishing an ordered image label data sequence, it is possible to completely preserve the state change information under different times and scenarios, providing a rich and reliable data foundation for subsequent data review from a time-series perspective, greatly improving the accuracy and comprehensiveness of the review.

[0008] Furthermore, the construction of the first enhanced prompt word includes: integrating information from the image tag data sequence and review requirements based on data type, annotation specifications, data timestamp information, and error type to construct the first enhanced prompt word.

[0009] In this solution, NLP technology can be used to automatically construct the first enhanced prompt word by integrating information from multiple dimensions such as data type. This can significantly improve the accuracy of the prompt word, task adaptability, and generation efficiency. Through the semantic understanding and structured integration of key information such as data type and annotation specifications by NLP, the first prompt word is upgraded from being driven by human experience to being driven by data features, which can promote the generation of more accurate task planning by large models.

[0010] Furthermore, the step of generating a directed acyclic graph of the calling toolchain based on the data audit plan includes: decomposing and analyzing the data audit plan, identifying statement anomalies in the data audit plan, and handling statement anomalies.

[0011] In this solution, after generating the data review plan using a large model, NLP syntactic analysis techniques, such as dependency parsing and constituent parsing, can be used to deconstruct and analyze the sentence structure of the data review plan. This allows for the checking of grammatical errors, such as missing sentence components, inconsistent tenses, and unclear pronouns. By identifying these anomalies, missing information can be inferred from the context, or a clearer data review plan can be regenerated using the large model, ensuring accurate understanding of the data review plan's logic and steps in subsequent implementations.

[0012] Furthermore, the step of generating a directed acyclic graph of the toolchain based on the data audit plan also includes: performing logical verification on the data audit plan, wherein the logical verification includes feasibility verification of the plan and feasibility verification of tool calls; If the data audit planning logic fails the verification, the data audit plan is adjusted and optimized based on the preset reconstruction strategy; if the data audit planning logic passes the verification, a directed acyclic graph of the calling toolchain is generated based on the data audit plan.

[0013] In this solution, before generating the directed acyclic graph of the toolchain, a comprehensive logical verification is performed on the data audit plan based on its syntax parsing. This verifies the feasibility of the plan in the actual audit process from a logical perspective, analyzing whether the audit steps in the plan meet the actual needs and logical relationships of intelligent driving data auditing. For example, it checks for unreasonable sequences such as performing result judgment before data inspection, or redundant execution of unnecessary operations. Simultaneously, the feasibility of tool calls involved in the plan is verified, checking for permissions to call relevant tools and resources, the availability of tool API interfaces, and whether the required input parameters match the actual data format. If the data audit plan fails the logical verification, the audit steps and tool call order are replanned, taking into account factors such as data characteristics, audit objectives, and available tool resources.

[0014] Furthermore, the step of dynamically calling various tools to review the labeled data and image data based on the directed acyclic graph and generating a data review report includes: collecting the review results of the tools reviewing the labeled data and image data, constructing a second enhanced prompt word by combining the labeled data and image data, and generating a data review report using a large model.

[0015] In this scheme, based on the constructed directed acyclic graph, the review task is decomposed into modular nodes according to the order and dependencies of the nodes. The API interfaces of each tool in the toolchain are dynamically called to complete the review tasks of different modules. The review results of each tool on the labeled data and image data are collected, and the results are sorted and analyzed to construct a second enhanced prompt word. Using a large model, a detailed data review report is finally generated. The data review report will point out the physical and semantic errors in the annotation, explain the causes of the errors, and provide corresponding correction suggestions, providing clear guidance for subsequent data correction.

[0016] Furthermore, the step of dynamically calling various tools to review the labeled data and image data based on the directed acyclic graph includes: monitoring the calling status of each tool, and automatically retrying the tool call or switching to a backup tool if a tool call anomaly occurs.

[0017] In this solution, the API call status of the monitoring tool is monitored in real time. If problems such as call timeout or interface abnormality occur, the system will automatically retry or switch to a backup tool to ensure the continuity of the audit operation.

[0018] Furthermore, the step of reading the tag file according to the data storage path includes: validating the data storage path and the tag file, and reading the tag file after the verification is successful.

[0019] This solution utilizes a technology that reads tag files based on data storage paths and verifies their validity. Through automated path parsing, multi-dimensional verification, and pre-emptive risk handling, it significantly improves the accuracy, efficiency, and security of data processing, laying a reliable foundation for subsequent tagging and review.

[0020] Based on the same concept, a data annotation and review system for intelligent driving scenarios is also proposed, comprising: an input layer that reads a tag file according to the data storage path, wherein the tag file contains annotation data; a perception layer that parses the tag file, maps the annotation data to image data one-to-one, and generates an image tag data sequence based on time sequence; constructs a first enhanced prompt word according to the image tag data sequence and review requirements, and generates the data review plan using a large model; a planning layer that generates the directed acyclic graph of the toolchain based on the data review plan; and an execution layer that dynamically calls various tools to review the annotation data and image data based on the directed acyclic graph, generates the data review report, and outputs it.

[0021] Based on the same concept, a computer storage device is also proposed, which stores computer-executable instructions that, when invoked and executed by a processor, cause the processor to perform the method described above.

[0022] Compared with the prior art, the beneficial effects of this application are as follows: This solution automates the entire process from data reading, parsing, planning, execution review, and report output, significantly reducing manual operations, improving data review efficiency, and minimizing manpower investment. It generates image label data sequences based on time series, ensuring the review process considers the time dimension, improving data quality and reliability. Leveraging the powerful semantic understanding and reasoning capabilities of a large-scale model, it generates data review plans that adapt to different review standards (such as the labeling specifications of different car manufacturers), avoiding the rigidity of the rule engine. Based on a directed acyclic graph (DAG) dynamic toolchain design, it can flexibly combine various review tools to adapt to changing needs. Furthermore, the knowledge enhancement characteristics of the large-scale model enable it to continuously learn new review rules and scenario knowledge, improving the system's ability to handle complex scenarios and new error types, exhibiting good scalability and long-term applicability. Attached Figure Description

[0023] Figure 1 This is a flowchart of the intelligent driving scenario data annotation and review method described in this application.

[0024] Figure 2 This is a schematic diagram of the directed acyclic graph of the calling toolchain described in this application.

[0025] Figure 3 This is a block diagram of the intelligent driving scenario data annotation and review system described in this application. Detailed Implementation

[0026] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application. Example 1

[0027] This embodiment provides a method for reviewing data annotations in intelligent driving scenarios. Based on the data requirements of intelligent driving scenarios, it automatically detects physical and semantic errors in the data annotations, generates an interpretable review report, and replaces some of the manual review work, thereby reducing labor costs while ensuring the quality of data annotation review. To better understand this method for reviewing data annotations in intelligent driving scenarios, the specific implementation steps are described in detail using the review of traffic light annotations in intelligent driving scenarios as an example.

[0028] like Figure 1 As shown, a method for data annotation and review in intelligent driving scenarios mainly includes the following steps S1 to S4.

[0029] S1, read the tag file according to the data storage path, where the tag file contains annotation data.

[0030] S2, parse the label file, match the labeled data with the image data one by one, and generate an image label data sequence based on time series.

[0031] S3: Based on the image tag data sequence and review requirements, construct the first enhanced prompt word and use the large model to generate a data review plan; based on the data review plan, generate a directed acyclic graph of the calling toolchain.

[0032] S4, based on a directed acyclic graph, dynamically calls various tools to review labeled and image data, generates and outputs a data review report.

[0033] In step S1, the tag file is read according to the data storage path, including: validating the data storage path and the tag file; and reading the tag file after successful verification. Validating the data storage path and tag file ensures the reliability of the data source by checking if the path exists, if the file is accessible, and if the file format is correct. If the data storage path is invalid, an error message is sent to the user to prompt them to enter the correct storage path. After successful verification, a stable connection is established with the data file, and multi-threaded or asynchronous reading technology is used to efficiently read the tag file content, preparing sufficient data for subsequent data parsing and review operations, and avoiding slow data reading that could affect review efficiency. The technology of reading the tag file according to the data storage path and performing validity verification significantly improves the accuracy, efficiency, and security of data processing through automated path parsing, multi-dimensional verification, and pre-risk handling, laying a reliable foundation for subsequent annotation review. Additionally, the tag file contains data information collected in intelligent driving scenarios, such as image data paths, timestamps, and annotation data.

[0034] It should be noted that in this embodiment, the data storage path adopts a JSON format tag path. The JSON format data file contains rich and crucial information. Taking the review of traffic light labeling in the intelligent driving scenario as an example, the position coordinates of the traffic light in the image are stored in the form of [x, y, width, height], which is used to accurately describe the geometric position of the traffic light in the image; the category field is clearly labeled as "red_light", "green_light", "yellow_light", etc., to distinguish traffic lights in different states; and the time sequence information related to the image is recorded in the form of timestamps, accurate to the millisecond level, such as "2024-12-01 10:30:25.123". This timestamp not only reflects the time point of image acquisition, but can also be used for subsequent analysis of the changes in traffic light status over time.

[0035] In step S2, the tag file is parsed by calling a professional JSON parsing library, such as the json library in Python. During the parsing process, each data entry in the JSON file is traversed, and the annotation data is precisely associated with the corresponding image data. It should be noted that the image data is stored as a separate file in a specified path. Specifically, the system establishes the correspondence between the two through the image data name or a unique identifier. For example, if the image data corresponding to a certain annotation record in the JSON file is named "image_001.jpg", the system will search for the image data in the preset image storage path and bind the two together. After parsing the tag file, the tag file also includes a timestamp corresponding to the labeled data; verifying the validity of the labeled data and image data; by verifying the validity of the labeled data and image data, abnormal image data and labeled data can be identified, improving data quality. Verifying the validity of labeled data and image data includes, but is not limited to: identifying low-quality abnormal image data caused by sensor noise or harsh environments, as well as labeled data with incorrect timestamp formats or obviously unreasonable data. In this embodiment, by verifying the timestamp information of the labeled data, after identifying abnormal data, the corresponding labeled data and image data are removed, and then the data is arranged according to the timestamp. The image data and the corresponding labeled data of the same route are arranged in an orderly manner to generate an image tag data sequence. By establishing an ordered image tag data sequence, the state change information under different times and scenarios can be completely preserved, providing a rich and reliable data foundation for subsequent data review from a time-series perspective, greatly improving the accuracy and comprehensiveness of the review. For example, when analyzing a continuous driving video data, the switching process of traffic light states can be clearly observed through this data sequence to determine whether there are abnormal switching situations.

[0036] In step S3, based on the image label data sequence and review requirements, a first enhanced prompt word is constructed. This includes integrating information from the image label data sequence and review requirements based on data type, annotation specifications, data timestamp information, and error type to construct the first enhanced prompt word. Specifically, Natural Language Processing (NLP) technology can be used to automatically construct the first enhanced prompt word by integrating information from multiple dimensions such as data type. This significantly improves the accuracy, task adaptability, and generation efficiency of the prompt word. Through NLP's semantic understanding and structured integration of key information such as data type and annotation specifications, the first prompt word is upgraded from being driven by human experience to being driven by data features, which can promote more accurate task planning in large models. The review requirements are related requirements pre-specified based on the labeled data, such as traffic light label review requirements. Taking traffic light label review in an intelligent driving scenario as an example, when constructing the first enhanced prompt word, information is integrated from multiple dimensions, including: Data type: Clearly specify the data type as "traffic light images and labeled data in intelligent driving scenarios", and list in detail the basic information such as image resolution and format, as well as the structure and meaning of labeled data, so that large models can quickly understand the basic characteristics of the data.

[0037] Labeling Specifications: This section details the definitions of traffic light categories, such as "red_light" representing a red light (when vehicles should stop) and "green_light" representing a green light (when vehicles can proceed). It also specifies coordinate labeling requirements, such as the origin being the top-left corner of the image, and coordinate values ​​being in pixels and non-negative integers. Furthermore, it introduces special labeling rules, such as labeling multiple traffic lights at an intersection in a left-to-right, top-to-bottom order.

[0038] Data timestamp information: The processed timestamp information is incorporated into the prompts to emphasize the need to pay attention to the changing patterns of traffic light status over time, such as the minimum time interval between traffic light transitions and the duration range of different statuses. For example, the duration of a red light is generally between 30 and 120 seconds, while the duration of a yellow light is typically 3 seconds. Error Types: Lists various common error types, such as traffic light misclassification (mislabeling a yellow light as a green light), excessive coordinate deviation (the target box fails to completely cover the traffic light), and unreasonable semantic information (two contradictory traffic light status labels appear in the same frame of the image), and provides specific error examples to help large models understand error features more intuitively. The following is an example of the first enhanced prompt word constructed for traffic light labeling review in an intelligent driving scenario.

[0039] You are a traffic light data auditing expert in the field of intelligent driving. Please analyze according to the following rules: 1. Timing consistency: Traffic light switching must conform to actual logic. The transition from red to green must be preceded by a yellow light, and the duration of the yellow light should be 3 ± 0.5 seconds. The time interval between adjacent states of the same traffic light shall not be less than 15 seconds (except in special circumstances).

[0040] 2. Spatial consistency: Traffic light labels at the same location must not conflict in adjacent frames. If the overlap of the coordinate frames exceeds 80%, the labels should remain consistent. The target frame for traffic lights should completely cover the actual traffic light area, and the deviation should not exceed 10% of the frame size.

[0041] 3. Physical rules: The height of a vertical traffic light frame should be greater than its width, with a ratio ranging from 2:1 to 3:1. Traffic light color information should be consistent with the category label, and RGB values ​​should be within the standard range (red light: R>200, G<50, B<50; green light: R<50, G>200, B<50; yellow light: R>200, G>200, B<50).

[0042] The input data is in a list. Question: Please improve my analysis rules and plan a reasonable review method.

[0043] Once the first enhanced prompt is constructed, it is encapsulated into a standard request format according to a pre-defined interface protocol, such as HTTP or gRPC, and sent to the large model interface. Taking HTTP as an example, the system will send the prompt as the request body via the POST method to the specified API address of the large model, and add necessary authentication information and data format specifications to the request header to ensure stable and reliable interaction with the large model, requesting the large model to generate a comprehensive and accurate data audit plan based on the prompt.

[0044] In step S3, based on the data audit plan, a directed acyclic graph of the calling toolchain is generated, including: decomposing and analyzing the data audit plan, identifying statement anomalies in the data audit plan, and handling statement anomalies.

[0045] After generating a data review plan using a large model, NLP syntactic analysis techniques, such as dependency parsing and constituent parsing, can be used to deconstruct and analyze the sentence structure of the data review plan. This allows for the detection of grammatical errors, such as missing sentence components, inconsistent tenses, and ambiguous referents. For example, a data review plan containing vague expressions like "first check the coordinates, then determine the color and this" is an example of a sentence anomaly. Identifying these anomalies allows for the inference of missing information from the context, or the use of the large model to regenerate a clearer data review plan, ensuring a precise understanding of the model's thought process and steps in subsequent data review processes.

[0046] Furthermore, based on the above statement exception handling, the data audit plan is logically validated, including feasibility validation of the plan and feasibility validation of tool calls; if the data audit plan fails the logical validation, the data audit plan is adjusted and optimized based on the preset reconstruction strategy.

[0047] A comprehensive logical validation of the data audit plan is conducted to verify its feasibility in the actual audit process. This includes analyzing whether the audit steps in the plan align with the actual needs and logical relationships of intelligent driving data auditing. For example, it checks for unreasonable sequences such as performing result judgments before data checks, or redundant execution of unnecessary operations. Simultaneously, the feasibility of tool calls involved in the plan is verified, checking for permissions to access relevant tools and resources, the availability of tool APIs, and whether the required input parameters match the actual data format. If the data audit plan fails the logical validation, the audit steps and tool call order are redesigned, taking into account data characteristics, audit objectives, and available tool resources. For instance, if a tool in the original plan is unusable due to version incompatibility, the system automatically finds a similar alternative tool and adjusts the corresponding calling logic to ensure a new, feasible audit plan is generated.

[0048] It should be noted that the preset refactoring strategy includes at least: tool replacement and call order optimization, and refactoring of the review process logic.

[0049] Tool substitution and call order optimization: When tool call feasibility verification fails, such as due to tool version incompatibility, API unavailability, or insufficient permissions, the system automatically searches for alternative tools with similar functions. For example, it switches from OpenCV to TensorFlowLite for image coordinate verification and re-plans the tool call order, such as moving later steps to earlier ones. Based on data characteristics and available resource libraries, the system ensures the feasibility of the audit plan.

[0050] The logic of the review process has been restructured: When the feasibility verification of a planning scheme fails, such as due to unreasonable step order, repetitive operations, or semantic contradictions, the logic of the steps is optimized through contextual inference or large model regeneration. For example, the process of "judging the result first and then checking the data" is changed to "data verification first and then result analysis." Review requirements (such as the continuity of timestamp sequences) have been considered to avoid logical loopholes.

[0051] like Figure 2 As shown, after the data review plan passes logical verification, the system transforms it into a directed acyclic graph (DAG) representing the toolchain. When constructing the DAG, the system treats each operation step and tool call in the review plan as a node in the graph, with directed edges between nodes representing the order and dependencies of the operations. For example, in the traffic light labeling review scenario of intelligent driving, the first step is to call a geometric calculation tool to check the accuracy of the traffic light target box coordinates; this operation serves as the starting node. Only after the coordinate check is completed and the result meets the requirements will the category verification tool be called to determine if the traffic light category is correct. At this point, the category verification tool's calling node is connected to the geometric calculation tool node through directed edges, explicitly indicating their dependency. By constructing the DAG, the system can clearly display the logical structure of the entire review process, facilitating the subsequent execution layer to dynamically call the toolchain in the correct order and according to rules, achieving efficient and orderly review operations and avoiding chaotic or repetitive operations.

[0052] In step S4, based on the directed acyclic graph, various tools are dynamically invoked to review the labeled data and image data, and a data review report is generated. This includes: collecting the review results of the tools reviewing the labeled data and image data, combining the labeled data and image data to construct a second enhanced prompt word, and using a large model to generate a data review report.

[0053] Based on the constructed directed acyclic graph, the API interfaces of each tool in the toolchain are dynamically invoked according to the order and dependencies of the nodes. During the invocation process, efficient task scheduling algorithms, such as priority scheduling or time-slice round-robin scheduling, are employed to rationally allocate system resources and ensure that the API interface calls of each tool can be executed efficiently and stably. Taking the review of traffic light markings in an intelligent driving scenario as an example, when calling the API interface of the geometry calculation tool, the parsed traffic light coordinate information is used as input parameters, encapsulated according to the format specified by the API interface, and then a call request is sent. After receiving the request, the geometry calculation tool calculates the center coordinates, area, aspect ratio, and other parameters of the target box using its internal geometric algorithm, and compares them with preset standard ranges to determine if there are any deviations in the coordinates. Throughout the entire invocation process, the system monitors the API interface call status in real time. If problems such as call timeout or interface abnormality occur, the system will automatically retry or switch to a backup tool to ensure the continuity of the review operation.

[0054] Intermediate results generated after the execution of various tools are collected, organized, and analyzed. Combined with the original image and label data, Natural Language Processing (NLP) techniques are used to construct second-level enhanced prompts. For example, a Transformer-based model can be used for text summarization and key entity extraction; combined with computer vision output, visual features are converted into text, and a rule-based template engine transforms the analysis results into second-level enhanced prompts. These second-level enhanced prompts include error detection information obtained from tool execution, original data features, and key details from the review process. Then, the second-level enhanced prompts are input again into the large model interface. The large model performs comprehensive analysis and judgment based on this information, ultimately generating a detailed data review report. Taking the review of traffic light labels in an intelligent driving scenario as an example, the data review report will clearly point out physical errors (such as coordinate deviations exceeding the allowable range) and semantic errors (such as traffic lights appearing on car windows) in the traffic light labels, explain the causes of the errors, and provide corresponding correction suggestions, offering clear guidance for subsequent data correction.

[0055] In this embodiment, the entire process from data reading, parsing, planning, execution review, and report output is automated, significantly reducing manual operations, improving data review efficiency, and minimizing manpower investment. Image label data sequences are generated based on time series, ensuring the review process considers the time dimension, improving data quality and reliability. The powerful semantic understanding and reasoning capabilities of the large model are used to generate data review plans, which can adapt to different review standards (such as the labeling specifications of different car companies), avoiding the rigidity of the rule engine. The design based on a directed acyclic graph (DAG) dynamic toolchain allows for flexible combination of various review tools to adapt to changing needs. Furthermore, the knowledge enhancement characteristics of the large model enable it to continuously learn new review rules and scenario knowledge, improving the system's ability to handle complex scenarios and new error types, exhibiting good scalability and long-term applicability. Example 2

[0056] This embodiment proposes an intelligent driving scenario data annotation and review system. The system defines multiple hierarchical functional layers or categories. Each functional layer has a feed-through relationship with at least one adjacent functional layer, and each functional layer includes one or more functional tasks. Functional tasks may include planning or executing actions. Figure 3 As shown, the system comprises an input layer 100, a perception layer 200, a planning layer 300, and an execution layer 400 arranged sequentially. Wherein: Input layer 100 reads a tag file according to the data storage path, wherein the tag file contains annotation data; The perception layer 200 parses the label file, maps the labeled data to the image data one by one, and generates an image label data sequence based on time sequence; according to the image label data sequence and review requirements, it constructs the first enhanced prompt word, and uses a large model to generate the data review plan; Planning layer 300, based on the data review planning, generates the directed acyclic graph of the calling toolchain; Execution layer 400, based on the directed acyclic graph, dynamically calls various tools to review the labeled data and image data, generates the data review report, and outputs it.

[0057] It should be noted that the input layer 100, perception layer 200, planning layer 300, and execution layer 400 may include hardware and / or software. The input layer 100 can access a memory storing label files and image data to read input data. The perception layer 200, planning layer 300, and execution layer 400 are communicatively connected to the large model system, and the execution layer 400 is also communicatively connected to the required software tools. Furthermore, the input layer 100, perception layer 200, and planning layer 300 may each include at least one processor, or share a single processor. Additionally, the system may include a memory for storing data review planning, directed acyclic graphs, and data review reports, among other related data. Example 3

[0058] This application also provides a computer storage device. The methods described in the embodiments of this application can be implemented in hardware or firmware, or implemented as computer instruction code that can be recorded on a computer storage device, or implemented as computer instruction code downloaded via a network and originally stored in a remote computer storage device or a non-transitory machine-readable computer storage device and then stored in a local computer storage device. Thus, the methods described herein can be stored in software processing on a computer storage device using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The computer storage device can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the computer storage device may also include combinations of the above types of memory. It is understood that the computer, processor, microprocessor controller, or programmable hardware includes storage components capable of storing or receiving software or computer instruction code. When the computer instruction code is accessed and executed by the computer, processor, or hardware, the above-described intelligent driving scenario data annotation and review method is implemented.

[0059] Although exemplary embodiments have been described herein with reference to the accompanying drawings, it should be understood that the above exemplary embodiments are merely illustrative and are not intended to limit the scope of this application. Various changes and modifications can be made therein by those skilled in the art without departing from the scope and spirit of this application. All such changes and modifications are intended to be included within the scope of this application as claimed in the appended claims.

[0060] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0061] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed.

[0062] The various component embodiments of this application can be implemented in hardware, or as software modules running on one or more processors, or a combination thereof. Those skilled in the art will understand that microprocessors or digital signal processors (DSPs) can be used in practice to implement some or all of the functions of some modules according to the embodiments of this application. This application can also be implemented as an apparatus program (e.g., a computer program and computer program product) for performing part or all of the methods described herein. Such an implementation of this application can be stored on a computer-readable medium, or can be in the form of one or more signals. Such signals can be downloaded from an Internet website, provided on a carrier signal, or provided in any other form.

[0063] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0064] Although the description of this application has been made in conjunction with the specific embodiments described above, it will be apparent to those skilled in the art that many substitutions, modifications, and variations can be made based on the foregoing. Therefore, all such substitutions, modifications, and variations are included within the spirit and scope of the appended claims.

Claims

1. A method for labeling and reviewing data in intelligent driving scenarios, characterized in that, The method includes: According to the data storage path, read the tag file, wherein the tag file contains annotation data and timestamps corresponding to the annotation data; The tag file is parsed to correspond the annotation data with the image data one-to-one, and an image tag data sequence is generated based on time sequence. Specifically, the validity of the annotation data and the image data is verified; based on the timestamp, the image data and the corresponding annotation data of the same route are arranged in order to generate the image tag data sequence. Based on the image tag data sequence and review requirements, a first enhanced prompt word is constructed, and a data review plan is generated using a large model; based on the data review plan, a directed acyclic graph of the calling toolchain is generated; the construction of the first enhanced prompt word includes: integrating information from the image tag data sequence and review requirements based on data type, annotation specifications, data timestamp information, and error type to construct the first enhanced prompt word; Based on the directed acyclic graph, various tools are dynamically invoked to review the labeled data and image data, generate a data review report, and output it.

2. The intelligent driving scenario data annotation and review method according to claim 1, characterized in that, The process of generating a directed acyclic graph of the toolchain based on the data audit plan includes: The data audit plan is decomposed and analyzed to identify and handle statement anomalies.

3. The intelligent driving scenario data annotation and review method according to claim 2, characterized in that, The process of generating a directed acyclic graph of the toolchain based on the data audit plan also includes: The data audit plan is logically validated, including feasibility validation of the plan and feasibility validation of tool calls. If the data audit planning logic fails the verification, the data audit plan will be adjusted and optimized based on the preset reconstruction strategy. If the data audit planning logic passes the verification, a directed acyclic graph of the calling toolchain is generated based on the data audit plan.

4. The intelligent driving scenario data annotation and review method according to claim 1, characterized in that, The process involves dynamically invoking various tools based on the directed acyclic graph to review the labeled data and image data, and generating a data review report, including: The review results of the annotation data and image data reviewed by various tools are collected. The annotation data and image data are combined to construct a second enhanced prompt word. A data review report is generated using a large model.

5. The intelligent driving scenario data annotation and review method according to claim 4, characterized in that, The process of dynamically invoking various tools to review the labeled data and image data based on the directed acyclic graph includes: Monitor the status of each tool call. If a tool call error occurs, automatically retry the tool call or switch to a backup tool.

6. The intelligent driving scenario data annotation and review method according to claim 1, characterized in that, The step of reading the tag file according to the data storage path includes: The data storage path and tag file are validated, and the tag file is read after the validation is successful.

7. A data annotation and review system for intelligent driving scenarios, characterized in that, include: The input layer reads a tag file according to the data storage path, wherein the tag file contains annotation data and timestamps corresponding to the annotation data; The perception layer parses the tag file, mapping the annotation data to image data one-to-one, and generates an image tag data sequence based on time sequence. Specifically, this involves: verifying the validity of the annotation data and image data; arranging the image data and corresponding annotation data for the same route in an ordered manner based on the timestamp to generate the image tag data sequence; constructing a first enhanced prompt word based on the image tag data sequence and review requirements, and generating a data review plan using a large model; the construction of the first enhanced prompt word includes: integrating information from the image tag data sequence and review requirements based on data type, annotation specifications, data timestamp information, and error type to construct the first enhanced prompt word. The planning layer reviews the planning based on the data and generates a directed acyclic graph of the toolchain. The execution layer, based on the directed acyclic graph, dynamically invokes various tools to review the labeled data and image data, generates a data review report, and outputs it.

8. A computer storage device, characterized in that, The computer storage device stores computer-executable instructions, which, when invoked and executed by a processor, cause the processor to perform the method according to any one of claims 1 to 6.

Images