Map generation method and apparatus

Abstract

This disclosure provides a map generation method and apparatus, relating to the field of artificial intelligence technology, particularly to the fields of multi-agent collaboration, natural language processing, computer vision, deep learning, and autonomous driving. One specific implementation of the method includes: constructing input pairs based on multi-view images, the input pairs including multi-view images and a current lane-level map; performing constraint review on the current lane-level map to generate structured evidence; performing action planning based on the input pairs and structured evidence to generate an action sequence; and editing and executing the current lane-level map based on the action sequence to generate a final lane-level map.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, particularly to the fields of multi-agent collaboration, natural language processing, computer vision, deep learning, and autonomous driving. Background Technology

[0002] The construction of autonomous driving maps involves multiple elements, with lane network generation being one of the core tasks. Its goal is to extract and organize structured geometric lane elements from multi-source perception data such as images and point clouds, including lane centerlines and lane boundary lines, to provide crucial foundational data support for autonomous driving and lane-level navigation. Furthermore, it must meet requirements for geometric accuracy, topological connectivity and legality, and consistency with mapping standards.

[0003] The current mainstream method for generating autonomous driving maps is to construct a BEV (Bird's-Eye View) representation through onboard multi-sensor or multi-view camera, or to construct a high-precision base map through point cloud fusion, and then use deep learning or rule-based methods to generate lane lines and topological elements. Finally, the map is generated by combining manual quality inspection and editing correction. Summary of the Invention

[0004] This disclosure provides a map generation method, apparatus, device, storage medium, and program product.

[0005] In a first aspect, embodiments of this disclosure propose a map generation method, comprising: constructing an input pair based on multi-view images, the input pair including the multi-view images and the current lane-level map; performing constraint review on the current lane-level map to generate structured evidence; performing action planning based on the input pair and the structured evidence to generate an action sequence; and performing editing and execution on the current lane-level map based on the action sequence to generate the final lane-level map.

[0006] Secondly, embodiments of this disclosure propose a map generation apparatus, comprising: a construction module configured to construct input pairs based on multi-view images, the input pairs including multi-view images and a current lane-level map; a review module configured to perform constraint review on the current lane-level map and generate structured evidence; a planning module configured to perform action planning based on the input pairs and structured evidence and generate an action sequence; and an execution module configured to edit and execute the current lane-level map based on the action sequence and generate a final lane-level map.

[0007] Thirdly, embodiments of this disclosure provide an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method as described in the first aspect.

[0008] Fourthly, embodiments of this disclosure provide a non-transitory computer-readable storage medium storing computer instructions for causing a computer to perform the methods described in the first aspect.

[0009] Fifthly, embodiments of this disclosure provide a computer program product including a computer program that, when executed by a processor, implements the method described in the first aspect.

[0010] The key or essential features of the embodiments disclosed herein are not intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0011] Other features, objects, and advantages of this disclosure will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings. The drawings are provided for a better understanding of the invention and are not intended to limit the scope of this disclosure. Wherein: Figure 1 This is a flowchart of an embodiment of the map generation method according to the present disclosure; Figure 2 This is a flowchart of yet another embodiment of the map generation method according to the present disclosure; Figure 3 It is a flowchart of closed-loop repair process for autonomous driving maps in a multi-agent collaborative manner; Figure 4 This is a schematic diagram of the structure of an embodiment of the map generation apparatus according to the present disclosure; Figure 5 This is a block diagram of an electronic device used to implement the map generation method or video retrieval method of the embodiments of this disclosure. Detailed Implementation

[0012] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0013] It should be noted that, unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other. This disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0014] In the disclosed embodiments, the key terms involved are defined as follows: Autonomous driving maps: Autonomous driving maps are generally defined as a high-precision electronic map dataset for intelligent vehicles. They not only record the basic road information in traditional electronic navigation maps, but also integrate lane-level geometry, traffic signs and markings, dynamic traffic events, environmental semantic rules and prior information for behavioral decisions through multi-dimensional and hierarchical expression. This provides autonomous driving systems with prior information and safety redundancy support that far exceed the real-time perception capabilities of onboard sensors.

[0015] BEV image (bird's-eye view): An image or feature representation obtained by converting multi-view vehicle-mounted camera images to a ground-level bird's-eye view through camera calibration, geometric transformation, and feature projection, used to uniformly express the spatial relationship between roads and lane lines.

[0016] Lane line elements: Vector objects representing lane boundaries or lane markings, usually consisting of a series of ordered points and polyline segments, and have attributes such as line type, color, and solid / darkness.

[0017] Lane topology: A description of the connections and turning relationships between lanes, including predecessor, successor, parallel, branching, merging, etc., often represented by a graph structure or adjacency relationship.

[0018] Multi-agent collaboration: a mechanism in which multiple agents with different responsibilities work together to complete a task, and the agents form a closed loop of division of labor and cooperation through information transmission.

[0019] Figure 1 A flow 100 of an embodiment of the map generation method according to this disclosure is shown. The map generation method includes the following steps: Step 101: Construct input pairs based on multi-view images.

[0020] In this embodiment, the entity executing the map generation method can construct input pairs based on multi-view images.

[0021] The multi-view images are captured by an onboard multi-view camera, covering a 360-degree scene around the vehicle and comprehensively capturing environmental information such as roads, lane lines, and traffic signs. The input pairs include the multi-view images and the current lane-level map. The initial lane-level map is the output of the lane-level vectorization prediction model, and subsequent iterations use updated lane-level maps.

[0022] In some embodiments, the input pair is constructed through the following steps: First, the multi-view images are projected and converted to a ground-level bird's-eye view to generate a bird's-eye view, which can be represented as follows: The projection conversion process involves camera calibration, geometric transformation, and feature projection to ensure that the bird's-eye view accurately reflects the spatial relationship between roads and lane lines.

[0023] Then, the bird's-eye view Inputting into a lane-level vectorized prediction model, the output is an initial lane-level map, which can be represented as: The lane-level vectorized prediction model is an artificial intelligence model, such as the DuMapNet network, which can output lane lines and lane network drafts after a single forward inference. Initial lane-level map. Composed of several lane line vector elements, it can be represented as Each lane line It includes geometric polyline point sequence and attribute information such as type, solid / dark, color, and lane category.

[0024] Finally, based on the bird's-eye view and the initial lane-level map The generated input pairs can be represented as This input is for... This provides basic data support for subsequent multi-agent collaborative processing.

[0025] Step 102: Perform constraint review on the current lane-level map to generate structured evidence.

[0026] In this embodiment, the aforementioned executing entity can perform constraint review on the current lane-level map and generate structured evidence.

[0027] Constraint review is performed by a review agent. The review agent is a multimodal large model fine-tuned based on data from various map scenarios. Map scenarios include, but are not limited to, highway scenarios, urban main road scenarios, urban secondary road scenarios, and special road scenarios. Through data from these map scenarios, the review agent forms its own thought process, enabling it to judge the correctness of the map status and provide corresponding reasons.

[0028] During the review process, the input to the review agent includes a pre-built set of constraints and the current lane-level map. The constraint set is pre-constructed based on large-scale data analysis of the corresponding map scene, breaking down map quality issues into a verifiable set of constraints: Constraint Set Includes category / attribute constraints Geometric constraints Structural / topological constraints Three categories. Among them, category / attribute constraints. This involves requirements such as lane line type and attribute consistency. Geometric constraints. It covers criteria such as smoothness, curvature anomalies, jitter, out-of-bounds errors, and overlap. Structural / topological constraints. Rules include connectivity validity, complete flow splitting and merging, broken / incorrect connections, redundancy / missing information, etc.

[0029] Structured evidence can be represented as .in, Error type labels include category errors, redundant lanes, geometric errors, structural / topological errors, etc. Locate specific elements, sections, and nodes. This represents the error confidence level. This represents the severity of the error. It provides recommended repair directions, such as modifying the type, removing redundancy, smoothing, and rebuilding. Furthermore, the evidence and repair suggestions are naturally aligned with the categories of subsequent deterministic editing tools, achieving a closed-loop connection from review to execution.

[0030] Step 103: Based on the input pairs and structured evidence, perform action planning to generate an action sequence.

[0031] In this embodiment, the aforementioned execution entity can perform action planning based on input pairs and structured evidence to generate an action sequence.

[0032] Action planning is performed by a planning agent. This planning agent is a large text model fine-tuned based on data from various map scenarios, capable of clearly defining how to correct map errors.

[0033] The inputs to the planning agent include input pairs Structured evidence This includes pre-defined rule bases, historical experience, and tool capability information. The rule base contains lane-level mapping specifications, topology validity constraints, and virtual lane insertion strategies. Historical experience covers historical correction patterns, typical errors, and corresponding repair strategies. Tool capability information includes a set of available editing tools, their parameter ranges, and applicable conditions.

[0034] The generated action sequence is a structured plan ordered by steps. Each action It should at least include the operation object (such as lane line ID, segment, topology relationship, etc.), operation type (such as category correction, deletion, smoothing, reconstruction, etc.), operation parameters (such as smoothing intensity, reconstruction range, etc.), and execution order and dependencies (such as deleting redundancy before smoothing, etc.). Furthermore, while meeting mapping specifications and topology validity, it prioritizes generating action chains with minimal changes to avoid excessive editing leading to new errors. The plan itself can be recorded, replayed, and audited, facilitating quality inspection, rollback, and continuous optimization in large-scale generation scenarios.

[0035] Step 104: Edit and execute the current lane-level map based on the action sequence to generate the final lane-level map.

[0036] In this embodiment, the aforementioned execution entity can edit and execute the current lane-level map based on the action sequence to generate the final lane-level map.

[0037] The editing process is completed by the executing agent. The executing agent is a collection of editing tools, which can be understood as the tools of the planning agent. It can call the corresponding tools to process according to the action plan, which is equivalent to the rule implementation of the planning problem.

[0038] The executing agent includes at least four types of deterministic editing tools: category correction tools, used to correct the category and attributes of erroneous elements (such as lane line type, solid / dark, color, lane category, etc.); deletion tools, used to delete redundant, repetitive, or obviously erroneous lane lines and segments; smoothing tools, used to perform geometric optimization such as skeletonization and smoothing on lane lines with geometric jitter, spikes, or discontinuities; and reconstruction tools, used to perform local reconstruction and completion of structural error scenarios such as missing lines, severe drift, and topological breaks, generating lane lines and connection relationships that meet the specifications.

[0039] During execution, the executing agent updates the current lane-level map based on the action sequence. Edit and modify the map to output the updated lane-level map. Then Returning to step 102 as the new current lane-level map, it undergoes constraint review and enters the next round of the "review-planning-execution" closed loop until the final lane-level map that meets the constraint verification is generated.

[0040] This disclosure provides a map generation method that completes a closed-loop process of "quality screening - constraint review - plan generation - tool execution - verification" through multi-agent collaboration. It does not replace the existing lane-level vectorized prediction model, but treats the model output as an editable map state. It automatically repairs category, geometric and topological structure problems within a limited number of iterations, and outputs a high-quality lane-level vector map that meets constraint verification. This reduces the cost of manual intervention and consistency risks, and enables large-scale and stable implementation.

[0041] Continue to refer to Figure 2 This illustrates a flow 200 of yet another embodiment of the map generation method according to the present disclosure. The map generation method includes the following steps: Step 201: Project the multi-view image to a ground-level bird's-eye view to generate a bird's-eye view.

[0042] In this embodiment, the aforementioned executing entity can convert multi-view image projection to a ground-level bird's-eye view to generate a bird's-eye view.

[0043] The multi-view images are captured by an onboard multi-view camera, covering a 360-degree scene around the vehicle and fully capturing environmental information such as roads, lane lines, and traffic signs. The projection conversion process involves camera calibration, geometric transformation, and feature projection to ensure that the bird's-eye view accurately reflects the spatial relationship between roads and lane lines, providing standardized view data support for subsequent lane-level map generation.

[0044] Step 202: Input the bird's-eye view into the lane-level vectorized prediction model and output the initial lane-level map.

[0045] In this embodiment, the aforementioned execution entity can input the bird's-eye view into the lane-level vectorized prediction model and output an initial lane-level map.

[0046] The lane-level vectorized prediction model is an artificial intelligence model, such as the DuMapNet network, which can output lane lines and lane network drafts after a single forward inference. Initial lane-level map. Composed of several lane line vector elements, it can be represented as Each lane line It includes geometric polyline point sequence and attribute information such as type, solid / dark, color, and lane category.

[0047] Step 203: Input the confidence level of the initial lane-level map into the filtering agent, and output the input pair or the final lane-level map.

[0048] In this embodiment, the aforementioned execution entity can input the confidence level of the initial lane-level map into the filtering agent and output the input pair or the final lane-level map.

[0049] The filtering agent is a piecewise function, with different confidence thresholds corresponding to different map scenarios. Map scenarios include, but are not limited to, highway scenarios, urban main road scenarios, urban secondary road scenarios, and special road scenarios. Through the piecewise function design, the filtering criteria can be flexibly adapted according to the characteristics of the scenario, ensuring the accuracy of the filtering.

[0050] In some embodiments, the filtering agent filters through the following steps: First, determine the confidence threshold corresponding to the map scene to which the multi-angle images belong. The confidence threshold is set based on scene complexity and quality requirements; for example, the threshold can be appropriately lowered for complex scenes and relatively higher for simple scenes. The confidence level of the initial lane-level map is then calculated. The score can be derived from the confidence score of the lane-level vectorized prediction model output, the result of the rule consistency check, or a combination of the two.

[0051] Then, if the confidence level of the initial lane-level map The confidence threshold corresponding to the map scene to which the multi-angle image belongs is less than Samples deemed low-confidence are marked as requiring correction. The bird's-eye view and the initial lane-level map are combined to generate input pairs. This leads to the core closed loop of the subsequent map-based intelligent agent system; if the confidence level of the initial lane-level map is... Not less than the confidence threshold corresponding to the map scene to which the multi-angle images belong. If the sample is identified as a high-confidence sample, fast passage is triggered, and the initial lane-level map is updated. As the final lane-level map is directly output without needing to enter a subsequent repair loop, this mechanism ensures that generation is initiated only for a small number of difficult samples, controlling the overall computational overhead and meeting the throughput requirements of city-scale operations.

[0052] Step 204: Input the current lane-level map into the review agent and output structured evidence or the final lane-level map.

[0053] In this embodiment, the aforementioned executing entity can input the current lane-level map into the reviewing agent and output structured evidence or the final lane-level map.

[0054] The review agent is a multimodal large model fine-tuned based on data from various map scenarios. These map scenarios include, but are not limited to, highway scenarios, urban main road scenarios, urban secondary road scenarios, and special road scenarios. Through data from these map scenarios, the review agent forms its own thought process, enabling it to judge the correctness of the map status and provide corresponding reasons.

[0055] In some embodiments, the reviewing agent reviews through the following steps: First, the current lane-level map is validated for errors using the constraint set corresponding to the map scene to which the multi-angle images belong. This constraint set is pre-constructed based on large-scale data analysis of the corresponding map scene, breaking down map quality issues into a set of verifiable constraints. Constraint Set Includes category / attribute constraints Geometric constraints Structural / topological constraints Three categories. Among them, category / attribute constraints. This involves requirements such as lane line type and attribute consistency. Geometric constraints. It covers criteria such as smoothness, curvature anomalies, jitter, out-of-bounds errors, and overlap. Structural / topological constraints. Rules include connectivity validity, complete flow splitting and merging, broken / incorrect connections, redundancy / missing information, etc.

[0056] Then, the input to the review agent includes the set of constraints. And the current lane-level map The current lane-level map is determined through constraint verification. Does it meet the quality requirements? If the current lane-level map... Upon verification, the current lane-level map will be updated. As the final lane-level map; if the current lane-level map The validation failed, and the number of iterations was not less than the maximum number of iterations. (For example, the maximum number of iterations can be set to 2 to avoid entering an infinite loop), select the historical lane-level map with the highest score as the final lane-level map; if the current lane-level map... The validation failed, and the number of iterations was less than the maximum number of iterations. It outputs structured evidence. This structured evidence includes bug reports, bug evidence, and remediation recommendations, and can be represented as follows: .in, Error type labels include category errors, redundant lanes, geometric errors, structural / topological errors, etc. Locate specific elements, sections, and nodes. This represents the error confidence level. This represents the severity of the error. It provides recommended repair directions, such as modifying the type, removing redundancy, smoothing, and rebuilding. Furthermore, the evidence and repair suggestions are naturally aligned with the categories of subsequent deterministic editing tools, achieving a closed-loop connection from review to execution.

[0057] Step 205: Input the input pairs and structured evidence into the planning agent and output the action sequence.

[0058] In this embodiment, the aforementioned execution entity can input input pairs and structured evidence into the planning agent and output an action sequence.

[0059] The planning agent is a large text model fine-tuned based on data from various map scenarios, capable of clearly defining how to correct map errors. The inputs to the planning agent include input pairs. Structured evidence This includes pre-defined rule bases, historical experience, and tool capability information. The rule base contains lane-level mapping specifications, topology validity constraints, and virtual lane insertion strategies. Historical experience covers historical correction patterns, typical errors, and corresponding repair strategies. Tool capability information includes a set of available editing tools, their parameter ranges, and applicable conditions.

[0060] The generated action sequence is a structured plan ordered by steps. Each action It should at least include the operation object (such as lane line ID, segment, topology relationship, etc.), operation type (such as category correction, deletion, smoothing, reconstruction, etc.), operation parameters (such as smoothing intensity, reconstruction range, etc.), and execution order and dependencies (such as deleting redundancy before smoothing, etc.). Furthermore, while meeting mapping specifications and topology validity, it prioritizes generating action chains with minimal changes to avoid excessive editing leading to new errors. The plan itself can be recorded, replayed, and audited, facilitating quality inspection, rollback, and continuous optimization in large-scale generation scenarios.

[0061] Step 206: Input the action sequence and the current lane-level map into the executing agent, and output the next lane-level map.

[0062] In this embodiment, the aforementioned executing agent can input the action sequence and the current lane-level map into the executing agent and output the next lane-level map.

[0063] The executing agent includes a set of editing tools, which can be understood as the tools of the planning agent. These tools are invoked according to the action plan, essentially fulfilling the rules of a planning problem. The executing agent contains at least four types of deterministic editing tools: category correction tools, used to correct the category and attributes of erroneous elements (such as lane line type, solid / dark, color, lane category, etc.); deletion tools, used to delete redundant, repetitive, or obviously erroneous lane lines and segments; smoothing tools, used to perform geometric optimization such as skeletonization and smoothing on lane lines with geometric jitter, spikes, or discontinuities; and reconstruction tools, used to locally reconstruct and complete structural error scenarios such as missing lines, severe drift, and topological breaks, generating lane lines and connection relationships that meet specifications.

[0064] During execution, the executing agent updates the current lane-level map based on the action sequence. Edit and modify the data, merging the changes from each tool into an updated lane-level map. This refers to the next lane-level map.

[0065] Step 207: Perform constraint review on the next lane-level map as the new current lane-level map until the final lane-level map is generated.

[0066] In this embodiment, the aforementioned executing entity can use the next lane-level map as the new current lane-level map for constraint review until the final lane-level map is generated.

[0067] This step forms a verification-driven iterative control mechanism. After each round of execution, the next lane-level map is sent back to the review agent for verification. If the verification passes, the corrected prediction is output; if it fails and the maximum number of iterations has not been reached, the prediction is not updated. Then proceed to the next round of "review-planning-implementation" revisions; if it fails to pass and reaches the iteration limit, it will be revisited from the historical lane-level map. The system selects the lane-level map with the highest review agent score as the output to prevent the system from getting worse with each repair, and to ensure system stability and cost control.

[0068] This disclosure provides a map generation method that refines the process of multi-agent collaboration, clarifies the core functions and interaction logic of each agent, and achieves automated and high-precision generation of lane-level maps. It effectively solves the problems of frequent structural errors, high labor costs, and difficulty in ensuring consistency in the prior art, and provides reliable basic data support for autonomous driving.

[0069] Further reference Figure 3 The document illustrates a flowchart of a closed-loop repair process for autonomous driving maps using multi-agent collaboration. This flowchart clearly demonstrates the core logic of the disclosed technical solution, forming a complete closed loop from the input stage, the core map agent system, to the output stage. The specific process is as follows: Input Phase): The core input consists of input pairs, including a BEV bird's-eye view and an initial lane-level map. The initial lane-level map is generated by a lane-level vectorization prediction model based on the BEV bird's-eye view. Figure 1 The result is obtained through forward reasoning and includes the sequence of geometric polylines and attribute information of lane lines.

[0070] The core map agent system comprises filtering agents, review agents, planning agents, execution agents, and knowledge and capability source modules. These agents collaborate to complete map repair. The filtering agent performs early screening based on confidence levels. It uses a piecewise function to set confidence thresholds for different map scenarios and scores the quality of the initial lane-level map. High-confidence samples trigger early acceptance and directly enter the output stage; low-confidence samples are marked as needing correction and proceed to the subsequent review stage.

[0071] The review process involves performing constraint verification and quality checks on the intelligent agent, loading the constraint set for the corresponding scenario, and reviewing the current lane-level map. Perform error validation and output structured evidence, if the current lane-level map... If the constraints are met, proceed directly to the output stage; otherwise, pass the structured evidence to the planning agent.

[0072] After receiving structured evidence, the planning agent combines the memory, rules, tools, and model information from the knowledge and capability source modules to perform constrained reasoning and plan generation, outputting an ordered action plan that clarifies the object, type, parameters, and execution order of the repair actions.

[0073] The agent performs deterministic editing based on tools, invoking four types of editing tools: category correction, deletion, smoothing, and reconstruction. According to the action plan, it makes predictable and reproducible modifications to the current map state and generates the updated state.

[0074] The knowledge and capability source module provides decision support for reviewing and planning agents, storing memories, rules, tools, and models. Agents can query and access these sources.

[0075] Output phase: After processing by the core map intelligent agent system, a lane-level map that has passed quality verification is output; if the verification is still not passed after reaching the maximum number of iterations, the lane-level map with the highest score is selected from the historical lane-level maps as the final output to ensure output quality.

[0076] In addition, the core map intelligent agent system also sets up a retry loop and an iteration counter, and sets a maximum number of iterations (e.g., 2 times). Through iteration control, it ensures that the computational cost of complex samples has an upper bound, avoids entering an infinite loop, and realizes a controllable and terminateable repair process.

[0077] Further reference Figure 4 As an implementation of the methods shown in the above figures, this disclosure provides an embodiment of a map generation apparatus, which is similar to... Figure 1 Corresponding to the method embodiments shown, this device can be specifically applied to various electronic devices.

[0078] like Figure 4 As shown, the map generation device 400 of this embodiment may include: a construction module 401, a review module 402, a planning module 403, and an execution module 404. The construction module 401 is configured to construct input pairs based on multi-view images, the input pairs including multi-view images and the current lane-level map; the review module 402 is configured to perform constraint review on the current lane-level map and generate structured evidence; the planning module 403 is configured to perform action planning based on the input pairs and structured evidence and generate an action sequence; the execution module 404 is configured to edit and execute the current lane-level map based on the action sequence to generate the final lane-level map.

[0079] In this embodiment, the specific processing of the construction module 401, review module 402, planning module 403, and execution module 404 in the map generation device 400, and the resulting technical effects, can be referred to respectively. Figure 1 The relevant descriptions of steps 101-104 in the corresponding embodiments will not be repeated here.

[0080] In some optional implementations of this embodiment, the construction module 401 is further configured to: project and convert the multi-view image to a ground bird's-eye view to generate a bird's-eye view; input the bird's-eye view into a lane-level vectorized prediction model to output an initial lane-level map; and generate an input pair based on the bird's-eye view and the initial lane-level map.

[0081] In some optional implementations of this embodiment, the construction module 401 is further configured to: input the confidence of the initial lane-level map to the filtering agent, and output the input pair or the final lane-level map. The filtering agent is a piecewise function, and different map scenarios correspond to different confidence thresholds.

[0082] In some optional implementations of this embodiment, the construction module 401 is further configured to: determine the confidence threshold corresponding to the map scene to which the multi-angle image belongs; if the confidence of the initial lane-level map is less than the confidence threshold corresponding to the map scene to which the multi-angle image belongs, combine the bird's-eye view and the initial lane-level map to generate an input pair; if the confidence of the initial lane-level map is not less than the confidence threshold corresponding to the map scene to which the multi-angle image belongs, use the initial lane-level map as the final lane-level map.

[0083] In some optional implementations of this embodiment, the review module 402 is further configured to: input the current lane-level map to the review agent and output structured evidence or the final lane-level map, wherein the review agent is a multimodal large model based on data fine-tuning of various map scenarios.

[0084] In some optional implementations of this embodiment, the review module 402 is further configured to: perform error verification on the current lane-level map using the constraint set corresponding to the map scene to which the multi-angle images belong; if the current lane-level map passes the verification, the current lane-level map is used as the final lane-level map; if the current lane-level map fails the verification and the number of iterations is not less than the maximum number of iterations, the historical lane-level map with the highest score is selected as the final lane-level map; if the current lane-level map fails the verification and the number of iterations is less than the maximum number of iterations, structured evidence is output, including error reports, error evidence, and repair suggestions.

[0085] In some optional implementations of this embodiment, the planning module 403 is further configured to: input input pairs and structured evidence into the planning agent and output an action sequence, wherein the planning agent is a large text model fine-tuned based on data from various map scenarios.

[0086] In some optional implementations of this embodiment, the execution module 404 is further configured to: input the action sequence and the current lane-level map to the execution agent, output the next lane-level map, the execution agent including a set of editing tools; and perform constraint review on the next lane-level map as a new current lane-level map until the final lane-level map is generated.

[0087] The acquisition, storage, and application of user personal information involved in the technical solution disclosed herein comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0088] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0089] Figure 5 A schematic block diagram of an example electronic device 500 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0090] like Figure 5 As shown, device 500 includes a computing unit 501, which can perform various appropriate actions and processes based on a computer program stored in read-only memory (ROM) 502 or a computer program loaded from storage unit 508 into random access memory (RAM) 503. RAM 503 may also store various programs and data required for the operation of device 500. The computing unit 501, ROM 502, and RAM 503 are interconnected via bus 504. Input / output (I / O) interface 505 is also connected to bus 504.

[0091] Multiple components in device 500 are connected to I / O interface 505, including: input unit 506, such as keyboard, mouse, etc.; output unit 507, such as various types of monitors, speakers, etc.; storage unit 508, such as disk, optical disk, etc.; and communication unit 509, such as network card, modem, wireless transceiver, etc. Communication unit 509 allows device 500 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0092] The computing unit 501 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 501 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 501 performs the various methods and processes described above, such as map generation methods. For example, in some embodiments, the map generation method may be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 508. In some embodiments, part or all of the computer program may be loaded and / or installed on device 500 via ROM 502 and / or communication unit 509. When the computer program is loaded into RAM 503 and executed by the computing unit 501, one or more steps of the map generation method described above may be performed. Alternatively, in other embodiments, the computing unit 501 may be configured to perform map generation methods by any other suitable means (e.g., by means of firmware).

[0093] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0094] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0095] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0096] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0097] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0098] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, distributed system servers, or servers incorporating blockchain technology.

[0099] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution provided in this disclosure can be achieved, and this is not limited herein.

[0100] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A map generation method, comprising: Based on multi-view images, an input pair is constructed, the input pair including the multi-view images and the current lane-level map; Perform constraint review on the current lane-level map to generate structured evidence; Based on the input pair and the structured evidence, action planning is performed to generate an action sequence; Based on the action sequence, the current lane-level map is edited and executed to generate the final lane-level map.

2. The method according to claim 1, wherein, The construction of input pairs based on multi-view images includes: The multi-view images are projected and converted to a ground-level bird's-eye view to generate a bird's-eye view; The bird's-eye view is input into the lane-level vectorized prediction model to output an initial lane-level map; The input pair is generated based on the bird's-eye view and the initial lane-level map.

3. The method according to claim 2, wherein, The process of generating the input pair based on the bird's-eye view and the initial lane-level map includes: The confidence level of the initial lane-level map is input into the filtering agent, and the input pair or the final lane-level map is output. The filtering agent is a piecewise function, and different confidence thresholds correspond to different map scenarios.

4. The method according to claim 3, wherein, The step of inputting the confidence level of the initial lane-level map into the filtering agent and outputting the input pair or the final lane-level map includes: Determine the confidence threshold corresponding to the map scene to which the multi-angle images belong; If the confidence level of the initial lane-level map is less than the confidence level threshold corresponding to the map scene to which the multi-angle image belongs, the bird's-eye view and the initial lane-level map are combined to generate the input pair; If the confidence level of the initial lane-level map is not less than the confidence threshold corresponding to the map scene to which the multi-angle image belongs, the initial lane-level map will be used as the final lane-level map.

5. The method according to claim 2, wherein, The constraint review of the current lane-level map to generate structured evidence includes: The current lane-level map is input into the review agent, which outputs the structured evidence or the final lane-level map. The review agent is a multimodal large model that is fine-tuned based on data from various map scenarios.

6. The method according to claim 5, wherein, The step of inputting the current lane-level map into the review agent and outputting the structured evidence or the final lane-level map includes: The current lane-level map is validated using the constraint set corresponding to the map scene to which the multi-angle images belong; If the current lane-level map passes verification, the current lane-level map will be used as the final lane-level map; If the current lane-level map fails the verification and the number of iterations is not less than the maximum number of iterations, the historical lane-level map with the highest score is selected as the final lane-level map. If the current lane-level map fails verification and the number of iterations is less than the maximum number of iterations, the structured evidence is output, which includes error reports, error evidence, and repair suggestions.

7. The method according to claim 1, wherein, The step of planning actions based on the input pairs and the structured evidence to generate action sequences includes: The input pair and the structured evidence are input into the planning agent, which outputs the action sequence. The planning agent is a large text model fine-tuned based on data from various map scenarios.

8. The method according to claim 1, wherein, The step of editing the current lane-level map based on the action sequence to generate the final lane-level map includes: The action sequence and the current lane-level map are input to the executing agent, and the next lane-level map is output. The executing agent includes a set of editing tools. The next lane-level map is used as the new current lane-level map for constraint review until the final lane-level map is generated.

9. A map generation apparatus, comprising: The construction module is configured to build input pairs based on multi-view images, the input pairs including the multi-view images and the current lane-level map; The review module is configured to perform a constraint review on the current lane-level map and generate structured evidence; The planning module is configured to perform action planning based on the input pair and the structured evidence, and generate an action sequence; The execution module is configured to edit and execute the current lane-level map based on the action sequence to generate the final lane-level map.

10. An electronic device, comprising: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-8.

11. A non-transitory computer-readable storage medium storing computer instructions for causing the computer to perform the method of any one of claims 1-8.

12. A computer program product comprising a computer program that, when executed by a processor, implements the method according to any one of claims 1-8.

Images