Search

CN-122023588-A - High-precision map generation system and method based on large model

CN122023588ACN 122023588 ACN122023588 ACN 122023588ACN-122023588-A

Abstract

The application discloses a high-precision map generation system and method based on a large model, and relates to the field of high-precision map generation, comprising a multi-agent simulation verification ring, a virtual agent generation system and a virtual agent generation system, wherein the multi-agent simulation verification ring comprises a high-fidelity simulation environment based on nerve rendering and a group of virtual agents with differentiated driving strategies; the multi-agent simulation verification ring is used for receiving initially generated high-precision map data, driving the virtual agents to conduct preset mileage driving comprising edge use cases in the simulation environment, collecting interaction conflict events and track smoothness indexes of the virtual agents and the environment, and generating a targeted optimization instruction, and the map utility elastic supply module is used for solving the technical problems that in the prior art, the high-precision map generation process lacks an active safety verification mechanism, cannot be elastically supplied according to user requirements, hidden safety defects in a map are difficult to mine and correct, and the overall resource utility is not optimal due to isolation and optimization of each link of a system.

Inventors

  • ZHANG KUNTING
  • ZHU LIJUN
  • Peng Baolei
  • ZHAO XIAOMIN

Assignees

  • 信动互联(北京)科技有限公司

Dates

Publication Date
20260512
Application Date
20251226

Claims (10)

  1. 1. A large model-based high-precision map generation system, comprising: The multi-agent simulation verification ring is used for receiving initially generated high-precision map data, driving the virtual agent to perform preset mileage driving comprising edge use cases in the simulation environment, collecting interaction conflict events and track smoothness indexes of the virtual agent and the environment, and generating a targeted optimization instruction; the map utility elastic supply module maintains a demand feature library containing mapping relations of different vehicle models, task types and driving styles, and dynamically predicts regional map element demand intensity spectrums according to real-time accessed vehicle request flows; The federal causality discovery and safety enhancement unit is used for locally anonymizing and analyzing the potential causality association between a collision event, a take-over request or a path planning concussion and a currently used map in the actual driving process at a vehicle end and generating an encrypted local causality map for uploading; The full-period game optimizer of system resources and model parameters is used for modeling map generation precision, system response delay, calculation communication cost and safety indexes output by the multi-agent simulation verification ring into a multi-target game framework, and dynamically adjusting the resource allocation strategy and the model super parameters of each module through online learning.
  2. 2. The system of claim 1, wherein the virtual agents of the differentiated driving strategy in the multi-agent simulation verification loop at least comprise aggressive agents, conservative agents, fatigue driving agents and abnormal vehicle dynamics model agents, and wherein the analysis of the interaction conflict event comprises comparing a decision trajectory of the virtual agents with a standard passable area provided by a map to identify areas with fuzzy map semantic labels, geometric boundary conflicts or missing topological connections.
  3. 3. The system of claim 1, wherein the prediction method of the demand intensity spectrum in the map utility elastic supply module is to cluster feature vectors of a request vehicle, and predict demand intensity values of specific map elements by using a graph neural network in combination with real-time traffic states and weather, wherein the feature vectors comprise vehicle type sizes, task types and historical rapid acceleration and deceleration frequencies, and the specific map elements comprise lane grade gradient curvature, precise height of a curb or temporary topology of a construction area.
  4. 4. The system of claim 1, wherein the workflow of the federal causal discovery and security enhancement unit comprises: The vehicle starts a lightweight counterfactual reasoning model when a preset type of safety related event occurs based on the local sensor data and the received map, and whether the current event can be avoided if the map elements are different is deduced; Converting the reasoning result into an anonymized local causal graph segment taking map elements as nodes and causal influence as edges; Cloud aggregation is carried out on the local causal graph fragments from a large number of vehicles, and a stable causal path with global property between map data and bad driving events is excavated through statistical significance inspection and causal structure learning.
  5. 5. The system of claim 1 wherein the full cycle game optimizer treats modules within the map generation system as game participants, each participant's strategic space is a selectable combination of computation accuracy, time and energy consumption operating points, the optimization objective is to maximize the overall map quality overall index and minimize the overall resource consumption, solve for nash equilibrium points by multi-agent deep reinforcement learning, and dynamically configure the system accordingly.
  6. 6. A large model-based high-precision map generation method applied to the large model-based high-precision map generation system according to any one of claims 1 to 5, the method comprising: S1, generating an initial high-precision map draft based on multi-source perception data; S2, inputting the map draft into the multi-agent simulation verification ring, and actively exposing and positioning potential defect areas possibly causing safety or experience problems in the map through high-intensity simulation driving of diversified virtual agents; s3, analyzing a real-time vehicle request, and dynamically generating a personalized map data packet which is matched with the current regional vehicle group demand and has optimal resources through the map utility elastic supply module; s4, utilizing the federal causal discovery and safety enhancement unit to mine deep causal relation between map data and driving behavior results from non-invasive data of massive real vehicles, and carrying out safety enhancement iteration on the map; And S5, continuously coordinating balance among simulation verification intensity, map supply quality, safety mining depth and system resource consumption through the full-period game optimizer.
  7. 7. The method according to claim 6, wherein in step S2, the multi-agent simulation verification loop uses a course learning strategy, and the multi-agent simulation verification loop uses a conventional driving agent to perform extensive testing in an initial stage, and then gradually introduces more extreme scenes and edge use case agents to perform concentrated stress testing on the identified weak areas.
  8. 8. The method according to claim 6, wherein in step S3, when the personalized map data packet is generated, lossless encoding and preferential transmission are adopted for elements with high demand intensity, and lossy compression or delayed update strategies are adopted or high-order semantic descriptions are adopted to replace accurate geometric data for elements with low demand intensity and limited resources.
  9. 9. The method according to claim 6, wherein the mining of deep causal links in step S4 includes, when a hidden security defect of a certain type is found, not only correcting the defect, but triggering targeted fine-tuning of the large model with respect to the targeted data collection task for the same type of scenario.
  10. 10. The method of claim 6, wherein in step S5, the full-cycle game optimizer dynamically adjusts utility function weights for games based on external conditions, including network congestion level, server load, or battery level.

Description

High-precision map generation system and method based on large model Technical Field The invention relates to the field of high-precision map generation, in particular to a high-precision map generation system and method based on a large model. Background The high-precision map is used as a core infrastructure of the automatic driving system, and the accuracy, the instantaneity and the safety of the high-precision map directly determine the perception, planning and control performance of the intelligent vehicle. The traditional high-precision map construction is highly dependent on professional acquisition vehicles and post-processing, and has the inherent problems of high cost and lagging updating. In recent years, with the development of large models and multi-mode sensing technologies, the generation and updating of high-precision maps in real time or near real time by using vehicle-mounted sensor data has become an important technical evolution direction, and the aim of realizing the fresh activation of map data and crowdsourcing of production modes is achieved. In the prior art, a high-precision map generation scheme based on a large model is mainly focused on improving the precision of perception and reconstruction. For example, by fusing multi-source data such as vision and laser radar, a depth learning model is used to directly output vectorized map elements. Further improvements include introducing timing information to promote stability of element tracking, or enhancing understanding of complex scenes using semantic segmentation techniques. At the system level, existing schemes typically follow a linear flow of "sense-fuse-drawing" and pursue map data that provides a uniform precision standard for all users. The prior art still has significant drawbacks. Firstly, map generation quality is seriously dependent on offline labeling data and model training, an active and closed-loop verification mechanism is lacked, and driving risks possibly caused by map data in extreme or long-tail scenes cannot be systematically exposed before deployment. Second, map data feeds are static and "cut-away", and the differentiated demands of different vehicles, different tasks, on map elements are not considered, resulting in the waste of precious communications and computing resources in transmitting and processing non-critical information. Moreover, the existing method is difficult to effectively mine hidden safety defects which may exist in map data and are difficult to find through conventional perception from mass real driving feedback. Finally, key links such as map generation precision, security verification strength, system resource consumption and the like are often optimized in an isolated manner, and a global collaborative framework is lacked to realize the optimal balance of the overall utility. Disclosure of Invention Based on the above, the invention aims to provide a high-precision map generation system and method based on a large model, so as to solve the technical problems that in the prior art, a high-precision map generation process lacks an active safety verification mechanism, the map cannot be flexibly supplied according to the needs of users, hidden safety defects in a map are difficult to mine and correct, and the effectiveness of the whole resource is not optimal due to isolation and optimization of each link of the system. In order to achieve the aim, the invention provides a high-precision map generation system based on a large model, which comprises a multi-agent simulation verification ring, a virtual agent generation system and a control system, wherein the multi-agent simulation verification ring comprises a high-fidelity simulation environment based on nerve rendering and a group of virtual agents with differentiated driving strategies; the multi-agent simulation verification ring is used for receiving initially generated high-precision map data, driving the virtual agent to perform preset mileage driving comprising edge use cases in the simulation environment, collecting interaction conflict events and track smoothness indexes of the virtual agent and the environment, generating a targeted optimization instruction, a map utility elastic supply module maintaining a demand feature library comprising mapping relations of different vehicle models, task types and driving styles, dynamically predicting regional map element demand intensity spectrums according to real-time accessed vehicle request flows, elastically adjusting the depicting granularity, updating frequency and physical precision of corresponding elements in the map data according to the demand intensity spectrums and current resources of the system, a federal discovery and safety enhancement unit for locally anonymously analyzing potential association between a collision event, a take-over request or path planning concussion and a currently used map at a vehicle end and generating an encrypted local causal map, the federal disco