Search

CN-121766149-B - Intelligent processing method for automobile data

CN121766149BCN 121766149 BCN121766149 BCN 121766149BCN-121766149-B

Abstract

The invention discloses an intelligent processing method for automobile data, and relates to the technical field of automatic driving. The method comprises the steps of S1, constructing a simulation model, S2, determining a test case, namely, setting at least one antagonism test intelligent agent in the dynamic parallel simulation model, performing interactive test through the antagonism test intelligent agent, and determining an effective test case, S3, performing case verification, namely, acquiring a corresponding comprehensive score according to a test result of the effective test case, screening the effective test case according to the comprehensive score, and performing live-action verification on the screened effective test case. Therefore, closed loop iterative optimization of simulation-live-action is formed, further, the improvement of the performance of an automatic driving system can be continuously promoted, and the large-scale safe application is accelerated.

Inventors

  • LI TINGTING

Assignees

  • 成都工业职业技术学院

Dates

Publication Date
20260508
Application Date
20260303

Claims (6)

  1. 1. An intelligent processing method for automobile data is characterized by comprising the following steps: S1, constructing a simulation model, namely constructing a dynamic parallel simulation model by multi-source data acquired by an automatic driving motorcade, wherein the simulation model comprises the following steps: The method comprises the following steps of S1.1, data acquisition, wherein a fixed node is arranged on a running path of an intelligent vehicle, a mobile node is arranged on the intelligent vehicle, and multi-source data in the running process of the intelligent vehicle are acquired and acquired through the fixed node and the mobile node, wherein the multi-source data comprise perception data, dynamic data, pose data and scene data; S1.2, environment reconstruction, namely carrying out data association and alignment processing on the perception data to obtain fusion perception data, fusing the fusion perception data with a nerve representation model to construct an implicit 3D model, and simultaneously injecting the dynamic data, the pose data and the scene data into the implicit 3D model to construct a dynamic parallel simulation model; S2, determining a test case, namely setting at least one antagonism test intelligent agent in the dynamic parallel simulation model, and performing interactive test through the antagonism test intelligent agent to determine an effective test case, wherein the method comprises the following steps: S2.1, setting an intelligent agent, namely generating at least one controllable vehicle model through simulation software, and importing the controllable vehicle model into a dynamic parallel simulation model for countermeasure training to obtain a countermeasure test intelligent agent, wherein the method comprises the following steps of: S2.1.1, setting a countermeasure target, namely monitoring the behaviors of the test agents in the dynamic parallel simulation model through a set reward function to obtain an operation total reward score of each test agent, obtaining a core reward score and a constraint punishment score in the simulation deduction process of the test agents through a simulation engine in the dynamic parallel simulation model, and determining the operation total reward score through the core reward score and the constraint punishment score, wherein the core reward score comprises an uncertainty score of a maximized host vehicle, a safety margin reward score of a minimized host vehicle and a provoking rule failure score, and the constraint punishment score comprises a collision punishment score and a traffic rule punishment score; S2.1.2 training the intelligent agent, namely loading corresponding reconstruction scenes in the dynamic parallel simulation model according to multi-source data, carrying out simulation deduction on a main vehicle and the intelligent agent under test in each reconstruction scene, and carrying out training test on the intelligent agent under test through a deep reinforcement learning algorithm to obtain the intelligent agent under resistance test; S2.2, monitoring records, namely setting an automatic driving algorithm to be tested on a cloud platform of a dynamic parallel simulation model through a simulation task scheduling cluster, injecting the contrast test intelligent agent as a background traffic flow into the dynamic parallel simulation model for interactive test, and acquiring a test result of each test case through a set failure standard; And S3, verifying the use case, namely acquiring a corresponding comprehensive score according to the test result of the effective test case, screening the effective test case according to the comprehensive score, and carrying out live-action verification on the screened effective test case.
  2. 2. The method for intelligent processing of vehicle data according to claim 1, wherein obtaining the total operating reward score for each test agent comprises: S2.1.1.1 track monitoring, namely setting a probability value of each simulation track according to the simulation track of the test intelligent agent by a prediction module of the host vehicle, and determining the change amount of the track entropy value according to the probability value of the simulation track; S2.1.1.2, collision monitoring, namely determining the collision time between the host vehicle and the test intelligent body through a safety core module of the host vehicle, and determining the change amount of the danger coefficient according to the collision time; s2.1.1.3 brake monitoring, namely comparing a brake response corresponding to a planning control module of the main vehicle with a set comfort threshold value, and determining a regulation failure value according to a comparison result, wherein the method specifically comprises the following steps: when the deceleration of the braking reaction is larger than the comfort threshold, the regulation failure value is a preset regulation failure value, otherwise, the regulation failure value is 0; S2.1.1.4, result monitoring, namely detecting behaviors of the host vehicle and the test agent after simulation deduction is finished through a simulation engine, and setting a collision penalty score and a traffic rule penalty score according to detection results; S2.1.1.5 determining a total rewarding score by combining the track entropy change, the risk coefficient change, the regulation failure value, the collision punishment score and the traffic rule punishment score.
  3. 3. The method for intelligent processing of vehicle data according to claim 2, wherein when a physical collision occurs to the host vehicle and/or the test agent, the collision penalty score is a preset collision penalty threshold, and conversely, the collision penalty score is 0; When the host vehicle and/or the test agent violates the traffic rule, the traffic rule penalty score is a preset traffic rule penalty threshold, otherwise, the traffic rule penalty score is 0.
  4. 4. The intelligent processing method of automobile data according to claim 1, wherein a recording result of a host vehicle failure moment is determined according to the failure standard, and the recording result is matched with a vulnerability type, a severity level and a scene label to obtain a vulnerability report of an oppositional test agent.
  5. 5. The intelligent processing method of automobile data according to claim 1, wherein the test results of the effective test cases are mapped into a risk matrix, the comprehensive score of each effective test case is determined through the risk matrix, the effective test cases are arranged in a descending order according to the comprehensive score, and meanwhile a preset number of effective test cases are screened out in the descending order according to preset number requirements.
  6. 6. The intelligent processing method of automobile data according to claim 5, wherein the risk matrix is constructed according to occurrence probability and severity of test cases, and the occurrence probability score and severity score are set to be 1-5 points.

Description

Intelligent processing method for automobile data Technical Field The invention relates to the technical field of automatic driving, in particular to an intelligent processing method for automobile data. Background With the rapid development of the automobile industry and the continuous promotion of intelligent traffic systems, modern automobiles have gradually evolved into mobile intelligent terminals integrating sensing, computing, communication and control. The wide application of vehicle-mounted sensors (such as radar, cameras, laser radar, GPS, IMU and the like), vehicle-mounted infotainment systems and vehicle networking (V2X) devices on vehicles enables the vehicles to acquire massive multi-source heterogeneous data in real time during running, including vehicle state data (such as vehicle speed, rotating speed, oil consumption, battery state and the like), environment perception data (such as road conditions, pedestrian recognition, traffic sign recognition and the like), driving behavior data, position and track information and the like. The data not only provides basic support for advanced functions such as automatic driving, intelligent cabin, remote diagnosis, fleet management and the like, but also creates important conditions for optimizing product design, improving service quality, carrying out big data analysis and artificial intelligent modeling for a fleet. However, in the face of increasing data size, diversified data types, and high requirements for real-time performance, accuracy, and safety, it has been difficult for the conventional automobile data processing method to meet the actual application requirements. The Chinese patent application with publication number of CN117666785A discloses an automatic driving man-machine interaction takeover training method and system based on digital twinning, which comprises the steps of firstly acquiring a plurality of static 2D images through a camera on multiple angles of a modeling object, estimating camera internal and external parameters corresponding to each image through COLMAP and other tools, providing a segmentation prompt for a target object on each image, segmenting the image by utilizing a SAM large model to remove background noise to obtain a segmented image of the target object, inputting the segmented image and camera parameters into a NeRF neural radiation field algorithm to perform three-dimensional reconstruction and rendering, and converting an implicit three-dimensional model into a display three-dimensional model by utilizing a point cloud reconstruction technology so as to obtain a digital twinning geometric model with high precision. The method can effectively utilize the powerful semantic understanding capability of the SAM large model and the high-efficiency three-dimensional reconstruction capability of NeRF nerve radiation field algorithm, and realize the automation and the accuracy of digital twin modeling. The technical scheme can realize the automation and the accuracy of digital twin modeling, but the construction of the virtual environment depends on the data acquisition and the reconstruction of the generated real scene. And automatic driving is high in real-way test cost and has risks, so that most of the automatic driving is tested and verified through a simulation environment, and the real scene data acquisition difficulty is high. Meanwhile, due to the fact that a real gap which is difficult to eliminate exists between the simulation environment and the real world, an automatic driving algorithm trained and verified in the simulation is difficult to directly migrate into the real world. Once deployed in a real environment, performance degradation and even decision failure may occur due to imperceptible small differences or long tail scenes, and even constitute a core technical obstacle for large-scale application. Disclosure of Invention The invention aims to provide an intelligent processing method for automobile data, which aims to solve the problems in the background technology. In order to achieve the purpose, the invention provides the following technical scheme that the intelligent processing method for the automobile data comprises the following steps: S1, constructing a simulation model, namely constructing a dynamic parallel simulation model through multi-source data acquired by an automatic driving motorcade; S2, determining a test case, namely setting at least one antagonism test intelligent agent in the dynamic parallel simulation model, and performing interactive test through the antagonism test intelligent agent to determine an effective test case, wherein the method comprises the following steps: S2.1, setting an intelligent agent, namely generating at least one controllable vehicle model through simulation software, and importing the controllable vehicle model into a dynamic parallel simulation model for countermeasure training to obtain a countermeasure test intelligent agent; S2.2, monitoring