Search

CN-121989960-A - Path evaluation method, storage medium and vehicle

CN121989960ACN 121989960 ACN121989960 ACN 121989960ACN-121989960-A

Abstract

The application provides a path evaluation method, a storage medium and a vehicle, which are applied to the technical field of intelligent driving assistance and comprise the steps of acquiring multi-dimensional vehicle state sensing data of the vehicle in the current driving environment, wherein the vehicle state sensing data is used for representing the internal running state and/or external environment information of the vehicle; and determining a traffic state prediction result of the vehicle in the current driving environment according to the target semantic features, the ground surface inclination angle value and the vehicle state sensing data, so that the traffic path is evaluated according to the traffic state prediction result. The method can realize the deep fusion of visual perception and physical parameters, so that the dynamic physical state under the complex terrain condition is comprehensively analyzed, and the accuracy of path stability evaluation under the off-road scene is improved.

Inventors

  • LI MING

Assignees

  • 长城汽车股份有限公司

Dates

Publication Date
20260508
Application Date
20260320

Claims (10)

  1. 1. A method of path evaluation, the method comprising: Acquiring multi-dimensional vehicle state sensing data of a vehicle in a current driving environment, wherein the vehicle state sensing data is used for representing internal running state and/or external environment information of the vehicle; determining target semantic features of a terrain in the current driving environment based on visual images in the vehicle state awareness data, and determining a surface inclination value of the vehicle passing path based on the visual images, wherein the target semantic features are used for representing the ground attachment performance of the terrain; And determining a traffic state prediction result of the vehicle in the current driving environment according to the target semantic features, the ground inclination angle value and the vehicle state sensing data, so that the traffic path is evaluated according to the traffic state prediction result.
  2. 2. The method of claim 1, wherein the determining target semantic features of terrain in the current driving environment based on visual images in the vehicle state awareness data comprises: inputting the visual image in the vehicle state sensing data to a friction coefficient self-supervision learning model, and controlling the friction coefficient self-supervision learning model to output target semantic features of the terrain in the current driving environment aiming at the visual image; the training process of the friction coefficient self-supervision learning model comprises the following steps: Constructing an initial friction coefficient self-supervision learning model, and acquiring a plurality of sample visual images of a sample vehicle in a driving environment; inputting each sample visual image into the initial friction coefficient self-supervision learning model, and training the initial friction coefficient self-supervision learning model; and in the training process of the initial friction coefficient self-supervision learning model, controlling the initial friction coefficient self-supervision learning model to conduct contrast learning on each sample visual image, and adjusting parameters of the initial friction coefficient self-supervision learning model according to a contrast learning result until the initial friction coefficient self-supervision learning model converges to obtain a trained friction coefficient self-supervision learning model.
  3. 3. The method of claim 2, wherein the determining, based on the visual image, a ground tilt value of the vehicle traffic path comprises: the corresponding stereoscopic structure perceived image is fused in the visual image, a second visual image is obtained, the second visual image is input into a ground surface inclination angle estimation model, and the ground surface inclination angle estimation model is controlled to output a ground surface inclination angle value of the vehicle passing path aiming at the second visual image; the training process of the earth surface dip angle estimation model comprises the following steps: constructing an initial earth surface inclination angle estimation model, and acquiring a plurality of second sample visual images of the sample vehicle in the driving environment, wherein each second sample visual image is a sample visual image fused with a sample three-dimensional structure perceived image, and each second sample visual image is provided with a standard earth surface inclination angle label; inputting each second sample visual image into the initial earth surface dip angle estimation model, and training the initial earth surface dip angle estimation model; And in the training process of the initial earth surface dip angle estimation model, controlling the initial earth surface dip angle estimation model to output predicted earth surface dip angle labels aiming at each second sample visual image, and adjusting parameters of the initial earth surface dip angle estimation model according to each predicted earth surface dip angle label and standard earth surface dip angle labels of each second sample visual image until the initial earth surface dip angle estimation model converges to obtain a trained earth surface dip angle estimation model.
  4. 4. A method according to claim 3, wherein said determining a traffic state prediction result of said vehicle in said current driving environment from said target semantic features, said surface dip value and said vehicle state awareness data comprises: Inputting the target semantic features, the ground surface inclination angle values and the vehicle state sensing data into a traffic state classification model, and controlling the traffic state classification model to output a traffic state prediction result of the vehicle in the current driving environment, wherein the traffic state prediction result comprises various traffic states and corresponding probability confidence degrees; the training process of the traffic state classification model comprises the following steps: Constructing an initial traffic state classification model, and acquiring a plurality of sample state data of the sample vehicle in the driving environment, wherein each sample state data at least comprises a sample target semantic feature output by the initial friction coefficient self-supervision learning model, a predicted surface dip angle label output by the initial surface dip angle estimation model and a plurality of sample vehicle state sensing data of the sample vehicle in the driving environment, and each sample state data is provided with a standard traffic state label; inputting state data of each sample into the initial traffic state classification model, and training the initial traffic state classification model; And in the training process of the initial traffic state classification model, controlling the initial traffic state classification model to output a predicted traffic state label aiming at each sample state data, and adjusting parameters of the initial traffic state classification model according to the predicted traffic state label and the standard traffic state label of each sample state data until the initial traffic state classification model converges to obtain a trained traffic state classification model.
  5. 5. The method of claim 4, wherein the controlling the traffic state classification model to output traffic state prediction results of the vehicle in the current driving environment comprises: controlling the traffic state classification model to output an initial traffic state prediction result of the vehicle in the current driving environment; And comparing the vehicle state sensing data with a preset risk constraint index, and verifying or correcting the initial traffic state prediction result according to a comparison result to obtain a traffic state prediction result of the vehicle in the current driving environment, wherein the risk constraint index is a safety limit value set based on the physical limit of the vehicle.
  6. 6. The method according to claim 4, wherein after the controlling the traffic state classification model outputs the traffic state prediction result of the vehicle in the current driving environment, the method further comprises: And scoring the traffic state prediction result based on the vehicle state sensing data to obtain a traffic state score of the vehicle in the current driving environment, and determining that the traffic state corresponding to the traffic state score in a preset scoring standard is a traffic suggestion for the vehicle.
  7. 7. The method of claim 6, wherein the method further comprises: And carrying out safety evaluation on the current driving environment based on the actual driving condition of the vehicle in a preset time window, if the current driving environment is judged to be a risk environment, tightening a passing threshold in the preset scoring standard, and if the current driving environment is judged to be a safety environment, relaxing the passing threshold in the preset scoring standard.
  8. 8. The method of claim 4, wherein the training process for each model further comprises: After the friction coefficient self-supervision learning model is trained, freezing an encoder corresponding to the friction coefficient self-supervision learning model, and correlating the processed friction coefficient self-supervision learning model with the ground surface dip angle estimation model as an upstream model to obtain a first correlation model; after the pass state classification model is trained, the pass state classification model is used as a downstream model to be associated with the first association model, and a second association model is obtained; And deploying the second association model in a computing platform corresponding to the vehicle.
  9. 9. A computer readable storage medium, characterized in that the computer readable storage medium stores a computer program which, when executed, implements the method according to any of claims 1 to 8.
  10. 10. A vehicle, characterized in that the vehicle comprises: A memory for storing executable program code; A processor for calling and running the executable program code from the memory, causing the vehicle to perform the method of any one of claims 1 to 8.

Description

Path evaluation method, storage medium and vehicle Technical Field The present application relates to the technical field of intelligent driving assistance, and more particularly, to a path evaluation method, a storage medium, and a vehicle in the technical field of intelligent driving assistance. Background When the trafficability of the path is evaluated, the current off-road driving auxiliary system mainly depends on a relatively static and simplified direct judging method such as a preset rule threshold value and the like, so that whether the vehicle can safely pass through the specific path or not is determined. The existing path trafficability evaluation method can provide basic path selection advice for a driver in a known environment, improves the safety of off-road driving to a certain extent, and still has the technical problems of large error, insufficient flexibility and the like. Disclosure of Invention The application provides a path evaluation method, a storage medium and a vehicle, wherein the method can realize the deep fusion of visual perception and physical parameters, so that the dynamic physical state under the complex terrain condition is comprehensively analyzed, and the accuracy and the adaptability of path stability evaluation under the off-road scene are improved. In a first aspect, a path evaluation method is provided, which includes acquiring multi-dimensional vehicle state awareness data of a vehicle in a current driving environment, wherein the vehicle state awareness data is used for representing internal running states and/or external environment information of the vehicle, determining target semantic features of topography in the current driving environment based on visual images in the vehicle state awareness data, determining a ground surface inclination value of a passing path of the vehicle based on the visual images, and determining a passing state prediction result of the vehicle in the current driving environment according to the target semantic features, the ground surface inclination value and the vehicle state awareness data so as to evaluate the passing path according to the passing state prediction result. The technical scheme of the first aspect has the beneficial effects that firstly, multidimensional vehicle state sensing data are collected, a comprehensive and three-dimensional vehicle-environment state sensing system is constructed, so that the system can comprehensively capture the internal running state and external environment information of the vehicle, the comprehensive data collection overcomes the defects of single sensor information surface and easy interference, a rich and accurate information basis is provided for subsequent accurate evaluation, and the overall sensing capability of the system on complex off-road scenes is improved; on the one hand, the method extracts target semantic features from the visual image, can effectively sense the ground surface materials and potential adhesion performance thereof, on the other hand, the method directly returns an accurate ground surface inclination angle value from the image, is favorable for more accurately evaluating the passing risk of the vehicle under different gradient conditions, the two items together form a 'visual-physical' fusion core, the deep analysis of the key physical attributes of the path is realized, finally, the accurate passing state prediction result can be output by fusing the target semantic features, the ground surface inclination angle value and multidimensional vehicle state sensing data, and the passing path is conveniently evaluated according to the passing state prediction result. In some possible implementation manners, the method further comprises the steps of constructing an initial friction coefficient self-supervision learning model, acquiring a plurality of sample visual images of the sample vehicle in a driving environment, inputting each sample visual image into the initial friction coefficient self-supervision learning model, training the initial friction coefficient self-supervision learning model, controlling the initial friction coefficient self-supervision learning model to conduct contrast learning on each sample visual image in the training process of the initial friction coefficient self-supervision learning model, adjusting parameters of the initial friction coefficient self-supervision learning model until the initial friction coefficient self-supervision learning model converges according to a contrast learning result to obtain a trained friction coefficient self-supervision learning model, and determining target semantic features of the terrain in the current driving environment based on the visual images in the vehicle state sensing data, wherein the steps of inputting the visual images in the vehicle state sensing data into the friction coefficient self-supervision learning model, controlling the friction coefficient self-supervision learning model to