CN-122009246-A - Driving track planning method and device and terminal equipment

CN122009246ACN 122009246 ACN122009246 ACN 122009246ACN-122009246-A

Abstract

The application provides a driving track planning method, a driving track planning device and terminal equipment, which are suitable for the technical field of data processing, wherein the method comprises the steps of generating driving track planning donor knowledge information and driving track planning acceptor knowledge information according to historical driving scene image information, a driving track planning knowledge donor model and a driving track planning knowledge acceptor model; obtaining driving track planning donor-acceptor alignment information according to the driving track planning donor knowledge information, the driving track planning acceptor knowledge information and the driving track planning donor-acceptor alignment model; and training the driving track planning knowledge acceptor model according to the driving track planning donor-acceptor alignment information to obtain a driving track planning model. The method effectively avoids the problems of feature extraction failure, track drift caused by blind inheritance of wrong knowledge and the like of the traditional model caused by noise interference, and greatly improves the track planning capability of the end-to-end automatic driving system in a complex unstructured scene.

Inventors

ZHAO RUI
CHANG SHIJIE
GAO FEI
GAO ZHENHAI

Assignees

吉林大学

Dates

Publication Date: 20260512
Application Date: 20260413

Claims (10)

1. A driving trajectory planning method, characterized by comprising: Acquiring current driving scene image information; generating current driving track planning information according to the current driving scene image information and a driving track planning model; The driving track planning model is obtained through the following steps: acquiring historical driving scene image information, a driving track planning knowledge donor model and a driving track planning knowledge acceptor model; Generating driving track planning donor knowledge information and driving track planning acceptor knowledge information according to the historical driving scene image information, the driving track planning knowledge donor model and the driving track planning acceptor knowledge model; obtaining driving track planning donor-acceptor alignment information according to the driving track planning donor knowledge information, the driving track planning acceptor knowledge information and the driving track planning donor-acceptor alignment model; and training the driving track planning knowledge acceptor model according to the driving track planning donor-acceptor alignment information to obtain a driving track planning model.
2. The driving trajectory planning method according to claim 1, characterized in that, The driving track planning donor-acceptor alignment model comprises a driving scene image characteristic alignment sub-model and a driving track planning semantic generation sub-model; The driving track planning donor-acceptor alignment information comprises driving track planning donor-acceptor image feature alignment loss information and driving track planning donor-acceptor semantic feature alignment loss information; The step of obtaining driving trajectory planning donor-acceptor alignment information according to the driving trajectory planning donor knowledge information, driving trajectory planning acceptor knowledge information and driving trajectory planning donor-acceptor alignment model specifically comprises the following steps: Generating driving track planning donor knowledge information, driving track planning acceptor knowledge information and driving scene image feature alignment sub-model according to the driving track planning donor knowledge information, the driving track planning acceptor knowledge information and the driving scene image feature alignment sub-model; calculating to obtain driving track planning donor-acceptor image characteristic alignment loss information according to the driving track planning donor-acceptor image alignment characteristic information and driving track planning donor knowledge information; calculating to obtain driving track planning donor knowledge weight information according to the driving track planning donor knowledge information and a preset driving track planning donor knowledge screening rule; generating driving track planning semantic information according to the driving track planning donor-acceptor image alignment characteristic information, driving track planning donor knowledge weight information, driving track planning acceptor knowledge information and driving track planning semantic generation sub-model; and calculating to obtain the driving track planning supply receptor semantic feature alignment loss information according to the driving track planning semantic information and the driving track planning donor knowledge information.
3. The driving locus planning method according to claim 2, characterized in that, The driving track planning receptor knowledge information comprises driving track planning receptor knowledge content information and driving track planning receptor noise information; the step of generating driving track planning donor image alignment feature information according to the driving track planning donor knowledge information, driving track planning acceptor knowledge information and driving scene image feature alignment sub-model specifically comprises the following steps: Generating driving scene image feature variance distribution information and driving scene image feature mean distribution information according to the driving track planning donor knowledge information, driving track planning receptor knowledge content information, driving track planning receptor noise information and driving scene image feature alignment sub-model; and carrying out reconstruction processing according to the driving scene image feature variance distribution information and the driving scene image feature mean distribution information to generate driving track planning for the receptor image to align with the feature information.
4. The driving trajectory planning method according to claim 2, wherein the step of calculating driving trajectory planning donor knowledge weight information according to the driving trajectory planning donor knowledge information and a preset driving trajectory planning donor knowledge screening rule specifically comprises: Carrying out probability distribution conversion according to the driving track planning donor knowledge information to obtain driving track planning donor knowledge probability distribution information; calculating to obtain a driving track planning donor knowledge information entropy value according to the driving track planning donor knowledge probability distribution information and a preset driving track planning donor knowledge information entropy calculation function; and calculating to obtain driving track planning donor knowledge weight information according to the driving track planning donor knowledge information entropy value and a preset driving track planning donor knowledge screening rule.
5. The driving trajectory planning method according to claim 2, wherein the step of generating driving trajectory planning semantic information according to the driving trajectory planning donor image alignment feature information, driving trajectory planning donor knowledge weight information, driving trajectory planning acceptor knowledge information, and driving trajectory planning semantic generation sub-model specifically comprises: Calculating to obtain driving track planning donor knowledge weighting information according to the driving track planning donor knowledge information and the driving track planning donor knowledge weighting information; And generating driving track planning semantic information according to the driving track planning donor-acceptor image alignment characteristic information, the driving track planning donor knowledge weighting information, the driving track planning acceptor knowledge information and the driving track planning semantic generation sub-model.
6. The driving trajectory planning method according to claim 2, wherein the step of training a driving trajectory planning knowledge acceptor model according to the driving trajectory planning donor-acceptor alignment information to obtain a driving trajectory planning model specifically comprises: Training the driving scene image feature alignment sub-model according to the driving track planning donor-acceptor image feature alignment loss information, and generating a trained driving scene image feature alignment sub-model; Generating a driving track planning model to be trained according to the trained driving scene image feature alignment sub-model and the driving track planning semantics; Calculating weights according to the driving track planning donor-acceptor image feature alignment loss information, the driving track planning donor-acceptor semantic feature alignment loss information and preset driving track planning feature alignment loss, and calculating to obtain driving track planning donor-acceptor feature comprehensive loss; And training the driving track planning model to be trained according to the comprehensive loss of the driving track planning donor and acceptor characteristics to obtain a driving track planning model.
7. The driving trajectory planning method according to claim 6, wherein the training of the driving scene image feature alignment sub-model according to the driving trajectory planning donor-acceptor image feature alignment loss information, and the generating of the trained driving scene image feature alignment sub-model specifically comprises: Calculating to obtain a driving track planning mapping characteristic coefficient according to the driving track planning donor-acceptor image characteristic alignment loss information and a preset driving track planning mapping characteristic coefficient calculation function; and training the driving scene image feature alignment sub-model according to the driving track planning mapping feature coefficient to generate a trained driving scene image feature alignment sub-model.
8. A driving trajectory planning device, characterized by comprising: the current driving scene image information acquisition module is used for acquiring current driving scene image information; and the current driving track planning information generation module is used for generating current driving track planning information according to the current driving scene image information and the driving track planning model.
9. A terminal device, characterized in that it comprises a memory, a processor, on which a computer program is stored which is executable on the processor, the processor executing the computer program to carry out the steps of the method according to any one of claims 1 to 7.
10. A computer readable storage medium storing a computer program, characterized in that the computer program when executed by a processor implements the steps of the method according to any one of claims 1 to 7.

Description

Driving track planning method and device and terminal equipment Technical Field The application belongs to the technical field of data processing, and particularly relates to a driving track planning method, a driving track planning device and terminal equipment. Background The current automatic driving technology is evolving from auxiliary driving to higher-order automatic driving, the limitation of the traditional modularized technology architecture is gradually highlighted, and the end-to-end automatic driving architecture becomes an important direction of technology development because of being capable of constructing direct mapping from original sensor input to vehicle control signal output, and a new path is provided for solving the problem of 'cognitive long tail' of automatic driving by introducing a large visual language model. However, the high-performance multi-mode large model has the problems of large parameter scale, high calculation cost and high memory occupation, has obvious contradiction with the calculation power resource, the power consumption limit and the real-time requirement of the vehicle-mounted edge calculation platform, takes knowledge distillation as a mainstream model light-weight technology, is widely applied to the related research of migrating the large model reasoning capacity to a light-weight model, and becomes a key exploration direction of end-to-end automatic driving landing. In the prior art, an end-to-end driving planning scheme based on knowledge distillation mainly adopts a mode of combining feature distillation and logic distillation, a feature distillation link uses mean square error as a measurement standard, rigidity alignment in Euclidean space is carried out on middle layer features of a lightweight model and a large model, the logic distillation link utilizes output distribution of a KL divergence constraint lightweight model to approach to large model output, meanwhile, partial scheme is introduced into a variation information bottleneck theory to carry out feature compression, fixed Lagrange multiplier is adopted to control compression strength, noise filtering is attempted to be realized in a feature extraction stage, and migration of large model driving planning knowledge to the lightweight model is completed in the mode, so that the vehicle-mounted platform deployment requirement is adapted. The method has the advantages that the prior art has various technical problems in practical application, the noise screening mechanism is lacked in feature distillation, the model robustness is insufficient, the uncertainty perception of large model output is lacked in logic distillation, the risk of error knowledge transfer exists, the parameter adjustment mode of the variation information bottleneck method is stiff and is difficult to adapt to a driving scene which dynamically changes, the situation of excessively fitting environmental noise is easy to occur in the process of inheriting the reasoning capacity of the large model by a lightweight model, the extraction capacity of key driving semantics is limited, the error planning logic of the large model is blindly inherited by the lightweight model in a long-tail driving scene, automatic driving potential safety hazards are caused, the problem that key details are lost or noise residues are easy to occur in the feature compression mode of fixed compression rate, and the balance between the reserved semantics and the filtered noise cannot be realized. Disclosure of Invention In view of the above, the embodiment of the application provides a driving track planning method, a driving track planning device and terminal equipment, which aim to solve the problems that a model in the prior art is easily interfered by environmental noise such as rain, snow and illumination, a light model blindly inherits a large model 'illusion' error due to lack of uncertainty perception, a dynamic driving scene cannot be adapted, a large model has high calculation load and a vehicle-mounted edge platform resource are limited and are difficult to be compatible, the light model is easy to be subjected to fitting of the environmental noise, the critical semantic extraction capability is weak, and the safety hidden trouble is easy to be caused by inheritance of error planning logic. A first aspect of an embodiment of the present application provides a driving trajectory planning method, including: Acquiring current driving scene image information; And generating current driving track planning information according to the current driving scene image information and the driving track planning model. One aspect of the embodiments of the present application provides a driving trajectory planning model generating step, including: acquiring historical driving scene image information, a driving track planning knowledge donor model and a driving track planning knowledge acceptor model; Generating driving track planning donor knowledge information