CN-121999450-A - Training method of lane selection model and related equipment

Abstract

The invention relates to the technical field of unmanned driving and the technical field of deep learning, in particular to a training method of a lane selection model and related equipment. The method comprises the steps of carrying out data mining on a target traffic road to obtain a static map data set corresponding to the target traffic road, obtaining static scene marking data of the target traffic road according to traffic attributes of elements in the static map data set, obtaining dynamic scene marking data of the target traffic road according to historical traffic data of the target traffic road, wherein the dynamic scene marking data comprises actual lane selection results between adjacent intersections, carrying out iterative training on an initial deep learning model according to the static scene marking data, the dynamic scene marking data and the static map data set, and determining that the iterative training is finished when preset conditions are detected to obtain the target lane selection model.

Inventors

• YU NING
• GAO LINGPING

Assignees

• 魔门塔(苏州)科技有限公司

Dates

Publication Date

20260508

Application Date

20241101

Claims (10)

1. 1. A method of training a lane selection model, comprising: performing data mining on a target traffic road to obtain a static map data set corresponding to the target traffic road; Obtaining static scene marking data of the target traffic road according to traffic attributes of elements in the static map data set, wherein the static scene marking data comprises an intersection type and ideal lane selection results between adjacent intersections; Obtaining dynamic scene marking data of the target traffic road according to the historical traffic data of the target traffic road, wherein the dynamic scene marking data comprises actual lane selection results between adjacent intersections; Performing iterative training on an initial deep learning model according to the static scene annotation data, the dynamic scene annotation data and the static map data set; and when the preset condition is detected, determining that the iterative training is finished, and obtaining the target lane selection model.
2. 2. The method of claim 1, wherein the elements in the static map dataset include different types of intersections and lane information between adjacent intersections.
3. 3. The method of claim 1, wherein the iteratively training an initial deep learning model from the static scene annotation data, dynamic scene annotation data, and the static map data set comprises: performing first iterative training on the initial deep learning model based on the static scene annotation data and the static map data set; Determining that the first iterative training is finished when the initial deep learning model is detected to meet a first finishing condition, so as to obtain a static scene lane selection model; Performing second iterative training on the static scene lane selection model based on the dynamic scene annotation data and the static map data set; And when the static scene lane selection model is detected to meet a second ending condition, determining that the second iterative training is ended, and obtaining the target lane selection model.
4. 4. The method of claim 3, wherein the initial deep learning model comprises a first classifier and a second classifier, an output of the first classifier being connected to an input of the second classifier; the first classifier outputs the selection probability of each lane under the barrier-free interference according to the input static scene annotation data; And the second classifier outputs the selection probability of each lane under the obstacle according to the input dynamic scene annotation data and the selection probability of each lane under the obstacle-free interference.
5. 5. The method of claim 1, wherein the historical traffic data includes first travel track information of a target vehicle in the target traffic road and second travel track information of at least one vehicle other than the target vehicle, wherein the obtaining dynamic scene annotation data of the target traffic road from the historical traffic data of the target traffic road comprises: Inputting the first driving track information, the second driving track information and the static map data set into a target algorithm so that the target algorithm outputs the actual lane selection result; And generating the dynamic scene annotation data according to the actual lane selection result.
6. 6. The method of claim 5, wherein the historical traffic data further includes environmental information, and the target algorithm outputs an actual lane selection result of the target vehicle under the environmental information based on the environmental information, the first travel track information, the second travel track information, and the static map data set.
7. 7. The method according to any one of claims 1 to 6, further comprising: acquiring target traffic scene data, wherein the target traffic scene data comprises map information and obstacle recognition results; inputting the target traffic scene data into the target lane selection model; Obtaining a lane selection result output by the target lane selection model; and executing unmanned driving according to the lane selection result.
8. 8. A training device for a lane selection model, comprising: The data mining module is used for carrying out data mining on the target traffic road to obtain a static map data set corresponding to the target traffic road; The static scene marking module is used for obtaining static scene marking data of the target traffic road according to the traffic attribute of each element in the static map data set, wherein the static scene marking data comprises an intersection type and an ideal lane selection result between adjacent intersections; the dynamic scene marking module is used for obtaining dynamic scene marking data of the target traffic road according to the historical traffic data of the target traffic road, wherein the dynamic scene marking data comprises actual lane selection results between adjacent intersections; The training module is used for carrying out iterative training on the initial deep learning model according to the static scene annotation data, the dynamic scene annotation data and the static map data set, and determining that the iterative training is finished when a preset condition is detected to obtain a target lane selection model.
9. 9. An electronic device, comprising: at least one processor; The processor is coupled with the memory, wherein: The memory stores program instructions executable by the processor, the processor invoking the program instructions to perform the method of any of claims 1-7.
10. 10. A computer readable storage medium storing computer instructions for causing a computer to perform the method of any one of claims 1 to 7.

Description

Training method of lane selection model and related equipment Technical Field The application relates to the technical field of unmanned driving and the technical field of deep learning, in particular to a training method of a lane selection model and related equipment. Background With the rapid development of artificial intelligence, big data and other technologies, unmanned driving is also becoming an important research and development point in the automobile industry. Unmanned technology is a computer system that integrates a variety of sensors to enable an automobile to be driven autonomously without human intervention. This technology involves the cooperation of artificial intelligence, visual computing, radar, monitoring devices and global positioning systems so that the vehicle can automatically recognize the surrounding environment, make decisions and perform operations to achieve automatic and safe operation of the motor vehicle. The current unmanned technology is often obtained after training based on the traditional rule, and an unmanned deep learning model obtained by training based on the traditional rule cannot adjust the speed, change the lane and avoid dynamic obstacles in complex road conditions. Therefore, how to obtain the intersection route selection generalization capability is strong, and the model of the lane is flexibly adjusted according to the actual road conditions, so that the improvement of the unmanned flexibility is a problem to be solved urgently at present. Disclosure of Invention In view of the above, the application provides a training method and related equipment for a lane selection model, which can improve the intersection line selection capability of the model and increase the unmanned flexibility. In a first aspect, an embodiment of the present invention provides a training method for a lane selection model, including: performing data mining on a target traffic road to obtain a static map data set corresponding to the target traffic road; Obtaining static scene marking data of the target traffic road according to traffic attributes of elements in the static map data set, wherein the static scene marking data comprises an intersection type and ideal lane selection results between adjacent intersections; Obtaining dynamic scene marking data of the target traffic road according to the historical traffic data of the target traffic road, wherein the dynamic scene marking data comprises actual lane selection results between adjacent intersections; Performing iterative training on an initial deep learning model according to the static scene annotation data, the dynamic scene annotation data and the static map data set; and when the preset condition is detected, determining that the iterative training is finished, and obtaining the target lane selection model. In one possible implementation, the elements in the static map data set include different types of intersections and lane information between adjacent intersections. In one possible implementation manner, the performing iterative training on the initial deep learning model according to the static scene annotation data, the dynamic scene annotation data and the static map data set includes: performing first iterative training on the initial deep learning model based on the static scene annotation data and the static map data set; Determining that the first iterative training is finished when the initial deep learning model is detected to meet a first finishing condition, so as to obtain a static scene lane selection model; Performing second iterative training on the static scene lane selection model based on the dynamic scene annotation data and the static map data set; And when the static scene lane selection model is detected to meet a second ending condition, determining that the second iterative training is ended, and obtaining the target lane selection model. In one possible implementation, the initial deep learning model includes a first classifier and a second classifier, an output of the first classifier being connected to an input of the second classifier; the first classifier outputs the selection probability of each lane under the barrier-free interference according to the input static scene annotation data; And the second classifier outputs the selection probability of each lane under the obstacle according to the input dynamic scene annotation data and the selection probability of each lane under the obstacle-free interference. In one possible implementation manner, the historical traffic data comprises first driving track information of a target vehicle in the target traffic road and second driving track information of at least one vehicle except the target vehicle, the dynamic scene annotation data of the target traffic road is obtained according to the historical traffic data of the target traffic road, and the method comprises the following steps: Inputting the first driving track information, the second