CN-121994242-A - Multi-task cooperative flight vehicle path track planning method and device

CN121994242ACN 121994242 ACN121994242 ACN 121994242ACN-121994242-A

Abstract

The invention discloses a method and a device for planning a path track of a flying car through multitasking cooperation, and belongs to the technical field of automatic driving. The method comprises the steps of training a path point planning model by means of sample data, extracting a plurality of path points from a running record, projecting the path points into a multi-view image, searching key areas of the multi-view image according to projection positions of the path points in the multi-view image, carrying out feature aggregation on multi-source features by means of a unified decoder according to feature requirements of different tasks in a multi-task in a shared feature space, generating path planning points under multi-task cooperation by means of the obtained aggregated features, carrying out track refinement on the path planning points by means of a track refinement network according to a multi-mode BEV feature map, predicting future states by means of a BEV world model, selecting optimal path tracks from the plurality of path tracks by means of a reward model, and carrying out supervision training by means of a BEV space supervision mode. The method and the device can improve the precision of the flight vehicle path planning.

Inventors

YANG SHICHUN
GUO CHENGJIE
CHEN FEI
YAN ZIYANG
Yue qingyang
LI AOJIE
LI RUIKAI
YAN XIAOYU
CAO YAOGUANG

Assignees

北京航空航天大学

Dates

Publication Date: 20260508
Application Date: 20260205

Claims (10)

1. The method for planning the path track of the flying car by the multitasking cooperation is characterized by comprising the following steps of: the method comprises the steps of obtaining sample data, wherein the sample data are perception data and running records of a sample aerocar in a historical running time period, the perception data comprise multi-view images and laser radar data, and the running process comprises a running track and a running state; Training a pre-constructed path point planning model by utilizing the sample data so as to enable the path point planning module to execute the following operations of extracting a plurality of path points from the running record, projecting the path points into a multi-view image, and searching a key area of the multi-view image according to the projection positions of the path points in the multi-view image; track refinement is carried out on the path planning points by utilizing a track refinement network aiming at a multi-mode BEV feature map generated by perception data in a historical driving time period to obtain a plurality of path tracks, future states are respectively predicted on the basis of each path track by utilizing a BEV world model, track quality is estimated according to a prediction result by combining with a reward model, and an optimal path track is selected from the plurality of path tracks according to an estimation result; and performing supervision training by using a BEV space supervision mode so as to output an optimal path track aiming at the current perception data of the target aerocar by using the trained path point planning module, the track refinement network and the BEV world model.
2. The method of claim 1, wherein the path points comprise at least one type of spatial class path points, temporal class path points, and driving style class path points, and wherein the different types of path points are extracted using different extraction strategies.
3. The method of claim 2, wherein the extracting a plurality of waypoints from the travel record comprises: if the space-class path points are extracted, extracting representative space position points from the running record according to an extraction strategy of the space-class path points based on the position distribution of the path points in the three-dimensional space so as to obtain a plurality of space-class path points; If the time-class path points are extracted, extracting the path points which can reflect the running state changes of the flying automobile in different time periods from the running record according to a time-class extraction strategy based on the change of the path points in the time dimension so as to obtain a plurality of time-class path points; if the driving style type path points are extracted, based on the statistical data of the driving behaviors, the path points which can represent different driving styles are extracted from the driving record according to the driving style type extraction strategy so as to obtain a plurality of driving style type path points.
4. The method of claim 1, wherein the retrieving the key region of the multiview image according to the projection position of the path point in the multiview image comprises: Taking the projection position of a path point in the multi-view image as a center point to sample in the peripheral range of the center point, and dynamically adjusting the distribution area of the sampling point according to the driving task on the path point by utilizing a deformable attention mechanism in the sampling process to adjust the sampling position so as to obtain a key area consisting of the sampling points.
5. The method of claim 1, wherein the unified decoder comprises a time interaction module, a collaborative interaction module, and a task deformable aggregation module; The feature aggregation of the multi-source features by utilizing the unified decoder in the shared feature space aiming at the feature requirements of different tasks in the multi-task comprises the following steps: the time interaction module is utilized to capture time sequence characteristics of the flying automobile in the driving process based on relevance of the path points in the time dimension; Extracting cooperative interaction characteristics among the aerocars based on the mutual relation of the path points in the space position, the time sequence characteristics of the aerocars in the running process and the cooperative relation among the aerocars and other traffic participants by utilizing the cooperative interaction module; and dynamically carrying out feature aggregation on the multi-source features according to the feature requirements of different tasks in the multi-task by using the task deformable aggregation module, wherein the multi-source features also comprise time sequence features and cooperative interaction features.
6. The method of claim 5, wherein the task deformable aggregation module comprises a plurality of aggregation sub-modules in one-to-one correspondence with a plurality of tasks; the task deformable aggregation module dynamically performs feature aggregation on multi-source features according to feature requirements of different tasks in the multi-task, and the task deformable aggregation module comprises the following steps: each aggregation sub-module is utilized to adjust attention weights of the characteristics related to the characteristic requirements of the corresponding tasks in the multi-source characteristics, so that the adjusted attention weights are utilized to conduct characteristic aggregation on the multi-source characteristics to obtain aggregation sub-characteristics, and the aggregation sub-characteristics are utilized to conduct execution of the corresponding tasks; and carrying out fusion treatment on the plurality of aggregation sub-features to obtain the aggregation features.
7. A multitasking, coordinated, flying vehicle path trajectory planning device, the device comprising: the system comprises an acquisition unit, a control unit and a control unit, wherein the acquisition unit is used for acquiring sample data, the sample data are perception data and running records of a sample aerocar in a historical running time period, the perception data comprise multi-view images and laser radar data, and the running process comprises a running track and a running state; the path point planning unit is used for training a pre-constructed path point planning model by utilizing the sample data so as to enable the path point planning module to execute the following operations of extracting a plurality of path points from the running record, projecting the path points into a multi-view image, and searching a key area of the multi-view image according to the projection positions of the path points in the multi-view image; The track planning unit is used for carrying out track refinement on the path planning points by utilizing a track refinement network aiming at the multi-mode BEV characteristic map generated by the perception data in the historical driving time period to obtain a plurality of path tracks, respectively predicting the future state based on each path track by utilizing a BEV world model, and evaluating the track quality according to the prediction result by combining with a reward model so as to select the optimal path track from the plurality of path tracks according to the evaluation result; The training unit is used for performing supervision training by using the BEV space supervision mode so as to output an optimal path track aiming at the current perception data of the target aerocar by using the path point planning module, the track refinement network and the BEV world model which are completed by training.
8. A computer device, characterized in that it comprises a memory for storing a computer program and a processor for executing the computer program stored on the memory for carrying out the steps of the method according to any of the preceding claims 1-6.
9. A computer-readable storage medium, characterized in that the storage medium has stored therein a computer program which, when executed by a processor, implements the steps of the method of any of claims 1-6.
10. A computer program product comprising a computer program which, when executed by a processor, implements the steps of the method of any of claims 1-6.

Description

Multi-task cooperative flight vehicle path track planning method and device Technical Field The invention relates to the technical field of automatic driving, in particular to a method and a device for planning a path track of a flying car through multi-task cooperation. Background With the rapid development of autopilot technology, aerocars are emerging as vehicles, and the problems of path planning and trajectory control are becoming increasingly a research hotspot. In an automatic driving system of a flying automobile, how to realize accurate and real-time path planning, especially path selection under a complex dynamic environment, becomes a key factor for guaranteeing safe and efficient flight. The traditional path planning method focuses on the navigation of ground vehicles, and the path planning of the flying automobile not only needs to consider the geometric structure of the road, but also needs to deal with complex three-dimensional space environments, such as avoidance of obstacles, control of flying height and dealing with dynamic traffic conditions. The existing method for planning the path track of the aerocar has poor planning precision. Disclosure of Invention The invention provides a method and a device for planning a path track of a flying car through multitasking cooperation. The technical proposal is as follows: In one aspect, a method for path trajectory planning of a flying car with multitasking cooperation is provided, the method comprising: the method comprises the steps of obtaining sample data, wherein the sample data are perception data and running records of a sample aerocar in a historical running time period, the perception data comprise multi-view images and laser radar data, and the running process comprises a running track and a running state; Training a pre-constructed path point planning model by utilizing the sample data so as to enable the path point planning module to execute the following operations of extracting a plurality of path points from the running record, projecting the path points into a multi-view image, and searching a key area of the multi-view image according to the projection positions of the path points in the multi-view image; track refinement is carried out on the path planning points by utilizing a track refinement network aiming at a multi-mode BEV feature map generated by perception data in a historical driving time period to obtain a plurality of path tracks, future states are respectively predicted on the basis of each path track by utilizing a BEV world model, track quality is estimated according to a prediction result by combining with a reward model, and an optimal path track is selected from the plurality of path tracks according to an estimation result; and performing supervision training by using a BEV space supervision mode so as to output an optimal path track aiming at the current perception data of the target aerocar by using the trained path point planning module, the track refinement network and the BEV world model. In another aspect, a multitasking cooperative apparatus for planning a path trajectory of a flying vehicle is provided, the apparatus comprising: the system comprises an acquisition unit, a control unit and a control unit, wherein the acquisition unit is used for acquiring sample data, the sample data are perception data and running records of a sample aerocar in a historical running time period, the perception data comprise multi-view images and laser radar data, and the running process comprises a running track and a running state; the path point planning unit is used for training a pre-constructed path point planning model by utilizing the sample data so as to enable the path point planning module to execute the following operations of extracting a plurality of path points from the running record, projecting the path points into a multi-view image, and searching a key area of the multi-view image according to the projection positions of the path points in the multi-view image; The track planning unit is used for carrying out track refinement on the path planning points by utilizing a track refinement network aiming at the multi-mode BEV characteristic map generated by the perception data in the historical driving time period to obtain a plurality of path tracks, respectively predicting the future state based on each path track by utilizing a BEV world model, and evaluating the track quality according to the prediction result by combining with a reward model so as to select the optimal path track from the plurality of path tracks according to the evaluation result; The training unit is used for performing supervision training by using the BEV space supervision mode so as to output an optimal path track aiming at the current perception data of the target aerocar by using the path point planning module, the track refinement network and the BEV world model which are completed by training. In another aspect, a computer devic