CN-122019337-A - Training evolution method for unmanned system real-time mapping and positioning capability

CN122019337ACN 122019337 ACN122019337 ACN 122019337ACN-122019337-A

Abstract

The invention discloses a training evolution method for real-time mapping and positioning capability of an unmanned system, which relates to the field of unmanned system simulation training and positioning mapping, and comprises the steps of constructing a city-level simulation scene set covering various city environment characteristics, and carrying out parameterization and randomization configuration; determining a super parameter set and a constraint domain thereof of a definition algorithm, driving an unmanned system to execute a multi-round repeated simulation test according to a preset task script, collecting related data, forming a multi-index evaluation vector, generating a comprehensive score, carrying out iterative updating on the super parameter set to obtain a candidate super parameter set, writing information corresponding to the current iterative round into a capacity evolution knowledge base, retrieving the super parameter set from the capacity evolution knowledge base as an initialization parameter for parameter migration, and carrying out continuous iterative optimization on the candidate super parameter set by adopting a course-type difficulty propulsion mechanism until the preset convergence condition and real-time constraint are met, and outputting the optimal super parameter set and an applicable scene domain thereof.

Inventors

JIANG CHUNMAO
CHEN ZIYU
ZHU HUI
KUANG FEI
XU LIWEI
ZHANG SHIJIN

Assignees

中国科学院合肥物质科学研究院

Dates

Publication Date: 20260512
Application Date: 20260127

Claims (14)

1. The training evolution method for the unmanned system real-time mapping and positioning capability is characterized by comprising the following steps of: Constructing a city level simulation scene set covering various city environment characteristics, and carrying out parameterization and randomization configuration on scene elements; Determining a real-time mapping and positioning algorithm of evolution to be trained, defining a hyper-parameter set of the algorithm and a constraint domain thereof, and establishing a feasibility check rule of the hyper-parameter set; In the urban level simulation scene set, driving an unmanned system to execute a plurality of rounds of repeated simulation tests according to a preset task script, and collecting track data, map data and run-time resource data; Calculating a plurality of evaluation indexes based on the acquired data to form a multi-index evaluation vector, and generating a comprehensive score according to the multi-index evaluation vector; Combining an automatic parameter adjustment rule based on an error mode and a multi-target optimization strategy, carrying out iterative updating on the super parameter set to obtain a candidate super parameter set, and writing scene configuration information, the super parameter set and performance indexes corresponding to the current iteration round into a capability evolution knowledge base; In the cross-scene training process, the super-parameter set is searched from the capability evolution knowledge base based on scene similarity and used as an initialization parameter to carry out parameter migration, and a course-type difficulty propulsion mechanism is adopted to carry out continuous iterative optimization on the candidate super-parameter set until a preset convergence condition and real-time constraint are met, and the optimal super-parameter set and an applicable scene domain thereof are output.
2. The training evolution method for real-time mapping and positioning capability of unmanned system according to claim 1, wherein the step of constructing a city-level simulation scene set covering a plurality of city environment features comprises: Constructing a basic topology model of a plurality of city simulation scenes, wherein the basic topology model at least comprises a road structure, an intersection type, a building group form and landmark distribution; The method comprises the steps of controllably generating illumination conditions and weather conditions, and carrying out parameterization setting on density, speed distribution and behavior modes of dynamic traffic participants; performing parameterization on a sensor degradation element, wherein the sensor degradation element comprises noise intensity, external parameter disturbance, time synchronization error, frame loss probability or shielding probability; And generating a random parameter set based on the random seed generator, and generating a class parameter set according to a preset challenge class to form a simulation scene set comprising multiple cities, multiple difficulty classes and multiple working conditions.
3. The unmanned system-oriented real-time mapping and positioning capability training evolution method according to claim 2, wherein the method is characterized in that: the random parameter sets of the illumination condition and the weather condition comprise cloud layer type, cloud layer height, solar azimuth or wind direction and wind speed; The grade parameter sets of the illumination condition and the weather condition comprise cloud cover, visibility, rain and snow haze intensity or road surface reflection intensity; And normalizing the parameters in the class parameter set, and dividing the parameters into a plurality of challenge classes for the course-based difficulty propulsion mechanism to call.
4. The training evolution method for real-time mapping and positioning capability of an unmanned system according to claim 1, wherein the super-parameter group at least comprises a front-end odometer super-parameter, a key frame or key point management super-parameter, a loop detection super-parameter, a back-end map optimization super-parameter and a map update super-parameter.
5. The training evolution method for real-time mapping and positioning capability of an unmanned system according to claim 4, wherein the training evolution method is characterized in that: the front-end odometer super-parameters comprise a feature extraction threshold value, a matching window size, voxel resolution or motion model weight; the back-end graph optimization super-parameters comprise sliding window size, robust kernel parameters, edge weight proportion, iteration frequency upper limit or convergence threshold.
6. The training evolution method for real-time mapping and positioning capability of an unmanned system according to claim 1, wherein the feasibility checking rule comprises: The real-time constraint is used for limiting the single frame processing time delay or the real-time factor not to exceed a preset threshold; The resource occupation constraint is used for limiting the occupation rate of the CPU and the GPU, the occupation amount of the memory or the occupation amount of the bandwidth not to exceed a preset threshold value; stability constraints for determining a set of infeasible parameters for trajectory divergence, map optimization non-convergence, or map structural anomalies.
7. The training evolution method for real-time mapping and positioning capability of an unmanned system according to claim 1, wherein the preset task script at least comprises a constant line cruising task, a random exploration task, a shielding traversing task, a dynamic interference task and a sensor degradation working condition task; And performing repeated tests on each task under different random seeds to obtain a statistically stable evaluation result.
8. The method for training and evolving the real-time mapping and positioning capability of the unmanned system according to claim 6, wherein the multiple evaluation indexes at least comprise absolute track error, relative pose error, drift rate, loop consistency error, map integrity, map redundancy, single frame processing time delay, real-time factor and resource occupation index.
9. The method for training evolution of real-time mapping and positioning capabilities for an unmanned system according to claim 8, wherein the step of generating a composite score comprises: weighting and summing all evaluation indexes to obtain weighted items; applying a penalty term according to violation conditions of the real-time constraint and the resource occupation constraint; The composite score is formed from the weighted term and the penalty term.
10. The training evolution method for real-time mapping and positioning capability of an unmanned system according to claim 1, wherein the automatic parameter adjustment rule based on the error mode comprises: When an error mode of track drift or accumulated error increase is identified, executing parameter adjustment actions of improving loop triggering strength, increasing rear-end constraint weight or adjusting repositioning threshold; when the error mode of map shake, repeated map building or closed loop conflict increase is identified, performing parameter tuning actions of tightening key frame insertion conditions, improving robust kernel strength or reducing noise sensitive parameters; and when the error mode that the single frame processing time delay exceeds the threshold value is identified, executing the parameter adjusting action of reducing the upper limit of the number of the characteristic or point cloud, reducing the loop detection frequency or reducing the back-end sliding window.
11. The training evolution method for real-time mapping and positioning capability of an unmanned system according to claim 1, wherein the multi-objective optimization strategy is a bayesian optimization, an evolutionary strategy or a hybrid optimization strategy.
12. The training evolution method for real-time mapping and positioning capability of an unmanned system according to claim 11, wherein the hybrid optimization strategy comprises: generating initial values of candidate super-parameter sets by the automatic parameter adjustment rules based on the error modes, and then executing black box search in the neighborhood of the initial values.
13. The training evolution method for real-time mapping and positioning capability of an unmanned system according to claim 1, wherein the course difficulty propulsion mechanism is specifically: training in a low-challenge-level scene with low dynamic state, low shielding and good illumination to obtain a basic super-parameter set; And then introducing high-dynamic traffic participants, strong shielding and high-challenge-level scenes of complex weather step by step, and performing iterative optimization on the basic super-parameter set.
14. The method for training and evolving the real-time mapping and positioning capability of the unmanned system according to claim 1, wherein the convergence condition comprises that the comprehensive score is improved by less than a threshold value in continuous N rounds of iteration, or a key precision index reaches a target and real-time constraint is met, or the maximum iteration number is reached, and the optimal super-parameter set and the applicable scene domain thereof are output.

Description

Training evolution method for unmanned system real-time mapping and positioning capability Technical Field The invention relates to the technical field of unmanned system simulation training and positioning mapping, in particular to a training evolution method for unmanned system real-time mapping and positioning capability. Background When the unmanned system performs the tasks of inspection, distribution, security protection and the like in the urban environment, the unmanned system is required to rely on a real-time map building and positioning algorithm to realize autonomous positioning and environment modeling. Urban environments generally have the characteristics of complex road structure, frequent building shielding, high repeatability of textures and geometric features, obvious illumination and weather changes, dense dynamic traffic participants and the like, and the factors all put high requirements on the robustness and the adaptability of real-time mapping and positioning algorithms. In the prior art, the real-time mapping and the parameter setting of a positioning system are mostly realized by repeated experiments and manual adjustment under a limited scene depending on expert experience. This approach has the following limitations: First, the parameter tuning process is usually based on a single or a few typical scenarios, which easily results in an over-adaptation of the parameter settings to a specific environment, and an insufficient generalization capability in a multi-scenario, cross-scenario situation. Secondly, the real-time mapping and positioning system also needs to meet strict real-time performance and computational resource constraint, and the current parameter adjustment method is often focused on precision optimization, so that the balance between real-time performance and resource occupation is easily ignored, and the algorithm is difficult to meet performance requirements in actual deployment. In addition, the existing method lacks a systematic recording and multiplexing mechanism for the relation of scene-parameter-performance, so that a large number of parameter adjustment tests need to be repeatedly carried out in different projects or new scenes, the setting cost is high, the period is long, and the knowledge accumulation and migration efficiency is low. Therefore, how to provide a method capable of systematically and automatically realizing cross-scene parameter optimization and evolution, and on the premise of guaranteeing timeliness and resource constraint, improve the adaptability and stability of unmanned system map building and positioning algorithms in different urban environments, and become a technical problem to be solved in the field. Disclosure of Invention In order to solve the technical problems, the invention provides a training evolution method for real-time mapping and positioning capability of an unmanned system, which comprises the following steps: Constructing a city level simulation scene set covering various city environment characteristics, and carrying out parameterization and randomization configuration on scene elements; Determining a real-time mapping and positioning algorithm of evolution to be trained, defining a hyper-parameter set of the algorithm and a constraint domain thereof, and establishing a feasibility check rule of the hyper-parameter set; In the urban level simulation scene set, driving an unmanned system to execute a plurality of rounds of repeated simulation tests according to a preset task script, and collecting track data, map data and run-time resource data; Calculating a plurality of evaluation indexes based on the acquired data to form a multi-index evaluation vector, and generating a comprehensive score according to the multi-index evaluation vector; Combining an automatic parameter adjustment rule based on an error mode and a multi-target optimization strategy, carrying out iterative updating on the super parameter set to obtain a candidate super parameter set, and writing scene configuration information, the super parameter set and performance indexes corresponding to the current iteration round into a capability evolution knowledge base; In the cross-scene training process, the super-parameter set is searched from the capability evolution knowledge base based on scene similarity and used as an initialization parameter to carry out parameter migration, and a course-type difficulty propulsion mechanism is adopted to carry out continuous iterative optimization on the candidate super-parameter set until a preset convergence condition and real-time constraint are met, and the optimal super-parameter set and an applicable scene domain thereof are output. Further, the step of constructing a city level simulation scene set covering a plurality of city environment features includes: Constructing a basic topology model of a plurality of city simulation scenes, wherein the basic topology model at least comprises a road structure, an i