首页|期刊导航|南京航空航天大学学报(英文版)|基于图像-轨迹融合的任务级无人机意图识别

基于图像-轨迹融合的任务级无人机意图识别OA

Mission-Level UAV Intent Recognition Using Image-Trajectory Fusion

中文摘要英文摘要

低空交通管理需要及时判断无人机(Unmanned aerial vehicle,UAV)平台类型及其潜在任务意图.对于协作目标,申报飞行计划能够提供一定的意图信息;但对于非协作、异常或仅能被部分观测的无人机,意图需要从可观测线索中推断.现有无人机监视方法主要关注目标检测、跟踪、平台类型识别或轨迹预测,当不同任务具有相似短时运动特征,且单帧图像缺少时间上下文时,难以支持任务级意图判别.针对这一问题,本文提出一种图像-轨迹融合框架MCFNet,将外部视角RGB图像与8 s轨迹历史相结合,用于任务级无人机意图识别.该方法通过双向交叉注意力融合模块,在分类前建立局部视觉区域与轨迹时间步之间的关联,使载荷状态、喷洒效果等局部视觉线索能够结合相应机动片段进行解释.同时,引入模态随机dropout以降低单一模态捷径学习,并采用类型-意图一致性损失约束物理上不合理的类型-任务组合.在包含3类无人机、15类任务意图和7 290个仿真样本的数据集上,MCFNet取得了89.63%±0.99%的意图识别准确率、89.41%±1.00%的类型-意图联合准确率和91.33%±1.05%的宏平均F1值.与最强的晚期融合基线相比,MCFNet的意图识别准确率提高了3.04%,并取得了略高于多模态瓶颈(Multimodal bottkneck transfomer,MBT)token级融合基线的平均意图识别准确率.结果表明,图像-轨迹融合能够为低空交通管理中的早期预警和差异化决策支持提供一种基于观测证据的技术路径.

Low-altitude traffic management requires timely inference of both unmanned aerial vehicle(UAV)platform type and likely mission intent.Declared flight plans can provide intent information for cooperative operations,but non-cooperative,abnormal or partially observed UAVs require inference from observable evidence.The existing surveillance methods mainly detect or track UAVs,classify platform type,or predict trajectories.They therefore remain limited when different missions share similar short-term motion and appearance alone lacks temporal context.This study introduces MCFNet,a multimodal framework that combines an external-view RGB image with an 8 s trajectory history for mission-level UAV intent recognition.Its bidirectional cross-attention fusion links local visual regions with trajectory time steps before classification,allowing cues such as payload state or spray effects to be interpreted with the relevant maneuver segments.Modality dropout reduces single-modality shortcuts,and a consistency loss penalizes invalid type-intent pairs.On 7 290 simulated samples covering three UAV types and 15 intents,MCFNet achieves 89.63%±0.99%intent accuracy,89.41%±1.00%type-intent joint accuracy and 91.33%±1.05%macro-F1,outperforming the strongest late-fusion baseline by 3.04%and achieving a slightly higher mean intent accuracy than the multimodal bottleneck Transformer(MBT)token-level fusion baseline.The results provide an observation-based route to earlier warning and differentiated decision support for low-altitude traffic management.

唐立;王梓洁;涂彭越

西华大学汽车与交通学院,成都 610039,中国||西华大学智能空地融合载具及管控教育部工程研究中心,成都 610039,中国西华大学智能空地融合载具及管控教育部工程研究中心,成都 610039,中国||西华大学航空航天与智能装备学院,成都 610039,中国西华大学航空航天与智能装备学院,成都 610039,中国

航空航天

无人机意图识别图像-轨迹融合低空空域轨迹消歧领域感知约束

UAV intent recognitionimage-trajectory fusionlow-altitude airspacetrajectory disambiguationdomain-aware constraints

《南京航空航天大学学报(英文版)》 2026 (4)

529-545,17

This work was supported by the Na-tional Natural Science Foundation of China(No.52102384),the Project for Enhancing R&D Capabilities in Green,Low-Carbon and New Energy Technologies(No.2025GH-JSHZ-02),and the Xihua University Education and Teaching Re-form Project(Special Initiative on Industry-Education Inte-gration)(No.xcjz2025001).

10.16356/j.1005-1120.2026.04.004

评论