基于航迹序列的4类低空目标识别

Four-class low-altitude target recognition based on track sequences

摘要

基于深度学习方法的航迹预测所进行的低空安全管控,近年来受到广泛关注。针对小型旋翼无人机、轻型旋翼无人机、鸟和空飘球等低空目标航迹特征相近、识别难度较大的问题,本文从低空目标的航迹特征出发,通过利用飞行目标在飞行过程中的态势特性,对比了不同飞行目标在运动轨迹上的差异,从飞行物的运动、时序、速率趋势特性中提取了多种高纬度特征。在此基础上,本文提出一种多头1D-CNN和多层双向的长短期记忆网络(Bi-directional long short-term memory,BiLSTM)神经网络架构,该架构能够通过航迹时间序列数据的前向和反向依赖关系,捕获小型旋翼无人机、轻型旋翼无人机、鸟以及空飘球等4类低空飞行物在飞行轨迹中多维特征的空间和时间的依赖关系,实现了对雷达航迹数据的精确区分策略,有效解决小型旋翼无人机、轻型旋翼无人机、鸟和空飘球等低空目标航迹特征相近、识别难度较大等问题。实验结果表明:对于4类低空目标识别,提出的算法加权平均识别准确率为91.59%,加权平均F1-Score为91.33%,加权平均虚警率和漏检率分别为3.07%和8.41%,均优于长短期记忆网络(Long short term memory,LSTM)、BiLSTM、卷积神经网络(Convolutional neural networks,CNN)等重要算法。

Abstract

Low-altitude safety management based on deep learning-based track prediction has received widespread attention in recent years.Addressing the challenge of distinguishing between low-altitude targets with similar trajectory characteristics,such as small rotor-wing drones,light rotor-wing drones,birds,and free-floating balloons,this paper approaches the problem by analyzing the inherent trajectory features of these targets.By leveraging the situational dynamics exhibited by these flying objects during flight,and by comparing the differences in their respective motion trajectories,the study extracts a diverse set of high-dimensional features encompassing their motion patterns,temporal sequences,and velocity trends.Building upon this foundation,the paper proposes a novel neural network architecture comprising a multi-head 1D-CNN integrated with a multi-layer Bidirectional Long Short-Term Memory (BiLSTM) network.This architecture is designed to capture the spatio-temporal dependencies of the multi-dimensional features embedded within the flight trajectories of the four aforementioned categories of low-altitude objects—small rotor-wing drones,light rotor-wing drones,birds,and free-floating balloons—by analyzing both the forward and backward dependencies within their trajectory time-series data.This approach enables a precise differentiation strategy for radar trajectory data,thereby effectively resolving the difficulties associated with identifying low-altitude targets that exhibit highly similar trajectory characteristics.Experimental results demonstrate that,for the identification of these four categories of low-altitude targets,the proposed algorithm achieves a weighted average recognition accuracy of 91.59% and a weighted average F1-Score of 91.33%; furthermore,its weighted average false alarm rate and missed detection rate stand at 3.07% and 8.41%,respectively.These performance metrics consistently outperform those of other prominent algorithms,including standard Long Short-Term Memory (LSTM) networks,BiLSTM networks,and Convolutional Neural Networks (CNN).

关键词

航迹检测 / BiLSTM / 多头1D-CNN

Key words

track detection / BiLSTM / multi-head 1D-CNN

[1] Xing Zehua,Hu Shengbo,Ding Ruxuan,et al.Multi-sensor dynamic scheduling for defending UAV swarms with Fresnel zone under complex terrain[J].ISA Transactions,2024,153:57-69.
[2] Ding Ruxuan,Hu Shengbo,Xing Zehua,et al.Multi-type radar deployment for UAV swarms defense coverage using Firework Algorithm with Determinantal Point Processes under complex terrain[J].Applied Soft Computing,2026,170:112681.
[3] 沈兵,胡啸峰,吴建松.社区治安高危人员异常轨迹识别与预警方法研究[J].中国安全生产科学技术,2021,17(4):171-177.
[4] 刘颢,陈世友,汪学东,等.一种自适应航迹关联算法[J].电子学报,2013,41(12):2416-2421.
[5] 李书盼,梁彦,张会霞,等.基于雷达航迹序列的无人机/飞鸟动态分类[J].航空学报,2026,47(3):103-118.
[6] 毛鑫骞,于烨,贾维敏,等.基于多波段雷达的低慢小目标微动特征提取方法[J].无线电工程,2025,55(6):1200-1206.
[7] 陈唯实,黄毅峰,卢贤锋.多传感器融合在无人机探测技术中的应用综述[J].现代雷达,2020,42(6):15-29.
[8] 宋志勇,许云涛.基于多普勒与微多普勒联合利用的弱小目标检测与估计方法[J].电子与信息学报,2023,45(11):4083-4091.
[9] 彭华甫,黄高明,田威,等.幅度及多普勒信息辅助的多目标跟踪算法[J].航空学报,2018,39(10):234-241.
[10] 杨勇,王雪松,张斌.基于时频检测与极化匹配的雷达无人机检测方法[J].电子与信息学报,2021,43(3):509-515.
[11] Torvik B,Olsen K E,Griffiths H.Classification of birds and UAVs based on radar polarimetry[J].IEEE Geoscience and Remote Sensing Letter,2016,13(9):1305-1309.
[12] 刘佳,徐群玉,陈唯实.无人机雷达航迹运动特征提取及组合分类方法[J].系统工程与电子技术,2023,45(10):3122-3131.
[13] 管康萍,冯正康,马小艳,等.一种基于航迹特征的无人机与飞鸟目标雷达识别方法[J].上海航天(中英文),2024,41(1):130-136.
[14] 李思维.基于微多普勒特性的低慢小目标识别方法研究[D].成都:电子科技大学,2024.
[15] 李冬梅.基于激光微多普勒效应的旋翼无人机识别方法研究[D].哈尔滨:哈尔滨工业大学,2022.
[16] Byung K K,Hyun-Seong K,Seongwook L.et al.Improved drone classification using polarimetric merged-doppler images[J].IEEE Geoscience and Remote Sensing Letters,2020,18(11):1946-1950.
[17] 黄虹玮,刘玉娇,沈卓恺,等.基于深度学习网络模型的端到端航迹关联[J].计算机科学,2020,47(3):200-205.
[18] 杨博宇,杨勇,李永祯,等.基于启发式搜索的极化雷达无人机检测方法[J].信号处理,2025,41(5):867-885.
[19] 吴济洲,张红敏.一种航迹数据高维特征矩阵提取方法[J].指挥控制与仿真,2022,44(1):51-57.
[20] 熊风,王干,白杨,等.基于卷积神经网络的雷达航迹起始方法[J].电子测量技术,2020,43(10):78-83.
[21] 吴琪.基于多维特征融合的“低慢小”目标自动识别关键技术研究[D].长沙:国防科学技术大学,2019.
[22] Cui Zhiyong,Ke Ruimin,Pu Ziyuan,et al.Stacked bidirectional and unidirectional LSTM recurrent neural network for forecasting network-wide traffic state with missing values[J].Transportation Research Part C:Emerging Technologies,2020,118:102674.
[23] 秦攀科,叶波.融合鲸鱼算法优化的BiLSTM-ARIMA期货套利价差预测方法[J].重庆邮电大学学报(自然科学版),2026,38(2):321-330.
[24] 王铁虎,刘平安,欧玉鹏,等.基于LSTM神经网络的渡槽温度场预测[J].灌溉排水学报,2026,45(2):73-81.

基金资助

贵州省科技厅项目(黔科合平台ZSSYS[2025]重大001)