| 摘要: |
| 在雷达对低慢小（LSS）目标的跟踪任务中，目标频繁的非线性机动以及复杂环境下的强杂波干扰，给航迹关联带来严峻挑战，传统关联方法缺乏对历史运动趋势的深层理解，在杂波逼近真实目标时容易发生误判。为解决上面的问题，本文提出了一种基于双向LSTM异构匹配网络（BiLSTM-based Heterogeneous Matching Network，BHMN）的航迹关联算法。首先，构建了一个面向低慢小目标的航迹数据库LSST,涵盖21种运动模式和6类杂波干扰，为网络训练提供数据支撑；其次，设计了BHMN异构双分支结构以适配航迹序列与量测点两种不同模态的输入，利用航迹编码器的双向LSTM提取历史观测的时序特征，量测编码器的多层感知机提取点位特征，通过网络训练能够学习历史航迹序列与当前量测的匹配规律以及目标运动的演化趋势；最后，编码器输出的两路嵌入经多种互补相似度度量融合后输出关联概率作为评分，结合匈牙利算法完成全局最优关联。实验结果表明，BHMN算法在二分类判别能力和航迹关联能力上均优于基于欧氏距离、马氏距离和运动学一致性的传统航迹关联方法。 |
| 关键词: 低慢小目标 航迹关联 双向长短期记忆网络 异构匹配网络 |
| DOI： |
| 分类号:TN953 |
| 基金项目:安徽省重点研究和开发计划 |
|
| A Track Association Algorithm Based on a Bidirectional LSTM Heterogeneous Matching Network |
|
|
|
|
| Abstract: |
| In radar tracking tasks for low, slow, and small (LSS) targets, frequent nonlinear maneuvering of targets and strong clutter interference in complex environments pose severe challenges to track association. Traditional association methods lack deep representation of historical motion trends, making them prone to misassociations when clutter approaches the true target. To address these issues, this paper proposes a track association algorithm based on a BiLSTM-based Heterogeneous Matching Network (BHMN). First, a track database (LSST) tailored for LSS targets is constructed, encompassing 21 motion modes and 6 types of clutter interference, providing robust data support for network training. Second, a heterogeneous dual-branch architecture is designed within the BHMN to accommodate two distinct input modalities: track sequences and measurement points. A BiLSTM-based track encoder is employed to extract temporal features from historical observations, while an MLP-based measurement encoder extracts point-wise spatial features. Through network training, the model learns the matching patterns between historical track sequences and current measurements, as well as the evolutionary trends of target motion. Finally, the dual-path embeddings output by the encoders are fused through multiple complementary similarity metrics to generate association probabilities as scores, which are then combined with the Hungarian algorithm to achieve globally optimal association. Experimental results demonstrate that the proposed BHMN algorithm outperforms traditional track association methods based on Euclidean distance, Mahalanobis distance, and kinematic consistency in both binary classification and track association performance. |
| Key words: Low, slow, and small (LSS) targets track association Bidirectional Long Short-Term Memory (BiLSTM) Heterogeneous Matching Network (HMN) |