| 摘要: |
| 机载雷达平台的自运动状态高精度实时感知是实现精确探测的关键前提。脉冲多普勒雷达的距离-多普勒（RD）图谱中蕴含了平台运动（如水平距离、高度、速度等）与地物散射等复杂耦合信息。然而，传统信号处理方法难以在复杂地形与噪声干扰下实现参数解耦，且分辨率受限。深度学习方法为解耦RD图的内蕴信息提供了新路径，但现有研究大多集中在单帧RD图的静态分析，忽视了平台运动这一连续动作的时序信息。针对上述问题，本文选取了专为时序RD图参数反演任务设计的RDMFNet，并将其与ResNet、UNet等通用图像网络进行对比。同时，为克服真实数据稀缺的问题，本文还构建了一个高保真仿真流程，通过地形镶嵌与建筑物嵌入构建复杂场景，并在回波仿真中引入阴影遮蔽算法，结合噪声注入得到高保真时序RD图。使用时序堆叠、时序融合等训练策略的实验结果表明，所选用的RDMFNet的运动参数估计性能显著优于ResNet和UNet，除此之外，时序信息的引入，使模型的抗噪能力大大提高的同时，进一步增强了模型的估计精度，为精度高、鲁棒性强的自运动估计提供了有效方案。 |
| 关键词: 运动参数估计 距离-多普勒图 时间序列 深度学习 |
| DOI： |
| 分类号:TN959.73 |
| 基金项目:国家自然科学基金项目（面上项目，重点项目，重大项目） |
|
| Temporal RD Map-Based Self-Motion Estimation for Airborne Radar Platforms |
|
何屹廷
|
|
|
| Abstract: |
| High-precision and real-time perception of the self-motion state of airborne radar platforms is essential for accurate target detection. The Range–Doppler (RD) maps of pulse-Doppler radar contain complex and nonlinearly coupled information related to platform motion parameters (e.g., horizontal distance, height, and velocity) as well as ground scattering characteristics. However, traditional signal processing methods struggle to decouple these parameters under complex terrain conditions and noise interference, and their achievable estimation accuracy is fundamentally constrained. Deep learning provides an alternative data-driven paradigm for exploiting the intrinsic information embedded in RD maps, yet existing studies mostly focus on static analysis of single-frame RD images, overlooking the sequential nature of platform motion. To address these challenges, we adopt RDMFNet, a network specifically designed for temporal RD map parameter estimation, and compare its performance with general-purpose image-based networks such as ResNet and UNet. Meanwhile, to overcome the scarcity of real-world data, a high-fidelity simulation pipeline is developed. The pipeline constructs complex scenes through terrain mosaicking and building embedding, introduces shadow masking in echo simulation, and injects additive noise to generate realistic temporal RD sequences. Experimental results demonstrate that incorporating temporal information significantly improves motion parameter estimation performance and noise robustness. The proposed approach provides an effective solution for high-precision and robust self-motion estimation of airborne radar platforms. |
| Key words: motion parameter estimation Range-Doppler map time series deep learning |