针对由时频分析引起的失真而导致的特征自动抽取质量低的问题,文中将一个自动抽取微多普勒特征过程转化为 一个?2 -范式凸优化问题, 并通过搭建迭代的卷积神经网络框架近似求解。文中仿真运用四种运动捕捉数据库的测量数据,通过仿真模型模拟了雷达视线方向5 m 处的目标的雷达回波。仿真与实验样本所提取的特征用支持向量机分类。仿真和实验的分类性能表明,该框架抽取的特征的分类性能明显优于时频图像主成分分析所自动抽取特征的分类性能。

To avoid the low-quality feature extraction problem caused by the distortion of time-frequency analysis, an automatic micro-Doppler feature extraction process is transformed into a ?2-norm convex optimization problem and approximately solved by an iterative convolutional neural networks (CNNs) framework. The simulated continuous wave radar echo is based on the motion capture data with 5 m between target and radar in light of sight (LOS) direction. The simulated and measured samples are applied in the feature extraction for the final support vector machine classification. The classification performance of simulation and measurement illustrates that the feature quality extracted by the iterative convolutional neural networks framework is significantly better than the feature quality extracted by principal component analysis from the time-frequency images.

TN957

国家留学基金委(CSC) 资助中法蔡元培项目(201806070002)