基于残差连接LSTM的雷达目标分类识别方法

Radar Target Classification and Recognition Method Based on Residual Connected LSTM

摘要:

近年来随着多种小型智能探测设备的出现（如无人机、小型智能车等），给传统雷达目标识别方法带来了巨大挑战。在使用雷达对此类小型目标进行探测时得到的信号回波能量通常较低，导致在复杂环境噪声与杂波影响下难以使用传统恒虚警( Constant False Alarm Rate，CFAR)目标检测方法对其进行识别。针对以上问题，结合深度学习的方法提出一种基于残差连接长短期记忆网络（Long Short-Term Memory，LSTM）的多类别雷达目标识别模型，以同一距离门的相邻时间点的回波序列数据作为样本来设计数据集，使用多层的LSTM网络提取雷达回波样本中的时序信息，并在网络中加入残差连接以避免网络层数增多出现网络退化问题，同时将用于多类别分类问题的CCE（Categorical Cross-Entropy）函数作为网络的损失函数来训练网络，实现对包括无人机、智能车、行人以及噪声在内的4类目标的识别和分类。试验结果表明基于残差连接LSTM网络的多类别雷达目标识别模型相比于传统恒虚警检测方法具有更高的识别准确率和F1值。

Abstract:

In recent years, the emergence of a variety of small smart detection devices（such as UAV, small smart car, etc.）has brought great challenges to traditional radar target recognition methods. The signal echo energy obtained when using radar to detect such small targets is usually low，which makes it difficult to identify them by using the traditional constant false alarm rate (CFAR) target detection method under the influence of complex environmental noise and clutter. In response to the above problems，this paper combines the deep learning method to propose a multi-class radar target recognition model based on residual connected Long Short-Term Memory (LSTM)，which takes the echo sequence data at adjacent time points at the same distance gate as the sample to design the data set，the multi-layer LSTM network is used to extract the timing information in the radar echo samples，and the residual connection is added to the network to avoid the problem of network degradation due to the increase of network layers. At the same time，the CCE (categorical cross entropy) function used for the multi-class classification problem is used as the loss function of the network to train the network and realize the recognition and classification of four types of targets, including UAV，smart car，pedestrian and noise. The experimental results show that the multi class radar target recognition model based on residual connected LSTM network has higher recognition accuracy and F1 value than the traditional CFAR detection method.