Ballistic Target Recognition based on Deep Learning by utilizing the micro-Doppler feature

The identification of ballistic targets is very important in the domain of radar. Dependent on the structure and dynamics, the ballistic targets are usually undergoing micro-motion, such as coning, rotation and tumbling. As an inherent characteristic of ballistic targets, the micro-motion features are usually applied for the recognition of ballistic targets. The micro-Doppler (m-D) feature which carries the time-varying Doppler information of the micro-motion is widely used for ballistic target recognition. Currently, convolutional neural networks (CNNs) are the main recognition methods. However, existing CNNs do not consider the characteristics of the m-D feature and ignore its temporal correlation. In the paper, a new network is proposed, which includes one-dimensional feature extraction and fusion (1D-MEF) module and time self-attention (TSA) module for ballistic target recognition. For the time-frequency (TF) spectrogram generated from radar echoes, the 1D-MEF module is designed to extract and fuse multi-level features and we further design the TSA module to draw the global temporal information among 1D feature sequences. Simulations are performed to verify the validity of our method and the results reveal that it can achieve good recognition accuracy.