DopNet: A Deep Convolutional Neural Network to Recognize Armed and Unarmed Human Targets

The work presented in this paper aims to distinguish between armed or unarmed personnel using multi-static radar data and advanced Doppler processing. We propose two modified deep convolutional neural networks (DCNNs) termed single channel DopNet (SC-DopNet) and multiple channel DopNet (MC-DopNet) for mono-static and multi-static micro-Doppler signature (μ-DS) classification. Differentiating armed and unarmed walking personnel is challenging due to the effect of the aspect angle and the channel diversity in real-world scenarios. In addition, the DCNN easily overfits the relatively small-scale μ-DS dataset. To address these problems, the work carried out in this paper makes three key contributions. First, two effective schemes including a data augmentation operation and a regularization term are proposed to train the SC-DopNet from scratch. Next, a factor analysis of the SC-DopNet is conducted based on various operating parameters in both the processing and radar operations. Third, to solve the problem of aspect angle diversity for the μ-DS classification, we design the MC-DopNet for multi-static μ-DS which is embedded with two new fusion schemes termed as greedy importance reweighting (GIR) and 121-Norm. These two schemes are based on two different strategies and have been evaluated experimentally. The GIR uses a win by sacrificing worst case approach, whereas 121-Norm adopts a win by sacrificing best case approach. The SC-DopNet outperforms the non-deep methods by 12.5% in average, and the proposed MC-DopNet with two fusion methods outperforms the conventional binary voting by 1.2% in average. Note that we also argue and discuss how to utilize the statistics of SC-DopNet results to infer the selection of fusion strategies for the MC-DopNet under different experimental scenarios.