A Multi-Domain Fusion Human Motion Recognition Method Based on Lightweight Network

Through-wall human motion recognition is suffered from the problems of too few samples and too large model parameters. In this letter, we propose a multi-domain fusion through-the-wall radar (TWR) human motion recognition model based on lightweight network and transfer learning. Specifically, in order to make full use of the target information, a multiple parallel feature pyramid network (FPN) is first proposed to extract the detailed feature information from the time–frequency map and range profile. After that, a lightweight network based on the MobileNetV3 network and transfer learning is proposed. The MobileNetV3 model is pre-trained on the public ImageNet database. To ensure the performance of transfer learning, a heterogeneous migration learning algorithm is used to cross-domain transform the obtained time–frequency map and range profile. Experimental results show that the proposed model has a better performance in accuracy, model size, training time, and robustness compared with the existing methods. It also has the potential to embed portable radar, which has important research value for the application of radar in real life.