Radar-Based Human Motion Recognition Using Semisupervised Triple-GAN

Radar-based human motion recognition (HMR) has drawn considerable attention in many social and military applications. Deep learning (DL) is one of the most promising methods in this field, but it requires large-scale labeled training data to ensure the generalization of the network, which is a challenge in radar applications due to the scarcity of labeled data. To overcome this problem, we propose a semisupervised DL algorithm for radar-based HMR with micro-Doppler signatures, which uses a small amount of labeled data and a large amount of unlabeled data to accurately classify typical human motions. We utilize a semisupervised triple generative adversarial nets (Triple-GANs) model and improve it by introducing a connection structure and a loss term related to the Mixup data augmentation. The performance of the proposed method is verified by applying a public radar micro-Doppler spectrogram dataset, including six motions. Experimental results demonstrate that the proposed method can identify six motions with an average accuracy of 90% using only 10% labeled data in the training dataset. Ablation experimental results verify the efficiency of two improvement strategies. Moreover, the proposed method has a higher average accuracy compared with state-of-the-art methods.