Lip-TWUID: Noninvasive Through-Wall User Identification Using SISO Radar and Lip Movement Micro-Doppler Signatures With Limited Samples
Kai Wei Yang, Dongsheng Zhu, Chong Han, Jian Bao Guo, Suyun Sun, Lijuan Sun · IEEE Transactions on Instrumentation and Measurement · 2025
User identification technology, driven by advancements in AI and smart devices, is crucial in smart homes and surveillance systems. Traditional static biometrics such as fingerprints, iris scans, and facial recognition are prone to security risks and privacy concerns due to spoofing and data misuse. Dynamic biometrics, such as lip movements, offer improved security for voice-interactive devices, while radar technology outperforms vision-based methods by being less affected by lighting and offering better privacy protection. However, the currently popular millimeter-wave radar has limited penetration capabilities and its recognition performance significantly decreases when obstructed by walls. Additionally, existing deep learning (DL)-based recognition models typically require large, precisely labeled datasets, which are difficult to obtain. Few-shot learning (FSL) offers a solution but still depends on extensive auxiliary datasets with similar distributions, complicating real-world applications. To overcome these challenges, this article proposes Lip-TWUID, a user identification model that uses single-in, single-out (SISO) through-wall (TW) radar and lip movement micro-Doppler signatures, which eliminates the need for auxiliary datasets and works with limited training data. The model employs a Siamese network for feature encoding and extends to multicategory recognition via the K-nearest neighbor (KNN) algorithm. To further enhance performance, dynamic data augmentation, a joint loss function that combines contrastive and center loss, and a full-permutation data pairing strategy are introduced. The experimental results show that the model achieves 91.08% accuracy through a 24-cm-thick wall with extremely small samples, outperforming existing methods. This demonstrates the feasibility of high-precision user identification in TW scenarios with limited data.