Human Identification Based on Short-Time Radar Spectrograms Using Transformer Network

• Published in: Proceedings of the IEEE Radar Conference pp. 1516 - 1519
• Main Authors: Lang, Yue, Yang, Yang, Zhou, Yatong, Ji, Haoran, He, Yuan, Wang, Qing
• Format: Conference Proceeding
• Language: English
• Published: IEEE 15.12.2021
• Subjects: Computational modeling; Data mining; Feature extraction; Micro-Doppler; Person Identification; Personnel; Radar; Time-frequency analysis; Transformer; Transformers
• ISSN: 2640-7736
• DOI: 10.1109/Radar53847.2021.10028261

As an important topic in the field of public security, human identification is of great practical significance in a variety of tasks. Radar-based human identification has gained some attention in recent years, but recognizing people with short-time spectrograms remains an open problem. Since radar spectrograms with short durations contain limited micro-Doppler signatures, it is hard to extract discriminative personnel information for the recognition systems, resulting in poor identification performance. In this paper, a Transformer-based network is presented to solve this problem. The proposed model utilizes the self-attention mechanism to extract the inter-frame sequential features in the time domain, which act as a supplement to the lacking frequency information in the short-time spectrograms. A dataset with 15 subjects is constructed to verify the model, and the experimental results show that the proposed Transformer-based identification strategy achieves improved accuracy than existing methods when recognizing people with short-time spectrograms.