M-Identity: User Identification Algorithm Based on Microphone Array

Research on user identification based on footsteps has gained increasing attention. However, existing studies have not achieved sufficient efficiency and effectiveness in recognizing untrained users. To address this, this paper proposes a novel user identification algorithm called M-Identity, which passively identifies users by recognizing the characteristics of the footprint in their footsteps. The algorithm extracts mel-frequency cepstral coefficients (MFCCs) from the footsteps and uses a deep learning network called Id-Net, which is based on the Mobile Vision Transformer (MobileViT), for classification. To address the issue of new users, M-Identity adopts transfer learning by freezing the feature extractor parameters of Id-Net and fine-tuning them with lightweight training, achieving high recognition performance. To adapt M - Identity to various environmental conditions, this paper applies spectral subtraction and autocorrelation-based methods to remove noise from the surroundings. The M-Identity model is implemented via a commercial microphone array and evaluated in indoor environments. The experimental results demonstrate that M-Identity achieves a user identification accuracy of 95.0%, with a superior performance in addressing the new user problem compared with existing approaches.