Abstract
In recent times seismic signal-based person identification has gained significant popularity in the field of biometrics. The existing techniques of person identification are extremely effective and achieve high prediction accuracy. However, these techniques are inefficient in accommodating new users in the system. It requires retraining the model with a large amount of data, which is time-consuming and labour-intensive. To address this issue, we proposed FootsNet, a transfer learning convolutional neural network. FootsNet has a two-stage training process; first, the model is trained using classes with many labelled samples. Then the model is fine-tuned using the novel classes (representing new users), which have a limited number of labelled samples. We perform extensive experiments using datasets collected from 15 human subjects and compare the performance of FootsNet with the existing state-of-the-art techniques. We also study FootsNet's performance by varying the number of labelled footstep samples of the novel classes.