Biometrics gait system based on motion sensors embedded in a mobile phone: a case study for a two-day training set

Authors

Abstract

This paper presents the results of a study on developing a gait biometrics system based on motion sensors (an accelerometer and gyroscope), embedded in a smartphone. The experiments were conducted using a publicly available 13-person data corpus, with subjects participating in three data collection sessions. The study used CNN, CNN with attention and Multi-Input CNN neural networks. The training scenario from the first day resulted in an accuracy of 0.66 F1 score, 0.71 F1 score for training with the samples from the second day and 0.90 F1 score in the combined sets. It has been shown that it is more profitable to combine historical data than to update it with newer samples. Enriching the training set with a set of 30% synthetic samples produced by the LSTM-MDN generative models allowed to increase to accuracy to 0.94 F1-score. It was shown that synthetic samples can improve the generalization properties of the CNN network.

Additional Files

Published

2025-05-30

Issue

Section

Biomedical Engineering