Behavioral Gait Biometrics in VR: Is the Use of Synthetic Samples Able to Increase Person Identification Metrics?

Abstract

In this paper, we present an approach to build a biometric system capable of identifying subjects based on gait. The experiments were carried out with a proprietary gait corpus collected from 100 subjects. In the data acquisition process, we used a commercially available perception neuron body suit equipped with motion sensors and dedicated to entertainment in the VR domain. Classification was performed using two variants of the CNN architecture and evaluated using cross-day validation. A novelty in the presented approach was the exploration of research areas related to the usage of synthetically generated samples. Experiments were conducted for two types of preprocessing—a low-pass filtering of the signals using a 3rd- or 1st-order Butterworth filter. For the first variant, the synthetic samples generated by the long short-term memory-mixture density network (LSTM-MDN) model allowed us to increase the F1-score from 0.928 to 0.966. Meanwhile, in the second case from 0.970 to 0.978 F1-score.