Smartphone-Based Biometric System Involving Multiple Data Acquisition Sessions

This paper presents an analysis of a mobile phone-based gait biometrics system that involves several data acquisition sessions. The conducted work verified the influence of the number of training sessions on the performance of the biometric system. Experiments were conducted using a publicly available 14-person database in which individuals attended three data acquisition sessions. A CNN with an attentional mechanism architecture was used as a baseline classifier. According to the experimental results, the system based on one motion tracking session achieved a performance of approximately 0.50 and 0.62 F1-score for the raw and processed data, respectively. In contrast, for two motion tracking sessions, the performance of the system was 0.60 and 0.85 F1-score for raw and processed data, respectively. The study showed that accuracy greater than 0.8 F1 score could be achieved when data preprocessing was applied and measured data from two motion sessions were incorporated into the training process.