Towards Continuous User Recognition by Exploring Physiological Multimodality: An Electrocardiogram (ECG) and Blood Volume Pulse (BVP) Approach

Work on pervasive systems for performing continuous identity recognition has become a major research line due to the more recent developments on wearable and more user-friendly sensors. In this paper we present a multibiometric system based on physiological signals and integrated in a human-computer interaction (HCI) setup, which includes the electrocardiogram (ECG) and blood volume pulse (BVP). Feature extraction was performed on waveform morphology, and k-Nearest Neighbors and Naive-Bayes decision level fusion classifiers were used to perform identification and authentication tests. Furthermore, our approach is based on signal windowing, targeting a near real-time and continuous recognition application scenario. Results show that the BVP signal did not add value to improve the performance of the multimodal approach, but the combined use of windows of different lengths for the ECG modality can yield an increase in the performance. Tests were performed using within- and across- session data, to assess the stability of the signals over time, and the generalization ability of the classifiers. Rank-1 Error of Identification (EID) values of approximately 2% and 8% were obtained respectively in within and across-session identification tests while, in authentication, Equal Error Rate (EER) values of approximately 4% and 13% were achieved.