Ubiquitous Face-Ear Recognition Based on Frames Sequence Capture and Analysis

摘要

Unimodal biometric systems performance is known to be easily affected by intra-class variations, noisy samples, spoofing techniques and environmental conditions. These problems get even more challenging whenever biometric data acquisition is performed "in-the-wild". Some of these limitations can notably be addressed by means of multi-biometric approaches, exploiting different biometric traits, multiple samples and multiple algorithms to establish the identity of an individual. To this regard, the present study describes a face+ear biometric system requiring just a single combined video capture of the subject's face to work in a ubiquitous operative scenario. Exploiting the video capture capabilities provided by most smartphones' built-in cameras, the proposed method acquires subject's face both frontally and sideways within a single video sample. The resulting frames sequence is then analyzed to find the ones most suited, quality wise, to feed the two parallel biometric pipelines. Different data-fusion strategies, working either at score level with quality-based adaptive weighting or at decision level, have been applied to the output of face and ear matching stages to the aim of improving system's accuracy and reliability. Preliminary experimental results show good recognition accuracy coupled to an unusual easiness of operation for a ubiquitous multimodal biometric system.