With the advancement of artificial intelligence, biometric authentication systems have become widely used. In particular, uni-modal biometric authentication such as face recognition has become standard in various devices such as mobile phones. However...
With the advancement of artificial intelligence, biometric authentication systems have become widely used. In particular, uni-modal biometric authentication such as face recognition has become standard in various devices such as mobile phones. However, the effectiveness of uni-modal systems sometimes decreases when applied in real-world environments including occlusion and viewpoint changes. To address this limitation, this dissertation introduces multimodal approaches for authentication through three studies. The three studies focus on the following scenarios: (i) occluded faces, (ii) gait under varying viewpoints, and (iii) combined occlusion and viewpoint variation in face and gait.
The first study presents DemoID, a multimodal authentication model for occluded faces. DemoID integrates facial, vocal, and demographic attributes to identify individuals. By combining voice and demographic attributes, DemoID compensates for occluded facial features and outperforms the unimodal recognition method by improving 9.00% accuracy. This study demonstrates the importance of demographic features and multimodal approaches in improving user authentication.
The second study presents DualGait, a gait-based demographic attributes recognition model. DualGait integrates gait energy image (GEI) and 3D pose data, including 14 viewpoints, to predict the age and gender of individuals. By fusing global and local features between GEI and 3D pose, DualGait captures high and low interactions between them. DualGait achieves state-of-the-art performance compared to existing models and shows great performance from all angles.
The final study introduces the Covariate-Aware Face and Gait (CAFG) dataset, which includes various real-world covariate conditions. To construct the CAFG dataset, I recruited 20 subjects across 65 different conditions, covering variations such as face coverings, items carried, and clothing styles. This dataset is designed to reduce the gap between controlled environments and the real world for face and gait recognition. The experimental results show that the CAFG dataset has challenging factors for face and gait recognition compared to previous benchmarks.
This dissertation extends previous biometric authentication methods from unimodal approaches to multimodal approaches. The three studies within this dissertation contribute to the field of biometric authentication by improving security and reliability in real-world scenarios under occlusion and changing viewpoints.