Using Adaptive Trackers for Video Face Recognition from a Single Sample Per Person

Abstract

Still-to-video face recognition (FR) is an important function in many video surveillance applications, allowing to recognize target individuals of interest appearing over a distributed network of cameras. Systems for still-to-video FR match faces captured in videos under challenging conditions against facial models, often based on a single reference still per individual. To improve robustness to intra-class variations, an adaptive visual tracker is considered for learning of a diversified face trajectory model for each person appearing in the scene. These appearance models are updated along a trajectory, and matched against the reference gallery stills of each individual enrolled to the system. Matching scores per individual are thereby accumulated over successive frames for robust spatio-temporal recognition. In a specific implementation, face trajectory models learned with a STRUCK tracker are compared to reference stills using an ensemble of SVMs per individual that are trained a priori to discriminate target reference faces (in gallery stills) versus non-target faces (in videos from the operational domain). To represent common pose and illumination variations, domain-specific face synthesis is employed to augment the number of reference stills. Experimental results obtained with this implementation on the Chokepoint video dataset indicate that the proposed system can maintain a comparably high level of accuracy versus state-of-the-art systems, yet requires a lower complexity.