Adaptive ensembles for face recognition in changing video surveillance environments

Recognizing faces corresponding to target individuals remains a challenging problem in video surveillance. Face recognition (FR) systems are exposed to videos captured under various operating conditions, and, since data distributions change over time, face captures diverge w.r.t. stored facial models. Although these models may be adapted when new reference videos become available, incremental learning with faces captured under different conditions may lead to knowledge corruption. This paper presents an adaptive multi-classifier system (AMCS) for video-to-video FR in changing surveillance environments. During enrolment, faces captured in reference videos are employed to design an individual-specific classifier. During operations, a tracker allows to regroup facial captures for individuals in the scene, and accumulate the predictions per track for robust spatiotemporal FR. Given a new reference video, the corresponding facial model is adapted according to the type of concept change. If a gradual pattern of change is detected, the individual-specific classifier(s) are adapted through incremental learning. To preserve knowledge, another classifier is learned and combined with the individuals previously-trained classifier(s) if an abrupt change is detected. For proof-of-concept, the performance of a particular implementation of this AMCS is assessed using videos from the Faces in Action dataset. By adapting facial models according to changes detected in new reference videos, this AMCS allows to sustain a high level of accuracy comparable to the same system that is always updated using a learn-and-combine approach, while reducing time and memory complexity. It also provides higher accuracy than incremental learning classifiers that suffer the effects of knowledge corruption.