Bayesian score level fusion for facial recognition
Marco F. Huber, Andreas Merentitis, Roel Heremans, Maria E. Niessen, Christian Debes, Nikolaos Frangiadakis · 2016
Partial occlusions, changing lighting conditions, or rapid motion of persons are some reasons why the recognition rate of a facial recognition (FR) system can be very low. One approach that has gained increased interest in the recent years for compensating the limitations of a single system is to fuse the detections of multiple FR systems. In this paper, a novel fusion algorithm operating on match score level is proposed that follows Bayesian inference and decision theory. It is designed for on-line recognition and facilitates the incorporation of temporal correlation between detections. The proposed approach is compared against the state-of-the-art by means of standard FR benchmarks and an extensive person detection experiment in an office environment.