Hybrid face recognition based on real-time multi-camera stereo-matching

Abstract

Multi-camera systems and GPU-based stereo-matching methods allow for a real-time 3d reconstruction of faces. We use the data generated by such a 3d reconstruction for a hybrid face recognition system based on color, accuracy, and depth information. This system is structured in two subsequent phases: geometry-based data preparation and face recognition using wavelets and the AdaBoost algorithm. It requires only one reference image per person. On a data base of 500 recordings, our system achieved detection rates ranging from 95% to 97% with a false detection rate of 2% to 3%. The computation of the whole process takes around 1.1 seconds. © 2011 Springer-Verlag.