Evidence for "liveness" in face biometrics

Abstract: Aliveness or "liveness" detection is a novel but crucial topic in biometrics. In this study we will investigate it for face biometrics. The problem is given by the real-time requirements of such a system to be able to avert advanced spoofing attempts, e.g. by photographs or replayed videos via a mobile device presented to the face authentication automat by the spoofer. We have been developing efficient methods for object detection and motion analysis/tracking, which we have integrated into effective anti-spoofing measures, focusing on I: Mouth movements, II: Eye-blinking and, III: 3D properties of a real face (head). A multiexpert approach - being designed to avert known and novel attacks - that accumulates evidence for "liveness" during a short time is suggested. This thesis is grouped into three main parts (besides the introduction), which also represent its main contributions: Part one revisits a landmark-based object detection method, after which it introduces a holistic object detection method that employs illumination invariant features, termed "quangles" (quantized angle features). By using the latter a histogram-related image preprocessing becomes obsolete, and also the training of the according classifier excels in rapidness. These detectors are employed for the (scale invariant) detection of faces and localization of interesting regions within them. In part two we introduce the OFL - Optical Flow of Lines - which, by evaluating only "motion of lines" represents an efficient algorithm for motion estimation being particularly appropriate here. It is employed to retrieve local measurements, on the one hand for steering the tracking, and on the other hand for "liveness" detection purposes, e.g. for detecting eye-blinking. This and other anti-spoofing measures are detailed in the third part of this thesis that is devoted to "liveness" detection in face biometrics. We have been using both several public databases (e.g. CMU-MIT frontal face test set, XM2VTS database, SOFA test sequences) and laboratory data (spoofing equipment) for the experiments.