Uncertainty of 3D facial features measurements and its effects on personal identification

• Used AAM to extract two 2D features masks from facial images and triangulated.
• Multi-camera system frames stereoscopic images of individuals in different positions.
• Used weights to compensate landmarks uncertainty.
• Threshold applied to reduce False Acceptances.

Most of face recognition algorithms work fine when applied under controlled lighting conditions, proper distance and orientation of the subject from the camera and neutral expressions. In this work, effect of partial face rotation on recognition rate has been evaluated using a multi-camera acquisition system which allows a synchronous stereoscopic acquisition from three points of view. A feature-based approach to the face-recognition will be presented, using a set of landmarks to outline facial biometrics. Recognition is performed by evaluating weighted measurement discrepancies between 3D geometric masks obtained from stereoscopic images. A database of facial pictures of 117 individuals, framed in three different positions has been used. Recognition tests have been repeated four times, considering natural expression changes in repeated acquisitions. The results evidence a 96.2% correct recognition rate in 117 * 117 * 4 = 54756 comparisons. At a later stage, recognition system has been experimented splitting the database into a population of 90 authorized users and 27 impostors with the intent to reduce False Acceptances and False Identifications with a measurement threshold. Reliability improvements of recognition based on redundant measurements as well as in repeated tests is also analyzed.