Project Detail

3D Face Authentication

3D Face Authentication for Biometric Access Control

Lead: Razdan, Anshuman
Collaborators: Farin, Gerald;Lockwood, Charles;Zhang, Liyan;Bae, Myungsoo
RA: Chaudhari, Mahesh
Sponsor: NSF-Directorate for Computer & Information Science and Engineering
Date: 08/15/2003  - 07/31/2007
Web site: http://3dface.prism.asu.edu/

Abstract

3D digital face scanning is now being investigated for biometric analysis and authentication at PRISM. Supported by a 3 year grant from the National Science Foundation, the new project focuses on the development of intelligent and fast algorithms for representation, extraction, segmentation, query and matching of 3D facial shapes for authentication. Unlike current biometric access and analysis technologies, the project is taking advantage of the 3rd dimension to greatly increase accuracy and reliability of the data, allowing for more complex algorithms and analysis to be put to work. Researchers are looking at face geometry and curvature to analyze facial features that are invariant to expressions and other changes, such as facial hair.

A multi camera based 3D face scanner used at the PRISM lab at ASU. The scanned data included both 3D and 2D image. Note the missing data corresponding to the hair of the subject.

Biometric access control focuses on measurable physiological or behavioral trait to automatically and accurately authenticate or verify the identity of that person. Biometric characteristics must ideally be distinctive to the individual, easily acquired/measurable and able to be compared or encoded for the security validation. The characteristic should change little over time (i.e. aging) and be difficult to change, circumvent, manipulate, or reproduce by other means. The mainstream biometric technologies use morphological feature recognition such as fingerprints, hand geometry, iris and retina feature scanning, and face recognition based on extracting characteristics of the face. Each of these except face recognition is either intrusive or fails in some cases (for example: fingerprint identification is not possible for about 10% of population with indistinct fingerprints). Current face recognition technologies use algorithms ranging from heuristics to artificial intelligence to locate facial features, such as the eye socket, cheekbones, and sides of the mouth, in two dimensional (2D) black/white or color space. However, even the most sophisticated algorithms in 2D domain have a high failure rate. There has been an emergence of 3D technologies to combat the shortcomings of biometrics based on 2D images. In this chapter we review the current state of the art in 2D and 3D face recognition/authentication followed by in depth description of 3D face biometrics being developed at PRISM (Arizona State University) for homeland security application.

Terrorist events of September 11, 2001 have brought about the realization that the United States needs stronger national security measures and protocols. The primary goal of access control systems is to restrict access to authorized personnel. The return on the investment provided is based on the benefits to personnel safey and offsetting other risks resulting from unlawful access, fraud, and theft. Access control applications include door access, time & attendance, keep-out during off-hours, and the control of sensitive and restricted access points.

Recognition in one-to-many searches seeks to identify an unknown individual based on comparison to a database of known individuals (e.g., law enforcement, surveillance, and recently driver licenses). Authentication involves performing verification based on a one-to-one search to validate the identity claim of the individual (i.e., access control for a building, room, or for making a transaction at an ATM terminal). Authentication is in one sense a simpler process: comparisons are made only to the claimed identity, and a threshold of similarity is used to accept or reject the claim. In another sense, authentication is more difficult, because of the need to determine this threshold rather than using a "best match" criterion as in many face recognition applications.

Several approaches have been promoted to recognize and authenticate an individual or a group of people. Access control applications authenticate by physical appearance (by guard personnel, receptionist); by something the individual knows (pins, passwords); by something the individual has (lock/key, card, badge, token); by biometric evidence (a unique physiological or behavioral characteristic of the individual); or by a combination of both ?what one has? (i.e., a card) and "what one knows" (i.e., their passcode). Most workplace entry points are typically controlled by a badge/card or by physical appearance. All of these methods, except biometrics, are fallible and can be circumvented, lost, or stolen. For that reason, biometrics have been explicitly cited in several pieces of U.S. legislation, including the USA PATRIOT Act (signed in October 2001), the Aviation and Transportation Security Act (signed in November 2001), and the Enhanced Border Security and Visa Reform Act (signed in May 2002). Each calls for the implementation of biometric technology to enhance homeland security.

Interest in authentication using biometrics is therefore growing dramatically. Biometric access control uses measurable physiological or behavioral traits to automatically authenticate a person's identity. Biometric characteristics must be distinctive of an individual, easily acquired and measured, and comparable for purposes of security validation. The characteristic should change little over time (i.e., with age or voluntary change in appearance) and be difficult to change, circumvent, manipulate, or reproduce by other means. Typically, high-level computer based algorithms and database systems analyze the acquired biometric features and compare them to features known or enrolled in the database. The mainstream biometric technologies use morphological feature recognition such as fingerprints, hand geometry, iris and retina scanning, and 2D and 3D face authentication. Each of these except face authentication is either intrusive or fails in some cases (e.g., about 10% of population do not have good enough fingerprints). 2D face authentication, though less intrusive, has simply not attained the degree of accuracy necessary in a security setting. Facial biometrics on one hand present opportunities but at the same time acquiring correct biometrics from facial images continues to be an open research problem. Some issues are biometrics invariant to pose and expression, excessive facial hair and glasses that cover large part of the face, data processing time, etc.


Our approach is a combination of feature and profile based methods. We focus on finding accurate and robust features without any user intervention. The system is divided into three parts. These are: data acquisition, feature extraction, and authentication or recognition.

The 3D dataset for this project is acquired with a combination of commercial scanning technologies and research software applications developed at PRISM. The 3D Face Scanner is a 2-pod scanning system from 3Q Inc.6 This scanning system is quick enough to take the 3D face scan in comparison with the other laser scanners. The time taken for the cameras to operate is a fraction of a second while rest of the work is done by the software provided by the same company to generate a triangle mesh of the face, which takes a few seconds to complete.

The scanner does not give proper data on hair and glasses, a limitation of this and almost all other methods in the area of facial biometrics. The triangle mesh representation is the most common way to represent 3D data from scanning systems. The feature extraction phase performs face classification using curvatures, registration of a face, finds symmetry plane, critical points, and profile curves, nose and biometrically relevant sub- face area extraction. Finally, authentication/recognition module performs comparison between two faces.

Feature Extraction There are 5 steps in feature extraction, which include (1) surface point classification by curvatures, (2) approximated nose tip extraction and registration, (3) finding symmetry plane, (4) critical points and profiles extraction, and (5) nose and sub-face extraction.

We do not assume any thing about the orientation (which way is up) or the tilt of the head in a given scan or input data. Many algorithms require this as additional information before performing comparison but we feel that our methods do not need this additional information and are able to overcome issues of pose/orientation. This leads to the point that we need to find specific features such as the nose tip from a give scan. The scan may include parts of the shoulder, neck, etc. therefore it is important that any automated method should zero in on facial features without user intervention.