| Traditional 2D face recognition systems are not tolerant to changes in pose, lighting and expression. This dissertation explores the use of 3D data to improve face recognition by accounting for these variations. A two step, fully automatic, 3D surface alignment algorithm is developed to correlate the surfaces of two 3D face scans. In the first step, key anchor points such as the tip of the nose are used to coarsely align two face scans. In the second step, the Iterative Closest Point (ICP) algorithm is used to finely align the scans. The quality of the face alignment is studied in depth using a Surface Alignment Measure (SAM). The SAM is the root mean squared error over all the control points used in the ICP algorithm, after trimming to account for noise in the data. This alignment algorithm is fast (<2 seconds on a 3.2GHz P4) and robust to noise in the data (<10% spike noise). Extensive experiments were conducted to show that the alignment algorithm can tolerate up to 15° of variation in pose due to roll and pitch, and 30° of variation in yaw. It is shown that this level of pose tolerance easily covers the normal pose variation of a database of over 275 cooperative subjects. By using the SAM as an initial matching score and automatically rejecting poor quality scans, an equal error rate of 1.2% is achieved on a database of 943 scans from 275 subjects. This surface alignment verification system is fast, fully automatic, and requires no additional training. By using the 3D face surface alignment algorithm, a Canonical Face Depth Map (CFDM) is defined to allow for automatic preprocessing of 3D face scans. The CFDM is shown to help in anchor point localization, reduce model memory requirements, and enables other algorithms, such as PCA or correlation, to be applied. |
