Increasing demands from security applications (e.g., surveillance, secure access, human/computer interface) and the availability of cheap and powerful hardware led to the development of many commercial face recognition systems. Most of the commercially available face recognition systems have used 2D images of human faces, the reason being the cost effectiveness and easy availability of 2D sensors. However, 2D face recognition techniques are known to suffer from the inherent problems of illumination, pose, and are sensitive to factors such as occlusion, change in human expression, and aging. The appearance of human faces is subject to several different factors mentioned above. As stated by Moses et al. “The variations between the images of the same face due to illumination and viewing directions are almost larger than image variations due to the change in the face identity”

Utilizing 3D face information was shown to improve face recognition performance, especially with respect to pose variations. Range images captured by 3D sensor provide much more information than a conventional 2D sensor. These models are more accurate because the range sensor captures absolute measurements invariant to camera distance. Since the complete geometry of a person’s face is available instead of just color and texture, the models are invariant to illumination change. Pose normalization in 3D space turns out to be a significant advantage of such a technology. This is in contrast to the pose normalization from 2D images, which is a significant challenge considering that information is lost in the transformation from the 3D world to a 2D image. Also, enough invariant information is present to cope with change in expressions and other occlusions such as glasses and beard.

We propose a technique that uses 3D geometric (point sets) face representations. The use of 3D point sets to represent human faces in lieu of 2D texture makes this method robust to changes in illumination and pose. The method first automatically registers facial point-sets through a criterion based on Gaussian force fields. The registration method defines a simple energy function, which is always differentiable and convex in a large neighborhood of the alignment parameters; allowing for the use of powerful standard optimization techniques. The new method overcomes the necessity of close initialization and converges in much less iterations as compared to the Iterative Closest Point algorithm (ICP) [1]. The use of an optimization method, the Fast Gauss Transform, allows a considerable reduction in the computational complexity of the registration algorithm. Recognition is then performed by using the robust similarity score generated by registering 3D point sets of faces.