Face recognition from 3D data using Iterative Closest Point algorithm and Gaussian mixture models

An approach to face verification from 3D data is presented. The method uses 3D registration techniques designed to work with resolution levels typical of the irregular point cloud representations provided by structured light scanning. Preprocessing using a-priori information of the human face and the Iterative Closest Point algorithm are employed to establish correspondence between test and target and to compensate for the nonrigid nature of the surfaces. Statistical modelling in the form of Gaussian mixture models is used to parameterise the distribution of errors in facial surfaces after registration and is employed to differentiate between intra- and extra-personal comparison of range images. An equal error rate of 2.67% was achieved on the 30 subject manual subset of the 3d_rma database.