3D facial recognition using local feature-based methods and accuracy assessment

As 3-dimensional point cloud can be obtained easily with laser scanning technology, three-dimensional face recognition have become popular against the facial recognition performed using 2D images that has many limitations.In this study, the facial data of 10 people were modeled in 3D using a laser scanner. A total of 30 point clouds were taken from 10 people-two natural facial expressions and one laughing facial expression. The algorithm consists of three steps. In the first step, 3D points are defined on the point clouds using ISS and LSP methods. In the second step, key points were described using PFH and FPFH methods to obtain feature histogram. In the third step, the key points in different point clouds were matched using the feature histograms via the Kullbeck-Leiber Divergence method. For accuracy analysis, point clouds are registered with Iterative Closest Point (ICP) method. For accuracy assessment, the Euclidean distance between the matching points was calculated. The ISS algorithm finds about 25% less points than the LSP algorithm. The correct matching rate for PFH is up to 60%, while FPFH histograms are around 25%-30%.