3D Face Recognition - Data Processing: Registration and Deformation
M.M.J. Gerlach, C.T. Rooijers · Research Repository (Delft University of Technology) · 2013
A 3D face recognition algorithm has been developed for the Microsoft Kinect in the scope of the final bachelor project at the TU Delft in 2013. The aim of the project is to develop a prototype face recognition system. The prototype system had to outperform the existing 2D face recognition system. The project is divided into three subgroups each designing a specific part of the system. The three subgroups are: the data acquisition group, the data processing group and the data comparsion group. This thesis provides the data processing part of the system. An overview of the different existing algorithms is made, followed by the requirements for the system. The algorithm should give reliable results in a reasonable time and should be pose, expression and illumination invariant. A 2D face recognition algorithm, PCA, was implemented first. The 3D face recognition algorithm follows a morphable model approach. Several algorithms are used to align and fit the 3D face scan from the Kinect. To align the scan with a model, ICP and spin-images are used also referred to as registration. Deformation is done by a nonrigid ICP algorithm to fit the model with the scan. From the fitted model a geometry image and a normal image is generated. For prototyping reasons Matlab was used to implement and test the algorithm. The developed algorithm is tested by several measurements. From the results of these measurements, it could be concluded that the algorithm is robust and reliable, but isn’t fast. The proposed alignment solution is able to deal with rotation between ?pi/2 and pi/2 about any of the three axes and with translation in any of the three direction in the range of 1 till 10 cm. The proposed solution is able to improve up till 12% overall in the case of rotation compared to conventional spin-images and up till 43% in the case of translation. A speed-accuracy trade-off have been made. Furthermore, many optimization can still be made. More research will be needed to fulfill the system.