This paper describes a three-dimensional (3-D) face recognition system based on two different 3-D sensors. These sensors were used to overcome pose variation problems that cannot be effectively solved when working with 2-D images. We acquired input data based on a structured light system and compared them with 3-D faces acquired by a 3-D laser scanner. Due to differing data structures, we generated a selection of probe images and stored images (not only for head pose estimation but also for face recognition). Given an unknown range image, we extracted invariant facial features based on facial geometry and utilized the previously developed error-compensated singular-value decomposition method to estimate a head pose. Distinctive facial shape indices were defined and extracted based on facial curvature characteristics. The extracted indices have a different number and different distribution on each face image. When multiple matching possibilities are involved, dynamic programming (DP) is useful matching algorithm. DP merges data points in order to achieve better point-to-point matching by finding a matching path at minimum cost. Experimental results show that the proposed method obtained a 96.8% face recognition rate when working with 300 individuals under different pose variations.
All Science Journal Classification (ASJC) codes
- Atomic and Molecular Physics, and Optics