3D face registration across pose variation and facial expression using cross profile alignment
In a 3D face recognition system, face registration is usually employed to compensate the pose variation in a 3D face model. Most previous methods in 3D face registration are based on the well known global-based approach, Iterative Closest Point (ICP). The experiments are usually conducted using clea...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English |
| Published: |
2011
|
| Subjects: | |
| Online Access: | http://psasir.upm.edu.my/id/eprint/42276/1/FK%202011%2081R.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | In a 3D face recognition system, face registration is usually employed to compensate the pose variation in a 3D face model. Most previous methods in 3D face registration are based on the well known global-based approach, Iterative Closest Point (ICP). The experiments are usually conducted using cleaned and frontalviewed face models, neglecting the facial variation that often occur in real-time scenarios, such as pose variation, facial expression, facial outliers and occlusion. The proposed thesis uses a local-based approach known as Cross Profile Alignment (CPA) as an alternative to the global-based approach, utilizing the facial feature of a face surface as an attempt to cater all the above problems. Among all features on a face surface, nose tip is the most commonly used feature for
facial feature landmarking. It is crucial to accurately detect the nose tip as it will affect the overall performance of the registration process. Most of the presented nose tip detection algorithms were developed merely based on the assumption that the nose tip is the highest point on a face, which is not robust enough for face model under large rotation variation and having large facial outliers. Thus, as the first step prior face registration, the thesis proposed a novel nose tip region detection algorithm
using localized point signature, developed specially to locate the nose tip region across various facial variation. The experiment conducted on challenging 3D face databases yields good results with
94.77% detection rate for the nose tip region detection algorithm. Based on the nose tip region location, a cross-profile is extracted and face model is compensated for rotation variation and translation displacement. The registration framework with CPA which gained accuracy rate of 93.9% when tested within 10 degrees error margin, outperforms the registration framework with ICP using Average Face Model (AFM) with accuracy rate of 87.7%, with lower processing time. The findings during this work indicate the accuracy and the reliability of the proposed registration framework towards 3D face model with challenging facial variation. |
|---|