Facial Feature Tracking: Proposed System
Tracking using LK
Position of all stored templates are estimated using image alignment (Lucas Kanade).
Reweight Trackers
The tracker confidence is estimated using tracking and model fit from previous frame. Trackers that closely agree with the model are assigned a higher weight. This reduces the effect of outliers cause in the tracking stage.
Robust Head Pose Estimation
The head pose is estimated by a method based on LM minimisation. Because the minimisation is sensitive to tracking outliers causing local minima, this technique is wrapped in a RANSAC framework. This prodives a robust estimation of head pose.
Update Tracker Positions
Tracker positions are updated based on confidence weights, the tracking position estimates and the model fit.
Online Addition of Templates
If the face is a previously unseen orientation, additional templates are collected to store the current appearance of the features.
Results
| Average Error (Deg) | |||
| Method | Pitch | Roll | Yaw |
| Proposed method | 3.9 | 3.1 | 4.2 |
| Jang and Kanade 2008 | 3.7 | 4.6 | 2.1 |
| Choi and Kim 2008, Cylinder | 4.4 | 5.2 | 2.5 |
| Choi and Kim 2008, Ellipsoid | 3.9 | 4.0 | 2.8 |
Further details are available in the published paper.
Last update: July 2009.