Animating Highly Constrained Deformable Head/Face Models Using Motion Capture
Degree GrantorUniversity of Canterbury
We investigate, implement and evaluate algorithms and techniques that attempt to solve three key challenges when animating highly constrained deformable models by way of optical motion capture. We present two techniques to solve the pose estimation problem for rigid bodies, one based upon 3 fiducial markers falling on a plane and another built upon the ARToolkit. Further, we compare three techniques for fiducial marker segmentation, namely HSV colour space segmentation, camera hardware filtering and UV illumination segmentation. Lastly, we propose a hybrid pupil tracking algorithm combining Haar face detection, anthropometric localisation, pattern matching and row vs column intensity histograms. We test the performance of our pupil detection algorithm on a 285 frame video displaying a variety of gaze directions. Our hybrid approach performs well, resulting in a very low pixel error and a reasonable frame rate. The three segmentation approaches are tested on four 1000 frame and four 100 frame videos containing inherently different movement. All approaches show excellent segmentation accuracy but illustrate the importance of understanding the limitations of each technique prior to implementation. Both pose estimation techniques are implemented and tested on objects placed in a variety of poses. Both algorithms show good approximations of pose.