next up previous
Next: Motion Field Histograms Up: FACEFACTS: Modeling Natural Facial Previous: Data Collection

Head Tracking and Feature Extraction

In order to analyze unconstrained video which includes significant head movement, it is necessary to know changes in the head pose as well as the facial features. We used three existing 3D model-based head-trackers developed by our group [2,10,13]. We use the output to normalize and warp the face to a frontal position. In our experience no existing head tracker is able to track unconstrained data for a extended period of time, so the output normalized images have errors in position and scale. Consequently, it is very important to build in robustness into the extracted features. Evidence suggests that when people are engaged in a conversation the most frequently occurring movements are raising the eye-brow, lowering the eye-brow and some form of smiling [9]. Thus, we decided to automatically extract the eyebrows, eyes and mouth region from the normalized images.



 

Tanzeem Choudhury
2000-01-21