A system that could enable fast and robust facial expression recognition would have many applications in behavioral science, medicine, security and human-machine interaction. While working toward that goal, we do not attempt to recognize prototypic facial expressions of emotions but analyze subtle changes in facial behavior by recognizing facial muscle action units (AUs, i.e., atomic facial signals) instead. By detecting AUs we can analyse many more facial communicative signals than emotional expressions alone. This paper proposes AU detection by classifying features calculated from tracked fiducial facial points. We use a Particle Filtering tracking scheme using factorized likelihoods and a novel observation model that combines a rigid and a morphologic model. The AUs displayed in a video are classified using Probabilistic Actively Learned Support VectorMachines (PAL-SVM). When tested on 167 videos from the MMI web-based facial expression database, the proposed method achieved very high recognition rates for 16 different AUs. To ascertain data independency we also performed a validation using another benchmark database. When trained on the MMI-Facial expression database and tested on the Cohn-Kanade database, the proposed method achieved a recognition rate of 84% when detecting 9 AUs occurring alone or in combination in input image sequences.
pubs.doc.ic.ac.uk: built & maintained by Ashok Argent-Katwala.