Emotion Recognition Using Feature Extraction and 3-D Models

This paper describes an integrated system for human emotion recognition. While other techniques extract explicit motion fields from the areas of interest and combine them with templates or training sets, the proposed system compares evidence of muscle activation from the human face to relevant data taken from a 3-d model of a head. This comparison takes place at curve level, with each curve being drawn from detected feature points in an image sequence or from selected vertices of the polygonal model. The result of this process is identification of the muscles that contribute to the detected motion; this conclusion is then used in conjunction with neural networks that map groups of muscles to emotions. The notion of describing motion with specific points is also supported in MPEG-4 and the relevant encoded data may easily be used in the same context. Key-Words : Expression recognition, 3-d muscle mesh, feature extraction, motion estimation CSCC'99 Proc.pp.5371-5376