Comparison of 2D&3D Performances of Facial Feature Analysis Using RGB-D Vision Sensor

This study was conducted to experimentally identify a better method of facial expression recognition using the AAM (Active Appearance Model), a two-dimensional image-based method and three-dimensional depth information method. Experiments were performed by com-paring facial feature points for happy facial expressions and neutral facial expressions, and analyzed the 5000 sets of 2D and 3D facial feature point data of university student subjects. As a result of the analysis, it is confirmed that the same facial feature vector change is more clearly distinguished when using the 2D image based method than the 3D depth information based method. It is confirmed that this phenomenon is caused by the fundamental problem of structured light type RGB-D camera which causes error up to 15 mm at 1 m depth. Consequently, the 3D method can be advantageous when facial expression recognition through AAM is frequent in the depth direction or facial pose variation is large, but 2D method has excellent performance for accurate facial recognition in a static situation respectively.