Abstract
This paper discusses a vision-based human understanding application in mobile devices and reports an efficient mouth boundary detection and mouth state estimation algorithm. The proposed algorithm is insensitive to the illumination changes. 400 head-&-shoulder face images are used for evaluation. The accuracy in mouth boundary detection and the mouth state estimation are 83.5% and 79.75% respectively. The computational time in Pentium III 700MHz PC using matlab implementation is less than one minute. It is reasonable to estimate that if the algorithm is implemented using low level language or assembly language, the processing time should be less than one second.