Finding Relevant Image Content for mobile Sign Language Recognition

Abstract

We are currently developing a vision-based sign language recognition system for mobile use. This requires operability in different environments with a large range of possible users, ideally under arbitrary conditions. In this paper, the problem of finding relevant information in single-view image sequences is tackled. We discuss some issues in low level image cues and present an approach for the fast detection of a signing persons hands. This is achieved by using a modified generic skin color model combined with pixel level motion information, which is obtained from motion history images. The approach is demonstrated with a watershed segmentation algorithm.