Semi-automatic sign segmentation from sign language video corpus

Many researches focus on the study of automatic sign language recognition. Many of them need a large amount of data to train the recognition systems. Our work addresses the annotation of sign language video corpus in order to collect training data. We propose a robust tracking algorithm for hands and head, a method to segment hands during occlusions and an approach to segment gestures using motion and hand shape features. In order to show the advantages and limitations of the proposed approaches, we have evaluated each one using international corpus. The full sign segmentation approach shows promising results.