Sign Segmentation Using Dynamics and Hand Configuration for Semi-automatic Annotation of Sign Language Corpora

摘要

This paper address the problem of sign language video annotation. Nowadays sign language segmentation is manually performed. This is time consuming, error prone and no reproducible. In this paper we intend to provide an automatic approach to segment signs. We use a particle filter based approach to track hands and head. Motion features are used to classify segments performed with one or two hands and to detect events. Events that have been detected in the middle of a sign are removed considering hand shape features. Hand shape is characterized using similarity measurements. Evaluation has been performed and has shown the performance and limitation of the proposed approach.