Movement Epenthesis Detection in Continuous Fingerspelling from a Coarsely Sampled Motion Vector Field in H.264/AVC Video

This paper presents a novel framework for movement epenthesis detection in continuous fingerspelling sequence for H.264/AVC compressed videos. The framework is based on the motion vectors partially decoded from the compressed videos which are already inherent in them. We have used a three-level filter block to remove the stray motion vectors present in the videos. Then the most significant moving part which is the hand and/or fingers in sign language videos are identified. After that the sign and movement epenthesis frames are separated out based on the observation that hand/finger(s) changes their positions faster compared to other objects present in the video during movement epenthesis phase. The proposed framework is tested on a number of American Sign Language video sequences encoded with H.264/AVC JM encoder.