Detail of publication

Abstract

This thesis addresses several issues of automatic sign language recognition, namely the creation of vision based sign language recognition framework, sign language corpora creation, feature extraction, making use of novel hand tracking with face occlusion handling, data-driven creation of sub-units and "search by example" tool for searching in sign language corpora using hand images as a search query. The proposed sign language recognition framework, based on statistical approach incorporating hidden Markov models (HMM), consists of video analysis, sign modeling and decoding modules. The framework is able to recognize both isolated signs and continuous utterances from video data. All experiments and evaluations were performed on two own corpora, UWB-06-SLR-A and UWB-07-SLR-P, the first containing 25 signs and second 378. As a baseline feature descriptors, low level image features are used. It is shown that better performance is gained by higher level features that employ hand tracking, which resolve occlusions of hands and face. As a side effect, the occlusion handling method interpolates face area in the frames during the occlusion and allows to use face feature descriptors that fail in such a case, for instance features extracted from active appearance models (AAM) tracker. Several state-of-the-art appearance-based feature descriptors were compared for tracked hands, such as local binary patterns (LBP), histogram of oriented gradients (HOG), high-level linguistic features or newly proposed hand shape radial distance function (denoted as hRDF) that enhances the feature description of hand-shape like concave regions. The concept of sub-units, that uses HMM models based on linguistic units smaller than whole sign and covers inner structures of the signs, was investigated in the proposed iterative method that is a first required step for data-driven construction of sub-units, and shows that such a concept is suitable for sign modeling and recognition tasks. Except of experiments in the sign language recognition, additional tool \textit{search by example} was created and evaluated. This tool is a search engine for sign language videos. Such a system can be incorporated into an online sign language dictionary where it is difficult to search in the sign language data. This proposed tool employs several methods which were examined in the sign language recognition task and allows to search in the video corpora based on an user-given query that consists of one or multiple images of hands. As a result, an ordered list of videos that contain the same or similar hand configurations is returned.