The quality of recognition systems for continuous utterances in signed languages could be largely advanced within the last years. However, research efforts often do not address specific linguistic features of signed languages, as e.g., non-manual expressions. In this work, we evaluate the potential of a single video camera-based recognition system with respect to the latter. For this, we introduce a two-stage pipeline based on two-dimensional body joint positions extracted from RGB camera data. The system first separates the data flow of a signed expression into meaningful word segments on the base of a frame-wise binary Random Forest. Next, every segment is transformed into image-like shape and classified with a Convolutional Neural Network. The proposed system is then evaluated on a data set of continuous sentence expressions in Japanese Sign Language with a variation of non-manual expressions. Exploring multiple variations of data representations and network parameters, we are able to distinguish word segments of specific non-manual intonations with 86% accuracy from the underlying body joint movement data. Full sentence predictions achieve a total Word Error Rate of 15.75%. This marks an improvement of 13.22% as compared to ground truth predictions obtained from labeling insensitive towards non-manual content. Consequently, our analysis constitutes an important contribution for a better understanding of mixed manual and non-manual content in signed communication.
Keywords: gesture information retrieval; learning systems; motion segmentation; neural networks; sign language; signal processing.