No abstract

The automatic recognition of sign language is very important to allow for communication by hearing impaired people. The purpose of this study is to develop a method of recognizing the static Mexican Sign Language (MSL) alphabet. In contrast to other MSL recognition methods, which require a controlled background and permit changes only in 2D space, our method only requires indoor conditions and allows for variations in the 3D pose. We present an innovative method that can learn the shape of each of the 21 letters from examples. Before learning, each example in the training set is normalized in the 3D pose using principal component analysis. The input data are created with a 3D sensor. Our method generates three types of features to represent each shape. When applied to a dataset acquired in our laboratory, an accuracy of 100% was obtained. The features used by our method have a clear, intuitive geometric interpretation.

Few hearing people know and use Mexican Sign Language (MSL). Consequently, this is the main barrier between deaf and hearing people. This study proposes a system that recognizes and animates in real time a set of signs belonging to the semantic field of a general medicine consultation service. Therefore, a linkage between a hearing doctor and a deaf patient can be established in a non-intrusive way and with an easy dynamic interaction. Our main contribution is a bidirectional translator system for Mexican sign language in the context of a general medical consultation. The recognition module uses a MS Kinect sensor to obtain the sign trajectories, and images to feed hidden Markov Models (HMMs) for processing signs samples in real time. The experiments show recognition of 82 different signs from 22 participants. As a result, accuracy and F1 scores average rates of 99% and 88% were obtained.