부산대학교 도서관

Sign Language uses manual hand and body gestures as well as non-manual facial expressions as a means of communication between deaf and hearing communities. A communication divide exists between the deaf and hearing communities, to the disadvantage of the deaf. This paper explores the design and implementation of a mobile application for South African Sign Language recognition. The application connects, via Bluetooth, to an instrumented glove developed by the University of Cape Town. The objective is to recognize the manual alphabet and manual numeric digits that have static gestures (31 signs in total). Two neural networks (one with a log-sigmoid, and one with a symmetric Elliott activation function) and a Support Vector Machine (SVM) were compared. The SVM was chosen for implementation primarily because of its high accuracy of 99% and its superior robustness. The mobile application is developed for Android and allows the user to connect to a Bluetooth glove, display and dictate the classification output and calibrate the connected glove. On a low-end smartphone, the classification time did not exceed 45ms, the memory usage did not exceed 15 MB, and the battery life during typical usage was approximately 11 hours.