South African sign language dataset development and translation : a glove-based approach

There has been a definite breakdown of communication between the hearing and the Deaf communities. This communication gap drastically effects many facets of a Deaf person’s life, including education, job opportunities and quality of life. Researchers have turned to technology in order to remedy this issue using Automatic Sign Language. While there has been successful research around the world, this is not possible in South Africa as there is no South African Sign Language (SASL) database available. This research aims to develop a SASL static gesture database using a data glove as the first step towards developing a comprehensive database that encapsulates the entire language. Unfortunately commercial data gloves are expensive and so as part of this research, a low-cost data glove will be developed for the application of Automatic Sign Language Translation. The database and data glove will be used together with Neural Networks to perform gesture classification. This will be done in order to evaluate the gesture data collected for the database. This research project has been broken down into three main sections; data glove development, database creation and gesture classification. The data glove was developed by critically reviewing the relevant literature, testing the sensors and then evaluating the overall glove for repeatability and reliability. The final data glove prototype was constructed and five participants were used to collect 31 different static gestures in three different scenarios, which range from isolated gesture collection to continuous data collection. This data was cleaned and used to train a neural network for the purpose of classification. Several training algorithms were chosen and compared to see which attained the highest classification accuracy. The data glove performed well and achieved results superior to some research and on par with other researchers’ results. The data glove achieved a repeatable angle range of 3.27 degrees resolution with a standard deviation of 1.418 degrees. This result is far below the specified 15 degrees resolution required for the research. The device remained low-cost and was more than $100 cheaper than other custom research data gloves and hundreds of dollars cheaper than commercial data gloves. A database was created using five participants and 1550 type 1 gestures, 465 type 2 gestures and 93 type 3 gestures were collected. The Resilient Back-Propagation and Levenberg-Marquardt training algorithms were considered as the training algorithms for the neural network. The Levenberg-Marquardt algorithm had a superior classification accuracy achieving 99.61%, 77.42% and 81.72% accuracy on the type 1, type 2 and type 3 data respectively.

Grantor dc:publisher.institution

Department of Electrical Engineering

Year dc:date.issued

2014

Author dc:creator

• Mcinnes, Ben

Advisor dc:contributor.advisor

• Tsoeu, Mohohlo Samuel

Language dc:language.iso

eng

Handle dc:identifier.uri

http://hdl.handle.net/11427/13310

OAI identifier oai:identifier

oai:open.uct.ac.za:11427/13310