The following research projects were performed under my supervision from July 2002 to December 2003:
Real-Time South African Signed Language Recognition Using Hidden
Student: Brett Moses (Honours project)
This project formed part of our research programme in real-time Sign Language translation, and dealt with recognition of (isolated) dynamic hand gestures. The student built models for recognising these using HTK (a hidden Markov model toolkit originally written for use in speech recognition) and performed experiments investigating the effect of various parameters (model topology, model size, number of Gaussian mixture components) on recognition accuracy.
Recognition of Facial Expressions Using Support Vector
Student: Martin Schulze (diploma project)
This project formed part of our research programme in real-time Sign Language translation. The student investigated the use of SVMs for the recognition of facial expressions that are meaningful in sign language. The project resulted in a paper published at PRASA.
Automatic speech recognition using support vector machines.
Student: Michal Kokoszczynski (diploma project)
This was an investigation into SVM-based techniques for speech recognition. Performance of SVM and standard HMM algorithms were compared.
Pattern recognition for Brain-Computer Interface.
Student: Yusuf Said (Honours project)
An exciting area of research is the creation of computer interfaces that can read signals directly from the electrical activity of the brain. These can be applied for example to help the disabled or as additional interfaces for pilots and drivers. Several electroencephalography (EEG) data sets have been made available as part of a competition. See the following website for details: http://ida.first.fraunhofer.de/~blanker/competition/
This project involved implementing and performing experiments with different classifiers for these data sets and comparing to published results from the competition. In particular, classifiers using both k nearest neighbours and Gaussian mixture models were investigated.
Machine learning for the game of Bao.
Student: Edwill Nel (Honours project)
The game of Bao is a popular traditional board game in East Africa, especially Zanzibar. It belongs to the Mancala family of games, in which players take turns to "sow" seeds around the board, aiming to capture the opponent's seeds. The aim of this project was to write a computer program to play Bao as well as possible. In addition to implementing a convenient user interface for playing and experimenting with the game, a standard search algorithm was implemented to search the game tree to a fixed depth. This allowed detailed experiments to compare various evaluation functions. Reinforcement Learning techniques were then implemented for improving the evaluation function.
Speech recognition for toll bypass bypass.
Student: Warren Davidse (Honours project)
Toll bypass bypass is a fancy name for the idea of sending speech over the internet by means of passing it through an ASR front end on the sending side, transmitting it as text, and passing it through a speech synthesiser on the receiving side. Of course the main problem with this idea is that the ASR performance severely limits the performance one can expect from the system. This project (along with the synthesis project below) involved implementation of a chat tool using the Nuance speech recognition and synthesis packages and evaluating the performance of the Nuance speech recogniser for such an application, both when using unconstrained large vocabulary speech recognition and when artificially restricting the grammar available to the speakers. The transmitting part of the system was also ported from IPv4 to IPv6.
Speech synthesis for toll bypass bypass.
Student: Elroy Julius (Honours project)
This project (along with the recognition project above) involved implementing a chat tool using the Nuance speech recognition and synthesis packages, followed by extending it to enable multicasting and using the Nuance text to speech tools to synthesise speech in different voices for different users. The receiving part of the system was also ported from IPv4 to IPv6 and the performance of the system on the network was evaluated.
Vision-Based Static Hand Gesture Recognition Using Support
Student: Suneshan Naidoo (Master's project, not yet completed)
This project forms part of our research programme in real-time Sign Language translation. The aim is to recognise static hand gestures from video data. We are investigating the use of SVMs for this problem.
Realtime Virtual Signing Avatars.
Student: Matthew Stallebrass (Master's project, not yet completed)
This project forms part of our research programme in real-time Sign Language translation. The aim is to develop a system that will take a gesture markup language as input and generate video output of a signing avatar. We are investigating the use of hierarchical control systems for producing natural output.
Articulated structure from motion.
Student: Carl Scheffler (Master's project, not yet completed)
This project forms part of our research programme in real-time Sign Language translation. We are investigating solutions to the problem of reconstructing the 3-dimensional shape and motion of articulated objects from video sequences. Solutions to this problem exist for the case of rigid objects, but this project has produced exciting novel results (which have been published at PRASA) for the previously unsolved case of articulated objects. The project has immediate applications to vocabulary acquisition for sign language.