Appendix 3 : Honours theses' abstracts

Below are the abstracts for al of the honours theses.

David A. Allen-Williams	Jamming on the World Wide Web
The Musical Instrument Digital Interface (MIDI) has been used since 1983 for the digital communication of musical data. Currently, specialised cables are used as the communication medium for most MIDI setups, but the increasing use of computers in music introduces the possibility for computer networks to replace them. Computer networks have many advantages over MIDI cables as a communication medium; they are fast, different types of data are easily mixed, long-distance communication is possible, and a large amount of network infrastructure already exists that can be used. This dissertation examines the feasibility of real-time MIDI communication using computer networks, and finds that the prominent issues lie in basic protocol design, rather than (rapidly improving) hardware capabilities. To demonstrate this feasibility, I have successfully developed interactive World Wide Web software designed to enable people to play music together, with potential applications in education and entertainment.
Michael Barrett-Lennard	Second Order Derivative Shapes: Calculation, Visualisation, and Application
Feature detection is important in computer vision. Many methods have been proposed that attempt to detect features in images automatically. The method proposed by Per-Erik Danielsson is investigated in this thesis. It uses second derivative information to extract local shape and orientation from a grey-scale image. The algorithm returns a parameter that classifies the types of shape including blobs, ridges, lines and saddles present in the image. One of the implementation issues addressed is the algorithm's many-to-one shape mapping, where different image features are mapped to the same shape parameter. The thesis explains how to determine correct results from the shape parameter. Another issue that is addresses is the choice of scale of the filters used. The proposed solution applies several filters at different sizes, then combines the responses using the "best scale for each pixel. The thesis also shows how to visualise the information returned by the algorithm. This is done by colouring an image according to the shape parameter. Canonical shapes, blobs, ridges, lines and saddles, being represented by red, green and blue. Finally, the method is applied to mammograph images to highlight important features of each image.
Andrew Berry	Solving Multiobjective Optimisation Problems Using the SEE Algorithm
Multiobjective optimisation problems (MOOPs) are problems that require the objective optimisation problems (MOOPs) are problems that require the simultaneous optimisation of number of objective functions which cannot be combined into a single function. These problems require the use of different techniques to standard optimisation problems. Genetic algorithms provide an optimisation technique based on the concept of natural selection. However, standard genetic algorithms can not solve MOOPs. The Simple Evolving Ecology (SEE) algorithm is an extension of the traditional genetic algorithm (GA) that can potentially solve MOOPs. There are two main differences between the SEE algorithm and standard GAs. One is the spatially distributed structure. Instead of a single population of solutions, there are multiple separated populations of solutions, each with a different fitness function. The second difference is the addition of the concept of a second fitness function to model the effect of ecology. The new ecological fitness function rewards solutions that are similar to the rest of the population of solutions. This dissertation applies the SEE algorithm to a number of MOOPs and compares the results with two other methods of solving MOOPs. The results show that the SEE algorithm is able to find multiple optimal points when they exist and can locate the Pareto-optimal set of solutions for a problem. When compared against the DAMO directed search dialogue algorithm, a manual method for solving MOOPs, the SEE algorithm is able to produce comparable results. However, MOBES, a Matlab implementation of a different, GA-based multiobjective optimiser, is able to produce considerably better results than the SEE algorithm in less time. We compare the SEE algorithm and MOBES to determine the source of these differences.
Thomas Brown	Finger Tracking for the DigitalDesk
A trend in computing environments today is to move towards more 'natural' interaction, another is to make hardware invisible to the user, both these ideas converge into ubiquitous computing, and the DigitalDesk is an example of this idea. Wellner's DigitalDesk, uses a camera for input and a projector to display output, so the user needn't deal with mouse nor monitor, and both paper and electronic documents can interact together. I concentrated on the input device for the DigitalDesk, namely the user's fingertip, which is made to act like a mouse. tracking such an input device is common to a number of augmented reality environments and involves various areas of vision and motion analysis. However, previous attempts have focussed more so on the vision aspect of tracking general objects, than on using the information already known about the user's hand, which is the approach that I have taken. As well as looking at how tracking is best performed, I have considered methods of comparing trajectory based input devices, such as finger tracking. To do this the goal of tracking the user's fingertip as fast as possible in real-time was adopted, so the finger tracking system could be compared to other input devices, using models such as Fitts' Law. My results show that my tracking system fits Fitt's Law adequately and compares favourably against the performance of a mouse.
Rainer Buschenhofen	A Graphical User Interface to Process Simulation Routines
Modelling an industrial process using any third generation language is difficult. Model code written in C++ is no exception. At Kaiser Engineers, a C++ library has been written for modelling alumina refinery. This library can be used to write a program for modelling a chemical process. Writing a simulation at this level is very powerful, however any potential user of the simulation libraries must be a competent C++ programmer. Additionally, eliminating coding errors is notoriously difficult. Some of the difficulties can be avoided by careful design of the models, but by no means does this make the task of creating a simulation easy. A graphical user interface to the process simulation models can be written. The interface will be able to transform a graphical representation of a process into an executable simulator. This paper covers my work towards creating that interface.
Chao Miin Tyi	Obstacle Avoidance For A Puma 560 Robot Using Minty Obstacle Avoidance Algorithm
Described here is a solution to the six degrees of freedom, path planning problem for a Puma 560 robot. it begins by planning the path for the waist joint. This is done by configuring the arm in such way that the following conditions are met: 1. Any payload and the end-effector are likely to collide with the surrounding obstacles and 2. The entire arm can fit into the region called "collision free zone". In this way, we can rotate the waist joint to its final configuration first, then the entire path planning problem can be broken down to a 2-dimensional (2D) exploration of free space. Instead of using external forces such as potential field or numerical cells to help the robot navigate a collision free path, the robot knows of several paths to get to the goal. It decides which path it should choose based upon the presence of obstacles and the goal position. Once a path is chosen, it proceeds to the goal along that path. While it is moving, it uses its two equipped cameras and an infra sensor to check for any collision.
Michael Cheng	Stereoscopic Motion Tracking in Biomedical Visualisation
Neurosurgery requires a high degree of accuracy to avoid damage to critical brain tissues. The drawback of the traditional navigational technique called stereotaxis has been the need for the frame on the patient, which can interfere with surgery and causes some patient discomfort. A solution to this problem is the use of frameless stereotaxis. Instead of a frame, cameras track a "wand" or a specially marked surgical instrument while a computer calculates its relative position to the patient's skull. The purpose of this project is to experiment with stereoscopic vision to determine its effectiveness for precision motion tracking.
Michelle Cho	3D Reconstruction Using Delaunay Triangulation
3D reconstruction is a technique used to reconstruct objects from contour data. This can be applied in areas such as surgical planning and geological prospecting, where there is a need to visualise objects that are enclosed within opaque barriers. An approach that uses 3D Delaunay triangulation to produce an initial set of tetrahedra between the contours is presented. Heuristics are applied to these tetrahedra to produce reconstructed objects that form simply connected components between each pair of adjacent contours. In some cases, the reconstructed object may not contain all the original contours. This occurs when tetrahedra containing contour edges are removed because they are only connected to other tetrahedra by an edge or vertex (non-solid connection). To correct this situation, a ``hinging'' process is presented that joins the tetrahedra in non-solid connections to the largest connected component of the reconstructed object. The results show final reconstructed objects that expose the original contours and form simply connected components between adjacent pairs of contours.
Voon-Li Chung	Reinforcement Learning in a Small Mobile Robot
Reinforcement learning is a machine learning technique designed to teach an agent a task through the use of reward and punishment, in conjunction with trial and error based environmental exploration. The aim of this work is to explore the use of reinforcement learning in a small mobile robot to achieve stationary light-seeking behaviour. Associative reinforcement learning was implemented both in the actual robot and in a simulation environment, which predicted the likely reinforcement that would be received as a function of its inputs; this was achieved with the use of a backpropagation-based neural network as the primary information storage unit. The neural network was found to suffer from catastrophic forgetting, a problem that was eliminated through the replacement of the linear-sigmoid hidden layer with an ALCOVE-style (topologically-fixed radial basis function) neural network. Q-Learning, a form of reinforcement learning that incorporates history, using an ALCOVE-style neural network was also implemented. This was done to determine the reasons for the associative reinforcement learning algorithm's failure to improve its performance with practice. This suggests that the cause of associative reinforcement learning's failure was the lack of long-term memory.
Peter Dreisiger	A Statistical Measure of Text Similarity
Identifying semantic similarities using probabilistic language models has received very little attention in the literature to date. While some articles have investigated the ability of statistical language models to generate semantic clusters, these models are traditionally restricted to word prediction and correction applications. This thesis investigates the suitability of using a probabilistic language model to provide an indication of conceptual similarity. The system uses mutual information-based spreading activation to provide a quantifiable measure of the similarity between two passages of text. Although not an absolute measure, it may be used to identify and rank semantically related passages. Results show that the system is able to correctly group a set of parallel passages found in the Bible, as well as identify paraphrased and simple metaphoric passages. Experiments have also shown that the similarity rankings generated by the program reflect the human perception of similarity.
Novak Elliott	Quantification of Geological Features Stored Within a Geographic Information System (GIS)
When viewed from within a Geographic Information System (GIS), geologists have found that mineral deposits of economic significance share some common geometric properties with respect to surrounding geological features. If the geometry of these significant deposits could be described by a set of metrics, then new areas could be sought out that have similar metric values --- and hopefully equally significant mineral potential. This thesis discusses several techniques for parameterising properties of shape. In particular, moments, curvature, and fractals are investigated as to their ability to encapsulate geologically significant shape characteristics. A small portion of the Kalgoorlie goldfields of Western Australia is used as a test-set. Results show that the moment-based metrics are an effective tool for quantifying a wide range of geological features.
Sebastian Glass	RUSTIC: Real-time Unconstrained Synchronous Tool for Internet Conferencing
An electronic whiteboard is a multiple-user conferencing tool that connects remote computers with a single visual display. All participants of a conference or informal meeting may place information and images on the whiteboard to highlight their argument or presentation. This dissertation proves the feasibility of the development of a Web-based, platform-independent electronic whiteboard tool. The whiteboard implements a real-time, unconstrained, and synchronous environment for Internet conferencing. The Java programming language is shown to be suitable for the implementation and the maintenance of the whiteboard given the requirements for visual display, communication and platform-independence. The whiteboard tool: RUSTIC, is developed using the Java programming language. RUSTIC shows that reasonable communication speeds, a high usability for the whiteboard interface, and useful maximum number of users are achievable. RUSTIC also maintains a globally consistent display for all user's interfaces against the three problems of inconsistency divergence, causality-violation, and intention-violation that are identified in this dissertation. This dissertation also investigates two established algorithms for consistency maintenance and implements the algorithm most ideologically suited to RUSTIC. This research culminates in the development of the electronic whiteboard tool: RUSTIC, showing the feasibility of developing a real-time, platform-independent, Web-based conferencing tool that can be used by anyone with Internet access and a Web-browser.
Simon Huband	Heuristics for Graph Colouring
The NP-hard graph colouring problem has many applications that involve some form of conflict resolution, such as exam timetabling and the channel assignment problem. Over the years, the literature on graph colouring heuristics has grown considerably large, with varying degrees of success indicated. To date there is no paper that adequately reports on a substantial number of graph colouring algorithms. This problem is addressed by this dissertation which considers many of the common, and some uncommon, algorithms. Some of the polynomial time algorithms considered include the classic fixed and dynamically ordered sequential algorithms, several sequential by colour algorithms, algorithms that consider pairs of vertices, and several others. In addition, some arbitrary running time heuristics are included, namely simulated annealing, tabu search, the exhaustive recursive last heuristic, and a graph partitioning algorithm. All algorithms are tested on random graphs with at least 1000 vertices, where the most promising results originate from algorithms which try to find independent sets of vertices that will leave as few edges as possible in the remaining, uncoloured portion of the graph.
Tristan Lewis	Techniques for Averaging Binary Images
The idea of computing the average of several binary images has not received much attention to date in the literature. Isolated articles have proposed techniques for use in binary image averaging, but none have performed a thorough analysis of a variety of techniques and attempted to determine the technique that produces the best average under a variety of circumstances. This thesis outlines the implementation and analysis of 38 binary image averaging techniques to determine which technique produces an average that is closest to a true image. The majority of the implemented techniques incorporate signed and unsigned distance transforms in one and two dimensions and make use of mean, median, and trimmed mean pixelwise combination to compute an average image. To obtain the final binary image, grey-scale images are thresholded at a value computed in a procedure analogous to histogram matching. As the size of the mask of values (used for the distance transform) increases, a closer approximation to the true Euclidean distance between two points in an image is obtained. Results show that the averaging procedure which incorporates a signed 2D distance transform using a 7x7 mesh of distance values and combines images via the pixelwise mean produces the best average image. Two areas of application of the binary image averaging techniques are pursued. These are computing average aircraft runway approach paths and averaging binary images of faces. The applications allude to the use of ridge tracing as an alternative method of obtaining a binary image from a grey-scale image.
Yuval Marom	Improvising Jazz with Markov Chains
Markov chains provide a simple way to model music. In this dissertation, I identify some important musical aspects of a jazz improvisation and present several techniques, based on Markov chains, that model these aspects to produce simulations based on h uman-generated input. Markov chains are stochastic processes that develop in time in a manner controlled by probabilistic laws and they deal with the interdependence of events. They are capable of capturing global as well as internal, local structures of a particular sequence of states. Because of these properties, Markov chains are suitable to model a jazz improvisation. Parallels exist between the activities undertaken by a human musician during a jazz improvisation, and the basic mechanisms of Markov chains. The results of this project are musically and statistically satisfying.
David McClure	Wanton Destruction -- Simulating Arbitrary Computer Models Shattering due to an Explosive Force.
Computer Generated Imagery (CGI) is increasingly being used to provide special effects for film and television where physical scale models were once used. However, physical models and pyrotechnics are currently still used in preference to CGI to create explosion special effects, due to their more realistic results. If realistic explosion effects could be created using CGI, both time and money would be saved. This thesis investigates the implementation of one aspect of simulating an exploding object: the shattering of the object. A computer model can be made to appear to shatter due to an explosive force by separating it into fragments and moving the fragments away from the centre of the explosion. Algorithms for performing these tasks on an arbitrary computer model that is defined by a mesh of polygons are investigated. The aim of the algorithms is to assist in producing a believable animation of a computer model shattering due to an explosive force. It is found that algorithms based on heuristics, and not physics, are able to produce convincing animations. In addition, such algorithms are not computationally expensive, and hence can be used in real time applications.
Jehann Mendis	A Computer Augmented Chessboard
Augmented reality is an exciting new stream of research. The possibility of integrating the real world and computer technology is becoming a popular area of discussion. Many techniques introduced by computer vision and image processing are now widely used in the development of such systems. The trend in augmented reality is to provide a "natural" interface to the machine so that the user may interact with the computer using well-developed motor skills. A computer vision based user interface to a chess playing program is presented. I have developed a system so that natural interaction with the chessboard can be recognised by the machine. In doing so, I investigated various techniques from the field of computer vision and explored the feasibility of applying such techniques in the context of augmented reality.
Matthew Miller	The Design and Implementation of a Component-based Real-time 3D Game.
Real-time 3D graphics applications, such as Virtual Reality systems, require fast computation and benefit from hardware acceleration. With recent advances in 3D rendering technology for desktop computers, there is now scope to use 3D graphics in applications for the World Wide Web. Deploying a 3D application on the Internet introduces the issue of how to access hardware acceleration on heterogeneous computer architectures. This thesis describes the use of an application program interface to provide rendering acceleration in a 3D computer game for the Web. The application is implemented as an ActiveX control to enable the game to be embedded directly into an HTML page.
Bruce Murphy	Geometric Methods in Spatial Data Mining
Spatial data mining encompasses many different processes which together provide a system for automated interpretation of data. It is the nature of data mining to deal with large sets of data and spatial data mining is no exception. This paper focuses on the problem of efficiently extracting the most significant elements from a large set of spatial data. Significance in a spatial set is restricted to those objects closest to the point of interest. The process of filtering the data set with cheap proximity approximation is examined and applied to a number of data sets. A system generating realistic pseudo-data is developed and used throughout. The extraction process is further generalized to include both spatial and non-spatial attributes.
Robert Nielsen	Evolving Artificial Life Forms
Evolution has been a powerful force in the development of life on Earth. Simulating this process on a computer can help us to understand it better. This project involves the creation of a computer based environment for virtual organisms to live and breed, governed by evolution and natural selection. The implementation, based on an object oriented model, achieves this to some degree but there are many problems associated with the task.
Sherrie Yap	Image Normalisation of Retinal Images
Information extracted from images taken at different times can be used for automated analysis. Before image registration, problems associated with the differences in brightness and contrast need to be solved. The thesis discusses mainly the histogram matching method. By performing histogram matching on the luminance of the image, we can yield normalised colour retinal images so that they can be used for bio-medical purposes such as the diagnosis of glaucoma. Past research has indicated that histogram matching works well for grey-level images but performing histogram matching methods on colour images is more complex.