Student Projects

For solving a number of business decision problems, it is necessary to attach values such as costs and prices to the frequent patterns and the association rules discovered by data mining. It will be more efficient if optimal patterns and rules can be found directly by associating cost or price to items during the mining process. This project is aimed at combining optimisation and mining techniques to find value based frequent patterns that are meaningful for decision making.

Classification in data mining has many applications. Associative classifiers based on frequent patterns are more accurate but not feasible to construct for many practical data sets when there are too many association rules at low support thresholds. Maximal frequent patterns and closed frequent patterns are far more concise and could form the basis of better classifiers. This project is aimed at mining maximal and closed patterns more efficiently using compressed frequent pattern trees and then using those patterns for classification.

Mining patterns from dynamic datasets (also known as mining from data streams) such as stock market data and web click streams is a focus of current research interest. The goal of this project is to mine frequent patterns in data streams using more compact data structures. For the same size and rate of data streams, more information can be stored in compact structures and the interval of mining extended resulting in more efficient stream mining.

The project combines the use of image processing and machine learning to detect abnormal behaviour in enclosed public areas such as underground passage ways. The project requires the development of models of normal behaviour which can then be used in conjunction with machine learning to determine whether illegal activities such as vandalism, loitering and fighting occur in the surveillance video. The project involves work in both object tracking and spatio-temporal pattern recognition.

This project is done in collaboration with DTI. The goal of the project is to be able to accurately count the passengers getting on/off trains using multiple cameras in order to optimise the train schedule. Each train door will be monitored by a single camera and the aim is to be able to determine at all times how many passengers are on the train. The project complexity comes from the fact that it must be able to handle crowds and be able to produce reliable results under varying lighting conditions.

In these days of heightened security awareness, both employers and employees have a major role to play in making the workplace secure and safe. Security and safety vulnerability in a work place plus the onset of threats determines the risk of these threats from occurring. Current risk assessment procedures perform on critical infrastructures are sometimes inadequate to deal with dynamic threat situations. This project aims to develop a computational simulation system which provides decision makers with the tools needed to access the state of security and preparedness to deal with security and safety related events. The system will interactively model dynamics of people or groups of people within the protected Virtual Space. Agent-based computational simulation techniques, popular in gaming technology will provide the key focus of this project on top of virtual environment modelling and simulation. Simulations created by the system can also be used for computer-assisted training purposes.

This project addresses the problem of pose and illumination variations affecting the performance of real world face recognition systems. First of all, we will explore the use of a shape-from-shading technique to extract the 3D structure of a face using multiple images taken of the person under varying illumination directions using an in-house developed mini light stage. This 3D face structure with the skinning model can then be used to generate further 2D representations of the face under varying pose and illumination variations. This analysis-by-synthesis technique will be used subsequently to address the problem of outdoor, unconstrained face recognition applications.

The problem of tracking human motion using video inputs is an active research area in computer vision. In addition to this, there is a parallel development in the computer graphics (CG) community to inject greater realism and efficiency in rendering dynamic 3D models (e.g. human kinematic motion, clothing, and etc.). From the perspective of video-based human motion tracking, it is becoming obvious that the utility of CG-generated dynamic models can be harnessed to improve its performance. For example, one can harness the body joints constraints whose information can be embedded into the CG human model to define the extent of the joints movement in the video. It is also interesting to determine how this model can be used to alleviate the problem of body self occlusions which is an unsolved problem in tracking. This project will investigate the use of CG human model to assist in the tracking of human motion taken from a single camera.

Investigate and implement a method of delivering the seismic tomography results from CO2 pressure vessel experiments to a desktop PC in the laboratory (or remote office) in realtime using HPC parallel architecture (i.e. SGI Altix). The solution will include a visualisation platform (can be existing software) for both Windows and Linux environments, and the processing of sensor (pressure, temperature, seismic, etc) data into suitable forms to be visualised.

Investigate and implement a webserver solution to host and serve 3D virtual models of rocks and minerals, with a local client option for the remote viewing of the collection via the Web. 3D models of rocks and minerals will be obtained via a desktop scanner at UWA and these will form the collection. Some optimisation and/or refinement of the models may be required before inclusion. Some of the 3D models will also be "printed" on the rapid prototype (3D printer) machine also at UWA.

Investigate the application of face recognition techniques to rock surface feature recognition and frame registration from video streams and, therefore, progress the capability for the auto generation of 3D textured geometric mesh of the relevant rock surface. This has application in the mining and geotech fields.

Write a C wrapper for the standard OpenGL32.dll such that stereo-enabled OpenGL-based visualisation software can export/write the left-right buffers as sequential movie frames to disk. Normal OpenGL calls will pass through the wrapper to the regular OpenGL32.dll. The wrapper is only to be invoked if placed in the directory of the relevant visualisation software or through some other non-intrusive selection procedure.

Investigate and implement a method of integrating the InterSense Wireless InertiaCube3 tracker system with a stereo (collaborative) visualisation environment, e.g. GDIS, VTK (including Mayavi), Drishti, or a Virtual Environment (e.g. EON, OpenMASK, OpenSceneGraph) to provide navigation and data interaction suitable to iVEC usage requirements.

Investigate and implement the integration of an open source haptic API (e.g. H3D.org) with existing open source scientific visualisation software (e.g. Drishti, VTK or ParaviewGeo), primarily for interpretation of geoscience (i.e. seismic) data sets. A PHANTOM® Premium 1.5 haptic device from SenseAble Technologies is available.

Streaming data is becoming quite common, but machine learning algorithms are only just being developed to handle such 'online' data streams. A major issue is not only clustering the incoming data to assign labels to groups of similar data points, but also detecting when the stream's data characteristics have fundamentally changed (indicating a change in the physical system that the data stream is measuring). Such data stream clustering is also becoming increasingly useful due to the data capture capabilities of modern mobile phones.