Research Strengths

The development of the Internet during the later half of the last century, and its subsequent exponential growth, played a defining role in the information revolution and established the Information Age that we all now inhabit.  The adoption of the Internet as a primary source of information, coupled with the geometrical growth of computing power, as predicted by Moore’s Law, has seen the amount and variety of information accessible through the Internet also grow exponentially, this exponential growth continues today and shows no sign yet of abating. We now have access to more information than ever before via the Internet, which raises the question, how do we find the information we want in this vast sea of information? Search engines, such as the ubiquitous Google, provide one solution. However, many search engines return additional, and often unnecessary, information, such as advertisements.  Moreover, the search may return incomplete data. Generally, users are only interested in retrieving the specific information they are searching for, and ideally they want that information to be as complete as possible.  In addition, users would like to find the required information simply and quickly, it becomes time consuming when the user has to manually search through pages of search results to locate the required information, in effect conducting a manual meta-search themselves.  The solution to this problem is Information Extraction, which can be used to automate this meta-search.  Methods for doing this are termed wrappers.  Wrappers search through and extract information from one or many web pages, collating relevant information.  Wrappers need to deal with semi-structured information and the many ambiguities of the HTML language.  Wrappers were initially constructed by hand to extract information from specific types of web page.  However, this method soon becomes impractical, virtually every type of pages needs its own wrapper, and the wrappers need maintaining and updating should their target web pages change their layout.  The solution of course is to automate the wrapper method.  Essentially, automatic wrappers attempt to find structure within the target web page and extract information accordingly, as data records, for example.  Early automatic wrappers used HTML tags to determine structure.  However, the accuracy and precession of these methods are affected by ambiguities in the HTML language, furthermore if the information in the web page is not uniformly presented.  To overcome these limitations wrappers used additional visual cues, such as context, in font, color, style, size and relative object positions with the web pages, pictures, for example. This research focuses on developing a new Information Extraction tool by combining statistical and visual knowledge, whereby the limitations in the current tools are minimized, notably in developing an automatic wrapper for extracting multiple record sections and providing ease of implementation that can be easily extendable in the future.

Tracking of features is crucial in areas such as computer vision and image processing. With the increased availability and affordability of web cameras, new features can potentially be detectable and utilized in manners that were not previously possible with other types of sensors. Therefore, we describe a model in which these potential features can be tracked both in an indoors and outdoors environment.  This feature tracking model is then applied to autonomous mobile robots to facilitate the process of Simultaneous Localization and Mapping (SLAM) in real-time. The implementation of feature tracking is intended to be done with binocular vision (two cameras) due to cheaper and smaller components.
In order to obtain sufficiently satisfying performances, many factors have to be considered. A robust and reliable world coordinate system is needed in order to allow the robot to track and traverse around real-world objects located in a given environment. The feature tracking ability of the cameras must not only be computationally inexpensive, but must also possess a high degree of accuracy. In order to compromise between computational complexity and acceptable results, it should also be known which types of features are allowed to be tracked, and using which specific algorithmic methods.

Biometrics are used to identify people in a non-intrusive way and should be obtained in a timely and accurate fashion. A typical scenario for automated recognition of humans is when people have to queue up to access a facility. They traverse a passage to be identified where commonly a single camera is placed. This provides a fronto-normal view of a person. The physical space needed is small and the application is natural, practical and useful. Since the video stream provides face and gait images, these provide spatial and spatio-temporal features respectively. Face features which are short-range biometrics, can be naturally combined with gait, which is a medium-range biometric using pattern recognition techniques. This combination is a relatively new field of research and provides a more robust identification as well as reducing error rates in recognition. Video streams can also provide various other biometrics like eyes, nose, and static gait features like body height and width to help in the identification process. By itself, gait is a relatively new biometric. Obtaining features from a video stream of fronto-normal gait provides dynamic data which can be subject to Time Series Analysis. Tracking body features from this viewpoint can be challenging because of the looming effect and occlusions.

Face recognition is a relatively mature field. We will use existing techniques from this field to validate the effects of biometric classifier combination. Initial experiments have shown promising results. We have used nonlinear dynamical analysis on fronto-normal gait, to improve face recognition rates.