Conventional face recognition research places much emphasis on still images and have shown to achieve a good level of success in mostly restrictive conditions such as near-frontal poses, controlled lighting and rigid face motions. However, recent significant interest in video-based face recognition has earmarked it as a key emerging topic. This research aims to present new approaches for the recognition of faces in continuous video stream. These approaches will utilise visual dynamics and spatio-temporal information for the tracking and recognition of human faces in video sequences. As video sequences have the luxury of having both spatial and temporal information, directions of this research will focus on manifold representation (spatial), temporal information extraction (temporal), and unification under a probabilistic framework.

Although face detection is a fairly simple task to the human eye, performing the similar task on a machine can be challenging. In a group of people, partial occlusion and passing objects are often encountered . The FACEFIND project aims to design a multiple-person face detector, which is able to reliably "find" all faces in a group of people. In Part I, a novel RGB-Y-CbCr skin colour model was designed to extract multiple faces in colour images.

Many edge detection schemes suffer from the lack of image quality at the global level. Global properties are more vital in grayscale images due to loss of hue and texture. This work proposes a novel fuzzy-based Gaussian edge detector that uses both global and local image properties for grayscale images.

Face recognition algorihtms have much developed over the years. In the light of present advances processing power, there is much interest in assimilating more features and classifiers together to improve the robustness and accuracy of algorithms. Here, a face recognition algorithm which incorporates a feature fusion method with a multi-tier classification scheme is proposed.

The task of detecting and classifying human motion is an important preliminary tool for many high-level applications. However, the lack of robust classification and proper motion cues is a common problem in many approaches. In this work, a novel human motion detection algorithm that uses a fuzzy rule-base classification scheme based on moving blob regions is proposed.