Date of Award


Publication Type

Doctoral Thesis

Degree Name



Electrical and Computer Engineering

First Advisor

Jonathan Wu


Applied sciences, Face tracking, Particle filters, Varying illumination




Automatic human face tracking is the basis of robotic and active vision systems used for facial feature analysis, automatic surveillance, video conferencing, intelligent transportation, human-computer interaction and many other applications. Superior human face tracking will allow future safety surveillance systems which monitor drowsy drivers, or patients and elderly people at the risk of seizure or sudden falls and will perform with lower risk of failure in unexpected situations. This area has actively been researched in the current literature in an attempt to make automatic face trackers more stable in challenging real-world environments. To detect faces in video sequences, features like colour, texture, intensity, shape or motion is used. Among these feature colour has been the most popular, because of its insensitivity to orientation and size changes and fast process-ability. The challenge of colour-based face trackers, however, has been dealing with the instability of trackers in case of colour changes due to the drastic variation in environmental illumination. Probabilistic tracking and the employment of particle filters as powerful Bayesian stochastic estimators, on the other hand, is increasing in the visual tracking field thanks to their ability to handle multi-modal distributions in cluttered scenes. Traditional particle filters utilize transition prior as importance sampling function, but this can result in poor posterior sampling. The objective of this research is to investigate and propose stable face tracker capable of dealing with challenges like rapid and random motion of head, scale changes when people are moving closer or further from the camera, motion of multiple people with close skin tones in the vicinity of the model person, presence of clutter and occlusion of face. The main focus has been on investigating an efficient method to address the sensitivity of the colour-based trackers in case of gradual or drastic illumination variations. The particle filter is used to overcome the instability of face trackers due to nonlinear and random head motions. To increase the traditional particle filter's sampling efficiency an improved version of the particle filter is introduced that considers the latest measurements. This improved particle filter employs a new colour-based bottom-up approach that leads particles to generate an effective proposal distribution. The colour-based bottom-up approach is a classification technique for fast skin colour segmentation. This method is independent to distribution shape and does not require excessive memory storage or exhaustive prior training. Finally, to address the adaptability of the colour-based face tracker to illumination changes, an original likelihood model is proposed based of spatial rank information that considers both the illumination invariant colour ordering of a face's pixels in an image or video frame and the spatial interaction between them. The original contribution of this work lies in the unique mixture of existing and proposed components to improve colour-base recognition and tracking of faces in complex scenes, especially where drastic illumination changes occur. Experimental results of the final version of the proposed face tracker, which combines the methods developed, are provided in the last chapter of this manuscript.