Face recognition is a fast growing field, with many different applications to its use in society. These include security access within buildings, protection of the home personal computer or in places such as shopping malls to keep an eye out for wanted criminals. The method of face recognition investigated in this thesis is the “Eigenface” method that was first investigated and proposed by Matthew Turk and Alex Pentland. This method involves selecting the most significant eigenfaces that are produced from the training set and to produce a face space using this data. New images that are acquired to test against are first projected onto the face space to calculate the weights of image and then to compare this against known face classes.
The Yale database was used in the basic recognition process. This consisted of fifteen different individuals with eleven different pictures, which shows variations in facial expressions as well as lighting conditions. It was found that the overall accuracy of the eigenface method with the Yale database was 75.1%. The largest error occurred from the differing lighting conditions contained in some of the images. Providing a larger training set of images for each individual instead of just the single facial image could have compensated for the error that occurred.
The face tracking method used in this thesis was a combination of the background subtraction method and skin tone recognition. This was done to reduce certain errors that could occur from using just one of the methods. For both recognition methods it was found that lighting conditions would affect the tracking to a certain degree. With the skin tone method it would often add objects in the background that are similar in colour to the skin. This became insignificant when combine with the background subtraction method as only new objects that appeared since the original background image would be shown.
Using the eigenface face recognition method and the tracking method, a simple real time recognition was created. This was tested to be quite accurate on just myself. The recognition process occurs at around 2 frames per second.
Overall the eigenface method was found to be a very efficient way of recognising people, with the combined tracking method being successful a majority of time. There is room for expansion on this thesis to improve accuracy and to create a very efficient recognition system for security in buildings.