Face Recognition
Master thesis
Permanent lenke
http://hdl.handle.net/11250/250093Utgivelsesdato
2007Metadata
Vis full innførselSamlinger
Sammendrag
Machine based face recognition has been a popular research area for several years, and has numerous applications. This technology has now reached a point where there already exists good algorithms for recognition for standardized still images - which have little variation in e.g. lighting, facial expression and pose. We are however in lack of good algorithms that are able to do recognition from live video. The low quality of most surveillance cameras, together with non-standardized imaging conditions, make face recognition from live videos quite a challenging task. This thesis represents a continuation of our project ``Face recognition''~cite{IOLMH}, performed in fall 2006. One of the goals is to finish the work on our previous project, and implement a functioning module for Trollhettas application TrollWatch. This module should work in real time, and is to be used for surveillance of a corridor at our university. The main focus will be on different techniques for face detection and normalization, as these represent crucial factors for recognition results. Detection of movement is the first important task. Further processing is only performed on frames where a person is included. Then we subtract the background and use appropriate colour models to extract skin regions. The largest remaining region is then found, and used as the basis of our two different face detection methods. Ellipse fitting tries to detect the face by fitting an ellipse to the remaining region. The template matching uses the same region to create a search space, which is scanned using a face template in order to find the face. Properties provided by the face detection methods are then used to normalize the detected face. The last step is recognition, which is performed using eigenfaces or fisherfaces. One of the goals of this project has been reached. We have finished the module for TrollWatch. The test results were however of varying quality. If the different variables are properly tuned, we are able to detect the face and normalize it in most cases. The quality of these faces is, however, often not good enough for the face recognition algorithms. This means that the recognition results could have been better.