Face & Behaviour Recognition
Face Recognition |
This is the part of the system the actually correlates the (pre-processed) input image with a database of faces and attempts to find a match. We have developed various algorithms over the years in an attempt to solve the face recognition problem using 2D images and 3D face models. These include algorithms based on neural networks, hidden Markov models, eignfaces, fisherfaces, kernel-based methods and support vector machines. Most of these involve the use of wavelets and Gabor wavelets for feature extraction. A complete automated face recognition system has now been developed encompassing three major software components - face detection, recognition and tracking. Though supported by a low cost camera, the system is able to automatically locate and recognise faces against a face database of several hundred subjects. Faces and other objects in the scene are also tracked and their trajectories recorded for behaviour analysis. The sample output from ourface recognition program can be viewed by clicking on the image to the right. The program is capable of recognising and tracking many faces at once in a real time video stream. |
Face Detection |
Techniques in face
detection range from looking for skin coloured objects in a scene, trying
to find features on a face within the image or searching for the general
contour of the head.
Indeed a combination of any or all of the above is possible. The reliability
and speed of the detection process is essential to the overall system
since the ability to recognise a face at later stages is very much dependent
on there being a face to recognise. Our
face detection system can be seen by clicking on the image above.
Another area of study is in developing applications to track movement and motion within a scene to provide additional data for analysis. Abody/motion tracking application can be seen by clicking on the image to the right. |
The 2005 International Face Verification Competition |
Our entry to the above face verification competition has won us one of the two top places in the competition. The error rates of our systems are far less than those entries from University College London, University of Surrey, Carnegie Mellon University, and German hi-tech companies. |