Face Recognition in Video Surveillance
Security in public spaces, such as squares, shopping centres, station and airport halls, has become an important issue over the last 5 years or so. Camera surveillance has been introduced in order to prevent threatening situations or criminal acts ranging from shoplifting and unwanted gatherings to riots and terrorist attacks. A recent development is to combine camera surveillance with automatic face recognition, with the purpose of spotting registered suspects before they can act. This is a typical watch-list application: a detected face is compared with faces of known suspects on a watch list.
Although commercial systems that combine camera surveillance with a watch list have been installed by security bureaus, this technology is far from mature and claims of success can be considered doubtful. The problem is that at present the error rates of automatic face recognition are still too high for this type of watch list application. There are a number of reasons why this is so:
- Subjects are non-cooperative and avoid the cameras.
- Poses are often non-frontal, which increases the error rate significantly.
- Illumination conditions are uncontrolled and usually bad.
- Image quality is often poor.
- Resolution is often low.
- Identification from a list is a hard problem, made harder by the fact that, in contrast to a standard identification task, a low false-reject rate is crucial because it determines the probability that a criminal slips through;
- Images contain multiple faces, some of which may be missed.
Nevertheless, an (semi-)automatic approach is the only way to deal with such a watch-list problem. It is unthinkable that a human can, instantly, compare camera views with the possibly more than a hundred images on a watch list.
A potentially successful approach in this situation is to reduce the error rates by exploiting the possibilities of the surveillance scenario. By combining person detection and tracking with face recognition a multitude of images of the same face become available for recognition. The research will focus on making use of this fact for improving the recognition performance to a level that is suitable for a serious watch-list application.
Team(for the moment)
- Raymond Veldhuis
- Luuk Spreeuwers
- Bas Boom
Last modified: 2008-03-28 (14:54) by Bas Boom