An integrated approach to face recognition in dynamic scenes was presented. The recognition tasks to be performed by such a system are typically characterised by poor resolution and variable lighting. In contrast to most previously developed face recognition methods, the data sets consist of many images of relatively small groups of known people. Four recognition tasks were defined: face classification, face verification, known/unknown and full recognition. All but face classification require consideration of the class of unknown people. As a consequence, identities should be modelled in a generic face space rather than a face space which is specific to the set of known people.
Mixture models provide an effective way to model identities as class-conditional probability densities in face space. Model complexity adapts to the structure of the data and simplified models are easily obtained when data is lacking. Face data used to compute a face space model for face detection were also used to compute a linear face space model for recognition. The eigenface method of [ 10 ] can be viewed as a special case and was outperformed by simple mixture models. It was shown that modelling identities using such models is beneficial given an appropriate level of mixture complexity. This approach to recognition results in a system which can learn and update identity models independently of one another. Gaussian mixture colour models were also used to provide an efficient and effective focus of attention for use in face detection and tracking.
Shaogang Gong