Abstract
The class of real interest (e.g. traffic incidents) is under-represented in our database. This is because the abnormalities are rare and difficult to collect in safety-critical applications. Conventional incident detection algorithms focus on identification and characterisation of a wide range of different incidents from historic data. In this paper, alternative approaches which estimate the probability density of incident-free data are proposed. The abnormality of an input vector is therefore identified by testing for novelty against the description of normality. Experimental results on both simulated and field data show that the techniques are capable of detecting incidents and can thus be used in dynamic traffic monitoring systems. (A)