Feature tracking in real world scenes (or how to track a cow)

In this paper we present a novel scheme for modelling and tracking complex real life objects. The scheme uses multiple models based on a variation of the point distribution model known as the vector distribution model. Inter and intra-class variation is separated using a variation on linear discriminant analysis known as 'Delta Analysis'. The tracking scheme is stochastic and is based on modelling model characteristics by a set of discrete probability distributions, which are updated in an iterative manner. Initialisation is performed using low level processing and a predictor is used to initialise characteristic probabilities on subsequent frames. This scheme has been applied to the task of tracking livestock in a realistic farmyard situation.