This thesis proposes methods for interpreting images of a known moving object by computer. The methods are intended to cope with image sequences for which an object may not be clearly visible, for whatever reason. The overall approach is highly predictive, involving the generation and evaluation of hypothetical states of an object. It is shown that a predictive approach is a feasible way of tracking a jointed object when its identity and actions are known. It is also shown that there are ways to focus the attention of a predictive procedure to cope with situations in which little is known about the position and configuration of an object; as may be the case for the first frame of a sequence. All the methods have been implemented as a computer program which serves to illustrate ideas throughout the thesis.
This paper is not available online