Using a multi-resolution method the Active Shape Models are able to locate accurately new examples of image structures, given a crude estimate of their position. In many cases the approximate position of an object is known in advance and can be used as the starting point for an Active Shape Model. However, in less constrained circumstances it can be difficult to find a suitable initial position. Various approaches have been tried, such as coarse exhaustive search, global optimisation (eg using a Genetic Algorithm  ), and cue-based methods . We will compare the different methods, examine how each can be made more efficient. Our intention is to devise a systematic approach to locating objects given little or no prior knowledge of their position.
It can be difficult to locate image structures when they are partially occluded, particularly when the shape of the structures can vary. We intend to increase the robustness of our techniques to occlusion by considering sub-parts of the model separately. Whereas the location of rigid objects can be accurately determined by a few unoccluded sub-parts, determining parts of a flexible object typically only applies loose constraints to the positions of other components. In the ASM each model point makes a contribution to the overall estimate of shape and pose. If we can calculate a reliable measure of the support in the image for each point or group of points, we can use this value to weight each point during search. The accuracy of location of partially occluded parts will be poor, but the occlusion will also reduce the evidence supporting the location of the part. The location of the whole model structure will be determined by the unoccluded areas and the statistical constraints the model places on the overall shape of the structure.
Where new search techniques are developed, careful attention will be given to assessing their efficiency, accuracy and robustness, where possible comparing them with existing methods.
To summarise, the intention is to investigate