We have described a method of using Point Distribution Models to elicit the behavioural trends of a group of animals within a suitable training set.
The standard PDM is used to describe the shape change of the outline of a flock of ducks, extracted via a simple segmentation scheme. The model is then extended to incorporate an additional entity, a robot, that affects how the flock moves. In order to describe this process of predator affecting prey, suitable parameters are chosen (robot position and flock velocity) for inclusion in the PDM.
A method has been described for balancing the influence of these extra parameters, and the results show that behavioural trends contained within the data are more accurately extracted. The applicability of the method need not be restricted to extending PDMs, such a technique may prove useful generally in applications based upon the use of principal component analysis in which various non-comparable parameters are considered simultaneously. We aim to further this work by formalising the method, in order that the desired scaling can be extracted mathematically.
Future work involves incorporating the positions of the individual birds within the flock into the model, as well as additional parameters (for example, distance to the boundary). A more complete extension of using the model to predict behaviour would be accomplished by incorporating an element of shape history.
N Sumpter