This paper presents a mechanism for analysing the deformable shape of an object as it moves across the visual field. An object’s outline is detected using active contour models, and is then re-represented as shape, location and rotation invariant axis crossover vectors. These vectors are used as input for a feedforward backpropagation neural network, which provides a confidence value determining how ‘human’ the network considers the given shape to be. The network was trained using simulated human shapes as well as simulated non-human shapes, including dogs, horses and inanimate objects. The network was then tested on unseen objects of these classes, as well as on an unseen object class. Analysis of the network’s confidence values for a given animated object identifies small, individual variations between different objects of the same class, and large variations between object classes. Confidence values for a given object are periodic and parallel the paces being taken by the object.
|Title of host publication||In: Procs of the 6th Int Conf on Engineering Applications of Neural Networks (EANN'2000)|
|Publication status||Published - 2000|