A neural networks based visual tracking system

A visual tracking system based on a neural architecture is presented. The approach to target identification is non-conventional in that it relies on an architecture composed of standard neural networks (multilayer perceptrons), and exploits the information contained in simple features extracted from the images, using a small number of operations. Therefore the tracking functions are learned by examples, rather than implemented directly. The training set is composed of various geometrical shapes, in various sizes, with and without a background, pre-processed to yield the data vectors. The system exploits a two-level neural networks hierarchy with a number of parallel networks and an “arbiter”. The selected hardware implementation is based on a cartesian arm and a Motorola VMEexec workstation, that hosts the system but does not take part in the actual computation. This allows a true real-time operation