Presented in this paper has been a method for controlling the gaze of an artificial, active vision system with respect to multi-featured, complex, non-synthetic objects. At present the scheme only uses low-level features, analogous to those in the primary visual cortex, and hence will be corrupted by non-target objects in the field of view. The proposal for future work to deal with this problem (and continue the analogy with the brain) is to include increasingly higher-levels of feature complexity that, with a similar feature count measure, will give added discrimination to the system.
There are many other research projects that endeavour to investigate and model the type of attentional, recognition behaviour discussed in this paper. Nordlund and Uhlin [ 5 ] and Tunley and Young [ 11 ], are concerned with optic flow and control to motion features and Weiman and Juday [ 12 ] with control to the foveal pixel count of binary shapes, all for tracking purposes. Hoad and Illingworth [ 3 ] and Pahlavan and Eklundh [ 6 ] describe methods for automatic control of stereo head-camera parameters, the former with open-loop fixation to single colour regions of interest in static scenes. Spratling and Cipolla [ 10 ] do describe a robotic, visual servoing system that has some goals in common with ours, but which operates by computing transformations between the error signal and specific motor actions, on outline contours in static scenes.
Static, open-loop fixation also features in many attention-based systems. Prime examples of such systems are Westelius [ 13 ] (symmetry and edge points), Milanese [ 4 ] (colour and local curvature), Culhane and Tsotsos [ 1 ] (intensity and edges), Rao [ 8 ] (image patches) and Grimson et al [ 2 ] (colour, edges and depth). The properties and benefits of the foveal representation have been recognised and discussed by , among others, Sandini and Dario [ 9 ], Tunley and Young [ 11 ], Weiman and Juday [ 12 ] and Westelius [ 13 ]. Although motion information is not explicitly incorporated into our model the importance of control to motion in animate systems for responding to things of potential significance is recognised. Extending the current system to include control to a maximum pixel count of motion features would be relatively simple.
In contrast to open-loop fixation systems, the work presented in this paper represents a move towards a more realistic system which is able to perform in real, dynamic environments. The system incorporates some of the features and ideas of the systems cited above, but extends and brings together the goals to develop an active, dynamic, versatile vision system that is able to respond to complex, real-world objects. The main feature described in this paper has been the foveal fixation measure. It provides a means by which the system is able to move to fixate a known object without having to pre-process the entire scene. The context of Perceptual Control Theory provides a strong foundation for a perception system, not only in terms of greatly simplified design of artificial architectures, but also in terms of simplified explanations for the motivation and operation of living systems.
Rupert J Young