Computer Vision and Computational Neuroscience Laboratory

active vision

	Active vision refers to a camera system which is actuated. This is an optional strategy for conventional computer vision, since the engineering problems associated with acutation are generally not worth the trouble. However, for space-variant systems, actuation is not optional: the existence of a dedicated high resolution (foveal) regions in the sensor is pointless without the ability to control the direction of view of the camera.
	All higher vertebrate vision systems use space-variant active vision (SVAV). At the present time, few computer visions have made this transition. One reason is that active vision requires solution of a much wider range of engineering and algorithmic problems than conventional space-invariant passive vision systems (SIPV's).
	For example, miniatured video cameras are desirable to exploit the inherent size and power savings provided by active vision.
	Continuing in this thread, it has historically been a difficult and largely unsolved problem to provide high speed, high accuracy camera actuation in a small, power efficient package, although one solution has been designed and constructed, the spherical pointing motor, which is capable of saccades at 1500 deg/sec with 10 arc minute accuracy and target acquisition and smooth tracking. The design of attentional algorithms, which are required to "point" the camera, is largely unsolved.
	Several other technical obstacles are the general difficulty of providing space-variant image sensors, and development of algorithms for early vision and classification that function in a space-variant environment.
	The transition from SIPV to SVAV occured with the rise of the higher vertebrates roughly 200 million years ago. It is expected that a similar transition will finally occur in computer vision within the next few decades.

search www search eslab.bu.edu