By Lynn Abbott, Subhoder Das and Narenda Ahuja

Active Vision Systems dynamically control attention in space, time and resolution by changing imaging parameters and processing algorithms.

Examples of Imaging Parameters are:

	Geometrical:	Optical:
	Position	Focus
	Orientation	Aperture
	Fixation	Zoom

Figure 1. illustrates the Active Vision System at the University of Illinois, Urbana-Champaign:

FEATURES:

2 cameras with zoom lenses for multiresolution stereo imaging;
5 degrees of freedom for position control (tilt, pan, independent vergence, translation);
6 degrees of freedom for lens control (zoom, focus, aperture for 2 cameras);
High accuracy and repeatability;
Constructed primarily from off-the shelf components (1987 vintage);
Stereo baseline: approx. 28 cm;
System is automatically controlled by Sun workstation.

OBJECTIVE:

Autonomous reconstruction of surfaces for real scenes having large size and large depth range by integrating depth information from stereo focus vergence under dynamically determined aperture zoom calibration

COMPUTATIONAL FORMULATION:

Two Steps Repeated:

· Visual Target Selection

· Surface Estimation in the Target Area to Accumulated

Global Surface Map from Local patches

Each Step Formulated as an Optimization Problem

DEMONSTRATION:

Estimation of surfaces in scenes using University of Illinois’ Active Vision System:

	Aim cameras at one part of the scene (video1, 983k )
	Fixate an object (video 2, 3.5M) , and acquire stereo images at reduced zoom.
	Analyze stereo images and extract local surface map
	Select a new target on the object using partial surface map (video 3, 1.23M), reorient the cameras to fixate the new target point (video4, 847k), extend the surface map of the object, and continue surface scan and reconstruction of this object (video5, 845k)
	Select a new object when scanning is completed (video6, 704k), and aim cameras at the new target (video7, 1.41M).
	Home on the target by iterating over the following steps , image registration to reduce horizontal disparity between image centers (video8, 1.68M). Stereo analysis of optically blurred images to derive a better estimate for the target point. Image plan adjustments to reduce blur until the target is fixated (video9, 2.38M ).
	Begin surface scan and reconstruction cycle for this object (video10, 1.12M).
	Continue to select and scan new objects, and update the composite surface map until the scene is completely scanned (video11, 3.2M).