Control in Robotics and Automation

5 PROPOSED APPROACH

95

of the minimization algorithm, gains for energy function, number of control points). When

the key factors have been identified, future work will focus on improving the critical elements

of the system.

4 PREVIOUS WORK

This work draws on two streams of research in the computer vision and robotics communi-

ties. We have combined techniques developed by the visual servoing community with contour

extraction techniques developed in the graphics and computer vision communities.

Several research efforts have focused on using vision information in the dynamic feedback

loop [7, 10, 13, 14, 19, 24, 27]. Weiss

et al.

[24] have proposed a model reference adaptive

control scheme for robotic visual servoing. In this work, servoing is performed with the goal

of reducing the error between the desired image attributes (center of mass, first or second

moment of the image) and the current image attributes. The verification of the proposed

algorithms has been limited to simulation. Allen

et al.

[1] have proposed an approach that

uses image-differencing techniques in order to track and grab a moving object. Dickmanns

[11, 12] has presented methods (Kalman filters) for the integration of vision information in

the feedback loop of various mechanical systems such as satellites and automobiles. Koivo

and Houshangi [17] have proposed an adaptive scheme for visually servoing a manipulator

based on the information obtained by a static sensor. Feddema and Lee [13] have proposed

a MIMO adaptive controller for hand-eye visual tracking. Their work has been used as the

basis for our approach. Several other researchers [3, 18] have proposed strategies for

vision-based exploration. Finally, Ghosh [14] has addressed several vision-based robotic

issues with the aid of new "Realization Theory" for perspective systems.

The concept of active deformable models, also called "snakes," was first introduced to the

field of computer vision by Kass

et al.

[16]. Snakes have been used in a number of

applications including image-based tracking of rigid and nonrigid objects. Using snakes

requires a minimization process of an energy function. Several techniques have been used to

solve this problem, including variational calculus [16], dynamic programming [2], and

greedy methods using heuristics [28, 29]. The latter method has the advantage of being fast

as well as numerically stable. Our method uses a greedy method similar to that used by

Williams and Shah [29] and Yoshimi and Allen [-28].

Other researchers have also combined elements of visual servoing and active deformable

model techniques to approach different problems than the one presented in this chapter.

Blake, Curwen, and Zisserman [4] have presented a different algorithm for contour

estimation and used it in a system that tracks a contour in an image (it does not include a

robotic component). Yoshimi and Allen [28] have used a greedy, iterative minimization

algorithm to track a robotic finger with a static camera and detect contact between the finger

and a stationary object. Finally, Colombo

et al.

[9] published a description of a system that

uses a spline contour model to plan and execute a movement that positions an eye-in-hand

robot so as to bring a known object into a canonical orientation relative to the camera. They

report initial simulation results.

5 PROPOSED APPROACH

We describe a system that tracks a moving, deformable object in the work space of a robotic

arm with an eye-in-hand camera. For these experiments, we have used a figure-ground

approach to object detection and identification. The figure-ground methodology allows