Rajeev Sharma's Research in Active Vision for Optimizing Visual Feedback

Active Vision for Optimizing Visual Feedback

Formalizing effect of camera position on control. An active camera can greatly enhance the scope of a robot control task. Incorporating active vision in the robot control loop requires an understanding of the effect of the camera position on control. An important issue is how the position of the camera effects its ability to observe the motion of an object (possibly a robot manipulator). We formalize a quantitative measure called motion perceptibility, which relates the magnitude of the rate of change in an object's position to the magnitude of the rate of change in the image of that object. Since the rate of change of the image directly effects the visual control, this measure provides a formal basis for optimizing the relative camera/robot position over time. Further, the motion perceptibility can be combined with the traditional robot manipulability to yield a composite measure which can then be used for formally defining various active vision based planning and control tasks. We are interested in exploring this and other formalisms for active vision.

Decoupling active vision and control by learning invariance. One key issue in active vision is establishing and maintaining calibration of a large set of imaging parameters. Calibration errors can nullify the advantage of an active camera. Another issue is defining the goal of a vision task from the changing perspective of an active camera. This motivates us to explore an approach for learning a spatial representation that is invariant to changes in the camera configuration. Further, by making the learning self-organized, one can eliminate explicit calibration and the dependence on idealized imaging and kinematic models. We have developed a novel neural network called SOIM (self organized invertible map) for learning an invariant mapping for 3D points. This mapping is used to define an active vision based robot control architecture. A salient feature of this architecture is that the active camera control can be completely decoupled from that of the robot control. Thus active vision strategies can be independently developed to aid a robot task such as assembly in a cluttered environment. We have conducted extensive experiments and simulations to supplement the theoretical analysis for establishing the feasibility of this approach. We have also shown how the invariant mapping can be applied to other problems, such as, saccade control, motion detection, and stereo/motion correspondence. We are interested in further exploring both the theory of such invariance learning and its robotic applications.

Related Publications:

"Motion perceptibility and its application to active vision-based visual servoing".
R. Sharma and S. Hutchinson.
IEEE Transactions on Robotics and Automation, 13(4):607-617, August 1997.
"Execution of saccades for active vision using a neurocontroller."
N. Srinivasa and R. Sharma.
IEEE Control Systems, 17(2):18--29, April 1997.
"SOIM: A self-organizing invertible map with applications in active vision."
N. Srinivasa and R. Sharma.
IEEE Transactions on Neural Networks, 8(3):758--773, May 1997.
"A framework for active vision based robot control using neural networks."
R. Sharma and N. Srinivasa.
Robotica, special issue on Intelligent Robotic Assembly, 1997 (to appear).
"Efficient learning of VAM-based representation of 3D targets and its active vision applications."
N. Srinivasa and R. Sharma.
Neural Networks (to appear).
"The role of exploratory movement in visual servoing without calibration".
H. Sutanto, R. Sharma, and V. K. Varma.
Robotics and Autonomous Systems (to appear).
"Early detection of independent motion from active control of normal image flow patterns."
R. Sharma and Y. Aloimonos.
IEEE Transactions on Systems, Man, and Cybernetics, 26(2):42--52, February, 1996.

Return to Rajeev Sharma's Home Page