LAMI-Mantra Day 		 LABORATORIE DE MICRO-INFORMATIQUE - EPFL  December 5th, 1997 

A quick, moving camera
for eye movement studies

Olivier Carmona 		  
No: 21 
Room:INF-113 
 
  

Statement of the problem

Mobility is a major feature for a robot evolving in a partially unknown environment. In order to interact efficiently with it, the robot needs to build a mental representation of its environment. In primates, the solution is to move the eyes with respect to the body to build a stable and coherent representation of the world, despite egomotion and spatially variant sampling.

I am taking inspiration from the biological mechanisms subtending the eye movements : from optics and mechanics up to anatomic architectures and psychophysics models, to implement robust and adapted algorithms permiting to update an internal representation of the world for a navigational task. In first approximation, I restrict myself to monocular vision.

This approach is closely related with the active vision paradigm. This concept, developed since the second half of the 1980's, can be defined as the voluntary control of perception and information processing parameters to carry out a specific action. Advantages of this approach are structure information from controlled motion, focused attention, enabling an efficient processing, and prediction. But it leads to voluminous and power consuming solutions not adapted to autonomous mobile robotics. Hence, we hope that the study of biological systems can lead us to efficient solutions.

  

Experimental Setup

We have built our own setup : the EPFL Vision Sphere (cf. figure 1) because no quick, miniature pan/tilt platform are commercially available. Here are its characteristics :
 
Video Signal Composite NTSC
Pixels 542 (H) x492 (V)
Field of View 110 deg (H) x 83 deg (V)
Excursion +/-40 deg
Speed 500 deg/s
Acceleration 40.000 deg/s^2
Dimensions 60 x 170 x 170 mm^3
Power 12 V DC, 0.5 A up to 1,2 A

Having such a wide field (cf. figure 2) permits to extract moving features and potentials features points in a bigger area. Moreover, the high acceleration ensures a short time to attain any points of interests in the scene.

Ongoing Experiments

We are mimicking human saccades by defining gaze targeting strategy. In order to complete this task, we are searching features at several image scales, proceeding in a coarse to fine strategy. We are also maintaining a ``craniocentric'' map based on the anatomy and functions of the posterior parietal cortex. Eventually, we are using the inherent, small motion of a mobile platform to obtain a crisp image and even better than the static one, a role ascribed to small drifts of the human eye. This processing also permits to extract motion information from the image.

In the future, we will embark this pan/tilt actuator on the new K-Team robot, the Koala, and we will cooperate with biologists to test theoretical models of saccades.

This research is supported by the CSEM Bioinspiration Group.


Olivier Carmona is an engineer from the School of Engineering at Marseilles, France. He received his Master Degree in Artificial Intelligence from University of Toulouse, France. He is now preparing his PhD at LAMI - EPFL. His research interests include Eye movements, Robotic Vision, Artificial Neural Networks, Autonomous Robotics, Machine Learning.

http://diwww.epfl.ch/lami/team/carmona
E-mail: Olivier.Carmona@di.epfl.ch Tel: ++41 21 693 39 08