On the design of visual behaviors for autonomous systems

We describe a set of visual behaviors developed over the years, in the general context of robot navigation. The first set of behaviors solve three basic problems in vehicle navigation: obstacle avoidance, docking to a surface and moving along corridors or walls. Another set of visual behaviors were developed for the active stereo head Medusa, with the purpose of tracking objects with arbitrary shape or motion. Finally, we present an approach for egomotion estimation assuming arbitrary motion for the camera. All behaviors use purposive visual inputs, based on similar image measurements (image flow) but computed on different regions of the visual field. Goals are attained without camera calibration or environment reconstruction