Visual contact estimation

A fundamental capability of any navigation system is the perception of potential contact with surfaces in the environment. The efficiency and robustness of natural vision has motivated the development of biologically-inspired approaches to achieve this. Biological studies have highlighted the importance of visual motion (as perceived via optical flow) in the guidance of animal action. However, the use of optical flow in robot navigation systems remains problematic, impeded by measurement noise, environmental assumptions, and real-time constraints. This thesis proposes new biologically-inspired visual cues and algorithms for robust visual control and contact estimation from optical flow. We consider this primarily in the context of robot navigation and control. We present a robust strategy for docking a mobile robot with near-frontal surfaces using optical flow divergence. Results show improved robustness during egomotion, allowing closer than previously reported stopping distances. We present a strategy for performing controlled approaches towards surfaces of arbitrary orientation, providing the first unified control law for landing and docking. Velocity and heading control is achieved using only the maximum flow divergence on the view sphere. We present an insect-inspired structure-from-motion scheme using spherical optical flow from a hemispherical fish-eye sensor, providing the first demonstration of real-time depth map recovery from dense optical flow estimation. In dynamic environments, we investigate the use of optical flow to predict the time and location of impact of incoming objects. We consider this in the context of a stationary camera, as well as for on-road driver hazard perception assistance. We conclude that robustness in flow-based control schemes can be improved if system dynamics are handled in the image domain. This can be achieved by prioritising visual cues conveying a relationship between self-motion and scene structure over explicit structure-from-motion recovery in the control loop. Results suggest a wide-angle spherical projection model is well-suited for visual contact estimation from optical flow