Visual navigation of wheeled robots : compensating floor undulations

Optical flow based navigation systems depend on the planar navigation constraint, which reduces search to two parameters-translation velocity in the headed direction and the rotational velocity about vertical axis. However, this constraint often fails in practice when motion undulations caused by floor variations or wheel kinematics generate dominating vertical flows, resulting in widely erroneous depth maps. In this work, these floor undulations are modeled as rotations about an axis derivable from the robot kinematics. These are then dynamically calibrated and compensated for, resulting in more accurate depth maps. Optical flow is now computed using the so called generalized dynamic image model [S. Negahdaripour, 1998] which results in a much less noisy depth map. Using fuzzy inference to ameliorate the effects of noise added in the differentiation process, we show an implementation on a Pioneer-II mobile robot that navigates successfully in unknown cluttered static environments