K.U.Leuven > ESAT > PSI > Visics > Research > Topics > Item 2.5

Autonomous, visual guided mobile robots

L. Van Gool, T. Tuytelaars, L. D'haene

Autonomous robots still often use special markers. This is especially the case in cluttered scenes, where the walls are difficult to pick up from laser scanner data. In this project the group has developed algorithms for picking up novel kind of landmarks that are available in most scenes. Hence, No special markers are required, as these landmarks can be extracted from the scene itself. Moreover, the characterization of these landmarks is based on the joint invariance under viewpoint and illumination changes. As a consequence, the robot can learn paths from a series of images taken along the taught paths. Later, the robot can repeat these trajectories by comparing its current position against those from where the example images have been taken. The robot can find its position even when it is off course and under different illumination conditions, due to the invariance in the description of the natural landmarks. The robot will automatically adapt its position to match it against the most similar example image. Early in the project, also experiments were carried out with 3D scene reconstructions from pure video, i.e., without taking odometry into account. But the 2D approach came out to be more efficient and more robust for in-house scenes.