Visual SLAM for Servoing for Appearance Based Topological Maps

In this paper we consider the problem of servoing wheeled vehicles in an indoor, initially unknown environment. The proposed approach relies on a hybrid (metric and topological) map built on visual cues. Navigation is planned using topological information to trace a path through viapoints that can be robustly performed by visual servoing control to accurately reach the goal positions. A map of an unknown environment is built as a collection of images taken by an exploratory robot. Images represent nodes in a navigation graph, in which edges represent feasible paths that the robot can execute by visual servoing. Metric and topological information are stored in a hybrid map, which can be shared and cooperatively updated in real time by groups of robots. The merit of the proposed approach is to combine the accuracy of visual servoing methods with a reliable representation of an unknown environment. As a result, the method provides purely visual-based solutions to two of the most relevant problems involved respectively in the field of localization, that is the kidnapped robot problem, and in the field of mapping, that is the closed path detection problem. Experimental results on a laboratory setup are reported, showing the practicality of the proposed approach.