Guiding and Localising in Real-time a Mobile Robot with a Monocular Camera in Non-flat Terrains

In this paper we present a real-time active motion strategy for a mobile robot navigating in a non-flat terrain and its 3D constrained motion model. The aim is to control the robot with measurements from only one camera that autonomously builds a visual feature map while at the same time optimises its localisation within this map. The technique chooses the most appropriate commands maximising the expected information gain between prior states and measurements, while performing 6DOF bearing-only SLAM at real-time. The constrained 3D motion model presented here is used to infer the position of the vehicle in order to evaluate the mutual information for all possible actions at the same time. To validate the approach, not only simulations over uneven terrain have been performed, but also experimental results are shown for the technique being tested with a synchro-drive mobile robot platform with a wide-angle camera.