Combining monocular and stereo vision in 6D-SLAM for the localization of a tracked wheel robot

摘要

Methods for the localization of tracked wheel robots in GPS-denied, unstructured environments, such as the ones encountered in Search and Rescue scenarios, meet several challenges. In most situations, one cannot rely on a planar ground, that would simplify localization and mapping. This paper addresses an online 6D-SLAM method for a tracked wheel robot, aiming at providing 6D pose estimates of the robot. While the robot pose is represented by a 3D position and a SO(3) orientation, the environment is mapped with natural landmarks in 3D space, autonomously collected using visual data from feature detectors. The observation model opportunistically employs features detected from either monocular and stereo vision. These features are represented using an inverse depth parametrization. The motion model uses odometry readings from motor encoders, together with orientation changes measured with an onboard IMU. A dimensional-bounded EKF (DBEKF) is introduced here, that keeps the dimension of the state bounded. A new landmark classifier using a Temporal Difference Learning methodology is used to identify undesired landmarks from the state. By forcing an upper bound to the number of landmarks in the EKF state, the computational complexity is reduced to up to a constant while not compromising its integrity. A real dataset from RAPOSA-NG, a tracked wheel robot developed for Search and Rescue missions, is used to experimentally validate the approach. This dataset encompasses a closed circuit, including stairs and non-planar ground segments.