Improving Localization of a Multicopter by External Tracking

Kurzfassung

Collaborative teams of heterogeneous robots will be a key technology for future planetary exploration missions. Therefore, a high degree of local autonomy is required. In this work, a concrete use case involving a Rover System (LRU) and a Micro Aerial Vehicle (ARDEA) is presented. It is assumed that during an exploration flight of the flying robot, an interesting spot is detected which is supposed to be further investigated by scientific tools of the ground robot. For that purpose, the relative orientation and translation between the two systems need to be determined with a high accuracy by the use of on-board sensor readings. The task of pose estimation can be formulated as a modified and temporal version of a Perspective-n-Point (PnP) problem. While in a classic PnP formulation, the pose can be estimated from known 3D-2D correspondences in one image frame, the proposed approach formulates the problem in terms of 3D-2D correspondences belonging to a flight trajectory over time. An essential part of the overall problem is given by detecting and tracking ARDEA in a sequence of images to retrieve 2D observations. To address this issue, two different approaches are proposed. One uses a combination of Background Subtraction and Correlation Filter based tracking (CFTs). The other approach is based on the RetinaNet detector which is trained on the task of detecting ARDEA in an image. Both approaches are evaluated on experimental data to assess and compare their performances. For the task of pose estimation, an existing approach aware of 2D uncertainties in the image plane is extended to also incorporate 3D uncertainties originating from VO measurements. The pose of ARDEA in the camera frame of the LRU is computed by minimizing weighted residuals in a reduced observation space which is spanned by tangent vectors to the unit sphere in the camera frame. As a reliability measure, covariance estimates of the resulting pose parameteres are derived. The proposed approach for fully uncertainty-aware pose estimation is validated in simulation and in a real-world experiment. Results indicate the large potential of the proposed approach by significantly increasing the accuracy of the estimated poses and by yielding meaningful covariance estimates. This thesis will introduce a framework for estimating the pose of ARDEA with respect to LRU based on image streams of cameras mounted on LRU and Visual Odometry readings from ARDEA.