A Framework for the Consistency Analysis of Relative Pose Sensors for Unmanned Aerial Vehicles (Uavs)

In autonomous multi-robot systems robot-to {robot/object} localization methods can be utilized to increase the robustness and to achieve a precise and robust localization of the individuals. This paper investigates on the performance of two promising systems: UVDAR, a vision-based mutual localization in the UV spectrum, which has shown to be effective in swarm formation and leader-following tasks, and PoET, which is a deep learning-based visual relative object pose estimator. To evaluate these methods, we collected datasets in a controlled indoor environment equipped with a motion capture system for precise ground truth measurements. Our evaluation considers two key aspects: the absolute error between measured and true relative poses, and the consistency of the provided measurement uncertainty estimates with the actual errors. We introduce a novel framework for evaluating the consistency of relative pose measurements. This framework supports various error definitions and leverages spline-based trajectory representations to generate smooth, $C^{2}$-continuous reference measurements. Both the UVDAR dataset and the evaluation framework are made publicly accessible to foster further research and development in this field.