Design and Implementation of an Odometry System for the ARDEA MAV Platform

Kurzfassung

The landing of the technical demonstration drone Ingenuity on Mars in February 2021 will mark the beginning of a new era of extraterrestrial exploration. Expeditions may now not only be conducted by rovers that are hindered by terrain traversability, but also by autonomous flying robots. For this robust self-localization and ego-motion estimation is key.Information-rich, yet passively sensing and lightweight cameras are ideal sensors. Dead reckoning the motion of a camera through its video stream is called visual odometry (VO). With the hexacopter Ardea, the DLR has a research platform at which new algorithm scan be explored. However, the VO currently deployed on Ardea only works when trackable features are not further away than about three meters. Therefore, this thesis gives an extensive overview of the building blocks of any VO method in order to reason about possible alternatives. Additionally the architecture of state-of-the-art methods is explained.Furthermore, comprehensive tests with benchmark datasets from different environments are carried out. Based on insights from those tests a fixed-lag smoothing odometry that uses sparse optical flow tracking for data association is proposed. Key features include the handling of arbitrary camera setups and the proper handling of pure rotational motion.To enable the operation in the monocular case, a new map initialization strategy is developed. Subsequently, new strategies for keyframe selection and feature point detection that drastically improve robustness are demonstrated and evaluated. The overall system performance is quantitatively measured on common benchmark datasets as well as on a dataset relevant for robotic exploration. In the later superior robustness over state-of-the-art methods becomes evident.To aid overall system robustness, novel and easy to interpret measures of uncertainty are explored. For this, the relative translation information matrix and the scale information matrix are introduced. They are then condensed into a scalar value via an algorithm inspired by Principal Component Analysis. Those scalar values can now be used to reason about the expected scale drift or detect the point at which a stereo system degrades to a monocular one, as the ratio between the distance of trackable landmarks and stereo baseline becomes too small.