Exploration for Micro Aerial Vehicles

• Applicant: Professor Dr. Cyrill Stachniss
• Subject Area: Geodesy, Photogrammetry, Remote Sensing, Geoinformatics, Cartography
• Term: from 2014 to 2020
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 166047863

In unknown environments, it is impossible to specify a flight trajectory for mapping a 3D structure such as building beforehand if the rough geometry is known. Therefore, an aerial vehicle need to ability to decide on the fly where to fly and which part to the scene to observe from which viewpoint. Through an iterative process, the platform selects new exploration goals and approaches them in order to update the map based on the new sensor input. The goal of this procedure is to reduce the uncertainty in the map as well as in the pose estimate of the platform. Humans perform a large number of tasks well because we have accumulated a huge amount of background knowledge. This project aims at exploiting background knowledge in the context of autonomous micro aerial vehicles exploring the environment. We will investigate how information collected from previous exploration missions, obtained form existing but incomplete models, and extracted through a semantic analysis of the scene can to improve the performance of an exploring copter. In the project P8, we will develop novel exploration techniques for micro aerial vehicles. A key challenge in this context is finding appropriate approximations to perform the involved computations efficiently. This involves, for example, efficiently estimating the expected gain of new observations. The central aim of this project is to develop a new generation of information-driven exploration approaches that combine information-theoretic concepts such as mutual information with background knowledge. The main difference of this proposal to existing exploration robots is the exploitation of background knowledge. We will investigate different types of background knowledge that we plan to exploit to more effectively explore the scene. This includes data from previous missions, existing but incomplete or outdated models as well as semantic information. A further aspect that we will investigate is to actively navigate the to support change detection with respect to previously acquired models.

• DFG Programme: Research Units
• Subproject of: FOR 1505: Mapping on Demand