(Georgia Institute of Technology, 2022-01)
Harris, Caleb M.; Kim, Seulki; Payan, Alexia P.; Mavris, Dimitri N.
Urban Air Mobility and Advanced Air Mobility require the certification of novel electrified, vertical takeoff and landing, and autonomous aerial vehicles. These vehicles will operate at lower altitudes, in more dense environments, and with limited recovery abilities. Therefore, emergency landing scenarios must be considered broadly to understand the risks in some situations of flight failures. This work provides a preflight planning tool to assist these vehicles when emergency landing is required in crowded environments by fusing geographic data about the population, geometric data from lidar scans, and semantic data about land cover from aerial imagery. The Pix2Pix Conditional GAN is trained on Washington D.C. datasets to predict eight classifications at a 1m resolution. The information from this detection phase is transformed into a costmap, or riskmap, to use in planning the path to the safest landing locations. Multiple combinations of the cost layers are investigated in three test scenarios. The Rapidly Exploring Random Tree (RRT) algorithm efficiently searches for an alternative path that minimizes risk during emergency landing. The tool is demonstrated through three scenarios in the D.C. area. The objective is that the tool allows for the safe operation of UAM and AAM vehicles through crowded regions by providing confidence to the local population and federal regulators.