Aerodynamics Analysis Tools for Flying Robots: Automated Tuft Recognition using Deep Learning

Abstract

For the design, analysis and monitoring of flying systems such as helicopters or stratospheric unmanned aerial vehicles (UAVs), understanding the physical phenomena of the airflow (aerodynamics) plays a crucial role. One of the oldest and simplest tool to study aerodynamics experimentally is tuft, which is a small wire attached to the flying system. By observing the direction of these wires during flight, many aerodynamic phenomena can be revealed. In this work, we present a system to detect, identify and segment the tuft using computer vision techniques. The idea is to make the analysis process automatic rather than relying purely on human experts. Using two custom dataset from the DLR – the DLR helicopter and the DLR stratospheric UAVs – we demonstrate a viability of our technical solution for the given tasks.