En säker landning av drönare via ett artificiellt neuralt nätverk

Abstract

Modern drones are often supplied with a built-in landing function. The function is usually primitive and rarely involves any advanced risk assessment. In this report, we propose a solution to transfer the human component of the risk assessment to the drone in the form of a neural network. It has previously been found advantageous to make valuable information more explicit before using it as an input in a neural network. Two different types of network models are compared in the work. One network only uses the drone’s camera frames as input variables, reducing its complexity. The second network performs more transformations on the drone’s camera frames to first extract valuable information from the images to make them explicit. The network then uses the explicit information in combination with the drone’s frames as input variables. In this report, a broader comparison is also presented about what impact on the network’s precision more explicit data has. The result of the work shows that networks that use explicit, valuable information have higher precision, but that the magnitude of the precision difference depends on the properties of the task. We also show that a neural network can, with good precision, perform a risk assessment of a drone’s landing.