Effects of light variations on drone’s visual positioning

Abstract

Positioning systems and algorithms play a crucial role in drone applications. Although Global Positioning Systems (GPS) are the most widely used method for drone localization, they are not always reliable and accurate in some scenarios. A recent study explores the visual-based positioning method, using Convolutional Neural Networks (CNNs) to match geometric features for drone positioning. The authors use an orthophotomap obtained from an actual drone to evaluate their algorithm. This can reduce the gap between research and practical operation. However, the approach overlooks the impact of lighting variations on positioning performance, i.e., brightness and color temperature. To address this limitation, we propose a novel CNN architecture to handle lighting variations. Our method improves reliability, accuracy, and computational complexity under varying lighting conditions by incorporating several critical components into the network. Remarkably, our architecture has only 51.35% trainable parameters and 83.97% floating point operations (FLOPs) of the existing one. Still, we can exceed it by 3.73% while not considering light variations and average 2.36% while considering light variations. The experimental results, also derived from an orthophotomap obtained via an actual drone, demonstrate that our approach effectively mitigates the challenges induced by lighting changes, ensuring reliable and accurate drone localization.