Refining Forest Mapping Drones for Swift Movement and Reliable Precision via Deep Learning

Loading...
Thumbnail Image

URL

Journal Title

Journal ISSN

Volume Title

School of Electrical Engineering | Master's thesis

Date

2024-09-30

Department

Major/Subject

Autonomous Systems

Mcode

Degree programme

Master's Programme in ICT Innovation

Language

en

Pages

95

Series

Abstract

As the demand for autonomous mapping flights in forest environments grows, the need for efficient and reliable quadrotor technologies becomes clear. This thesis, conducted at the Finnish Geospatial Research Institute (FGI), explores deep learning methods to enhance forest mapping drones, building on traditional Visual Inertial Odometry approaches. The aim of this research is to improve the safety and precision of drones navigating through forests, with a focus on avoiding collisions with small branches and enhancing the agility of the drone. The methodology involved a comprehensive literature review to identify recent advancements in autonomously flying drones within forested landscapes, focusing particularly on deep learning-based navigation strategies. Two of these models were critically evaluated and compared to select the most suitable approaches for implementation. RotorS, Flightmare, and Aerial Gym simulators were employed for training and fine-tuning deep learning models to optimize performance in the target environment. Real-world experiments were then conducted in forest environments using custom drone hardware specifically designed for this thesis to evaluate the robustness and applicability of the system. These experiments were performed in three different environment, and the performance of the system was evaluated based on mission success rates. The final conclusion of the thesis was that the tests yielded excellent results in simulation and sub-optimal results in real forest conditions, indicating room for improvement. Four experiments were conducted, each comprising four flights: one in an open space environment, one in a lightly vegetated forest, and two in a densely cluttered forest. The first three experiments yielded successful results, with the drone consistently reaching the target destination. This demonstrated a very robust response of the system in all three different environments. However, the final experiment was classified as an outlier due to odometry data issues, which caused the drone to collide with most obstacles. While the majority of real-world forest tests were successful, the goal of avoiding all small branches was only partially achieved. The maximum test distances were 50 meters in real forests and 30 meters in the simulator.

Description

Supervisor

Zhou, Quan

Thesis advisor

Honkavaara, Eija
Karjalainen, Väinö

Keywords

autonomous drones, deep learning-based navigation methods, under-canopy forest navigation, vision-based navigation, quadrotor hardware design and development, real-time embedded software

Other note

Citation