Autonomous Visual Navigation of a Quadrotor VTOL in complex and dense environments

Date

2021

Authors

Stevens, Jean-Luc

Journal Title

Journal ISSN

Volume Title

Publisher

Abstract

This thesis presents a system design of a micro aerial vehicle platform, specifically a quadrotor, that is aimed at autonomous vision-based reactive obstacle avoidance in dense and complex environments. Most modern aerial system are incapable of autonomously navigating in environments with a high density of trees and bushes. The presented quadrotor design uses leading-edge technologies and inexpensive off-the-shelf components to build a system that presents a step forward in technologies aimed at overcoming the issues with dense and complex environments. Several major system requirements were met to make the design effective and safe. It had to be completely autonomous in standard operations and have a manual override function. It had to have its computational capability completely on-board along with vision processing ability. As such, all state estimation and visual guidance had to be performed on-board the vehicle, removing the need for remote connection which can easily fail in forest-like environments. The quadrotor had to be made from mostly off-the-shelf components to reduce cost and make it replicable. It also had to remain under 2kg to meet Australian commercial aerial vehicle regulations regarding licencing. In order to meet the system requirements, many design decisions were developed and altered as needed. The main body of the quadrotor platform was based on off-the-shelf hobby assemblies. A Pixhawk 2.1 was the flight controller used due to its open-source code and design which included all sensors needed for state estimation, has manual override for control, and control the motors. A leading-edge computational device called the NVIDIA Tegra TX2 was used for vision processing on the quadrotor. The NVIDIA Tegra TX2's embedded NVIDIA Graphics Processing Unit (GPU), is compact and consumes low amounts of power. It also is capable of estimating dense optical flow on the GPU at rates of 120Hz when using a camera that outputs grey-scale images at a resolution of 376x240. The vision processor is responsible for providing directional guidance to the on-board flight controller. A design decision during the project was to include a 3-axis gimbal to stabilise the camera. The quadrotor was shown to be able to hover and locally move both indoors and outdoors using the optical flow measurements. Optical flow measurements give a sense of velocity which can be integrated to get a position estimate, though it was susceptible to drift. The drift was compensated using a combination of recognisable targets and positioning systems such as GPS. The experimental data obtained during the project showed that the algorithms presented in this thesis are capable of performing reactive obstacle avoidance. The reactive obstacle avoidance experiments were performed in both simulation and in real world environments, including the dense forest-like environments. By fusing vehicle speed estimates with optical flow measurements, visible points in 3D space can have their distance estimated relative to the quadrotor. By projecting a 3D cylinder in the direction of travel onto the camera plane, the system can perform reactive obstacle avoidance by steering the cylinder (direction of travel) to a point with minimal interference. This system is intended to augment a point to point navigation system such that the quadrotor responds to fine obstacle that may have otherwise not been detected.

Description

Keywords

Citation

Source

Type

Thesis (MPhil)

Book Title

Entity type

Access Statement

License Rights

DOI

10.25911/1BKG-8967

Restricted until