Dissertation

Deploying Deep Learning Models Using Embedded Computing on Autonomous Robot

Steele, Jonathan

2021

Overview

This paper is set out to investigate the optimization of sensor operation data cost and to deploy deep learning models on a robot for autonomy. The goal of this research was achieved through implementation of a regression based deep learning model which takes advantage of sensor fusion for collision avoidance (LiDAR+ STEREO+CSI). For these goals to be met, an autonomous robot, named “Phoenix One,” was built from recycled materials and purchased parts. Though several examples exist in literature for the use of autonomous systems that employ embedded computing or embedded systems, there is little literature available on how to effectively deploy deep learning algorithms on embedded computers in a step-by-step format. In our research, data was collected through a CSI camera on our Phoenix One robot. Images captured had a size of 224 × 224 pixels and were labelled manually for training our deep learning model ResNet18. I developed a deep learning model for collision avoidance and a deep learning regression model for lane following, using a Jetson Nano embedded computer. Its graphical processing unit gives the Jetson Nano an advantage over other embedded computers due to its added computational processing power and low cost. Through our research I proved that deep learning + regression (DL+R) can effectively be trained and deployed on autonomous vehicles for road following. In future work, I hope to achieve an added goal of Phoenix One autonomously driving between two buildings on campus by integrating GPS and waypoints.

• Publisher: ProQuest Dissertations & Theses
• Subject: Artificial intelligence / Cameras / Computer Engineering / Decision making / Engineering / Neural networks / Robotics / Sensors / Software / Unmanned aerial vehicles / Wireless networks
• ISBN: 9798516956621