Estimation of Bio-inspired Visual Cues using Onboard Camera Images for Landing a Nano-Quadcopter
Open Access
- Author:
- Huff, Wesley
- Graduate Program:
- Mechanical Engineering
- Degree:
- Master of Science
- Document Type:
- Master Thesis
- Date of Defense:
- June 24, 2022
- Committee Members:
- Daniel Haworth, Professor in Charge/Director of Graduate Studies
Bo Cheng, Thesis Advisor/Co-Advisor
Katie Fitzsimons, Committee Member - Keywords:
- Tau Theory
Inverted Landing
Micro Aerial Vehicle
Optical Flow
Horn Schunck
Reinforcement Learning
Visual Odometry - Abstract:
- Unmanned aerial vehicles have become increasingly prevalent in a multitude of industries across the globe. However, many challenges prevent their wide adoption, specifically their limited flight time. In order to alleviate this problem, previous work done by Habas et al implemented an inverted landing strategy using bio-inspired visual cues that were calculated with external positioning. This thesis extends the previous work by estimating these visual cues on-board the micro unmanned aerial vehicle. Additionally, the limits of the developed on-board visual cue estimation algorithm are explored in regards to ceiling pattern, frame rate, and computational load. Once the visual cue algorithm was explored, it was run through the same reinforcement learning algorithm from Habas et al to determine the landing robustness. In nature inverted landing maneuvers are observed in several organisms such as flies, bees, and bats. In this thesis, the maneuvers observed in blue bottle flies are used as inspiration to achieve the inverted landings for a small quad-copter. In order to implement the inverted landing strategy, two biological cues are used, time to contact and optical flow. Specifically, for the blue bottle fly, time to contact determines when the flip maneuver should initiate, while optical flow helps with the final landing phase. To emulate these biological cues the Horn and Schunck method is employed to segment image data from an onboard camera. The major contribution of this work is to explore the main variables that introduce error due to the onboard noisy sensors and how they affect landing performance. In order to explore these variables the algorithm testing and the landing policy learning were conducted in a physics-based simulation. Future work will then be done in order to experimentally determine the landing robustness of the implemented algorithm.