Collision Detection with Bio-inspired Image Processing for Flying Robots in Dynamic Environments

Zhao, Jiannan (2021) Collision Detection with Bio-inspired Image Processing for Flying Robots in Dynamic Environments. PhD thesis, University of Lincoln.

Documents
Collision Detection with Bio-inspired Image Processing for Flying Robots in Dynamic Environments
Published thesis

Request a copy
[img] PDF
JiannanZhao_PhD_Thesis_Final.pdf - Whole Document
Restricted to Repository staff only

24MB
Item Type:Thesis (PhD)
Item Status:Live Archive

Abstract

Detecting collision in dynamic environments is a vital capability for both animals to survive and robots to serve the human society. No customer would like to have an autonomous robot around to collide with objects or humans once in a while, especially for flying robots.
Since the Unmanned Aerial Vehicles (UAVs) are vulnerable to crashes and the onboard computational resources are limited due to payload capability, it is a huge challenge for UAVs to sense and avoid collision in complex environments. Traditional collision avoidance approaches for ground robots that require a heavy sensor system or massive computational power are not affordable for small UAVs.
On the other hand, bio-inspired collision detection and avoidance (CDAA) methods, which make use of strategies sourced in nature, are usually advanced in efficiency. In particular, fly insects inspired strategies are perfectly appropriate for small UAVs due to the compact size of their sensor system and the amazing "sense and avoid" ability in flight. For example, locusts are able to migrate in thousands of millions, they swarm in fields and forests without colliding into the environment or each other. Their ability to efficiently perceive a looming danger is mainly relying on a single visual neuron named the Lobula Giant Movement Detector (LGMD). Studying the visual pathways of flying insects and their underlying mechanisms is hence not only critical to neuroscience but also beneficial for developing artificial CDAA systems in a both efficient and robust manner.

This study aims to explore CDAA systems that are advanced in both efficiency and robustness for autonomous flying robots, with the inspiration from locust's LGMD neuron. Firstly, a research-oriented quadcopter platform is designed to implement LGMD based CDAA methods in real-time flight. In order to direct avoidance behaviour in 3D space, a competitive control strategy is introduced by separately processing information from competitive parts of the field of view. The LGMD collision detection algorithm and competitive control strategy are systematically tested in both simulation and real-time autonomous flight. Then, in order to overcome dynamic interferences caused by ego-motion during agile flight, a presynaptically distributed LGMD model is proposed, namely (D-LGMD). This model selectively enhances the barrier against low speed image edges from backgrounds, by setting a filter with spatial-temporal distribution. This filter mimics the synaptic topological mapping in the locust LGMD pathway by involving novel presynaptic strategies such as: distributed excitation, self-inhibition, and radially distributed latency. The D-LGMD model presented strong robustness in collision detection in flight against irrelevant background motion and responded only briefly to receding stimuli in experiments. Finally, to implement the D-LGMD in real-time flight, an OpenMV platform is connected with the quadcopter platform to run the D-LGMD algorithm. The system is able to quickly avoid dynamic obstacles in real-time flight and exhibits robustness even during attitude motion, ambient light changes, or towards complex background.

Keywords:UAV, Collision detection, LGMD
Subjects:G Mathematical and Computer Sciences > G700 Artificial Intelligence
H Engineering > H671 Robotics
Divisions:College of Science > School of Computer Science
ID Code:47317
Deposited On:19 Nov 2021 11:50

Repository Staff Only: item control page