Bio-inspired vision-based evasion control: collision avoidance without distance measurement

In the past decade, the appearance of multiple autonomous vehicle platforms - such as self-driving cars (SDC), unmanned aerial vehicles (UAVs), delivery robots and precision agriculture drones - have emerged with an accelerated pace. Their development has been driven by enormous research at the intersection of control technology and machine learning, enabling autonomous operations, and minimizing human intervention. While the minimization of human intervention has the objective of minimizing the impact of potential human errors, it comes at the price of rigorously formulating and solving challenging problems as collision avoidance that humans quite often do subconsciously and with ease. This dissertation introduces a framework for collision avoidance, where the measurement of the distance to objects and obstacles is not available. This limitation is common to all low-cost and small, ground, or flying vehicles that are not equipped with expensive cameras. The proposed solution takes inspiration from biological systems and the mechanisms that invertebrates and birds use to evade predators. Psychological evidence shows that animals are capable of evading eminent collisions without using depth information, relying instead on looming stimuli. In contrast, the field of robotics has solved collision avoidance among uncooperative vehicles by using depth (the relative distance) to the obstacles as feedback, measured e.g. by lidar, which can be very expensive. To bridge this gap, this works presents a different paradigm in the sensor measurements required for collision avoidance. Relying solely on information that can be directly acquired from a monocular camera this dissertation outlines three control strategies suitable for unicycle-like vehicles avoiding a single, unknown, dynamic uncooperative obstacle: (i) using a line-of-sight (LOS) only measurement, (ii) using a LOS measurement and time-to-collision, and (iii) a LOS, LOS rate and time-to-collision based algorithm. These quantities can readily be estimated from a monocular camera vision system on board the vehicle. Under reasonable assumptions theoretical guarantees are obtained that ensure collision avoidance with an uncooperative moving obstacle.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Mechanical Engineering

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2019

Author dc:creator

• Marinho, Thiago

Contributors dc:contributor

• Hovakimyan, Naira
• Voulgaris, Petros
• Stipanovic, Dusan
• Salapaka, Srinivasa

• Collision Avoidance, Evasion Control, Autonomous Systems.

Statement dc:rights

• Copyright 2019 Thiago Marinho

Language dc:language

en

Handle dc:identifier

http://hdl.handle.net/2142/105788

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/105788