Hardware Considerations of Visual Odometry Algorithms

Abstract

This thesis explores the wide world of visual odometry algorithms through the lens of the highly resource constrained environment of deployment on a microrobotic bee. Early on, these constraints lead to the choice to use an event detection camera instead of a more traditional frame based camera. After discussion of a subset of the broad body of existing work in visual odometry, it is hypothesized that results from the well studied field of frame based visual odometry can be applied to event based data after the creation of ``event images". A wide range of results are tested and their performance is discussed and reasoned about. A novel feature detector, FASTER, made by slight modification of the FAST corner detector to improve its performance on event based data is introduced and tested with good results. Different approaches to event tracking are discussed and tested. Pose estimation from this data is considered and tested with limited success. Finally, a high level analysis of a complete visual odometry pipeline for event data is performed and the feasibility of running the full algorithm on the microrobotic bee is considered and ultimately rejected for this stage of the project.