Learning visual odometry primitives for computationally constrained platforms

Abstract

Autonomous navigation for robotic platforms, particularly techniques that leverage an onboard camera, are of currently of significant interest to the robotics community. Designing methods to localize small, resource-constrained robots is a particular challenge due to limited availability of computing power and physical space for sensors. A computer vision, machine learning-based localization method was proposed by researchers investigating the automation of medical procedures. However, we believed the method to also be promising for low size, weight, and power (SWAP) budget robots. Unlike for traditional odometry methods, in this case, a machine learning model can be trained offline, and can then generate odometry measurements quickly and efficiently. This thesis describes the implementation of the learning-based, visual odometry method in the context of autonomous drones. We refer to the method as RetiNav due to its similarities with the way the human eye processes light signals from its surroundings. We make several modifications to the method relative to the initial design based on a detailed parameter study, and we test the method on a variety of challenging flight datasets. We show that over the course of a trajectory, RetiNav achieves as low as 1.4% error in predicting the distance traveled. We conclude that such a method is a viable component of a localization system, and propose the next steps for work in this area.