Author ORCID Identifier

https://orcid.org/0000-0002-5481-3969

Semester

Fall

Date of Graduation

2023

Document Type

Dissertation

Degree Type

PhD

College

Statler College of Engineering and Mineral Resources

Department

Mechanical and Aerospace Engineering

Committee Co-Chair

Jason N. Gross

Committee Member

Yu Gu

Committee Member

Guilherme A. S. Pereira

Committee Member

Natalia A. Schmid

Committee Member

Nicholas Szczecinski

Committee Member

Ryan M. Watson

Abstract

State estimation is an integral component of any autonomous robotic system. Finding the correct position, velocity, and orientation of an agent in its environment enables it to do other tasks like mapping and interacting with the environment, and collaborating with other agents. State estimation is achieved by using data obtained from multiple sensors and fusing them in a probabilistic framework. These include inertial data from Inertial Measurement Unit (IMU), images from camera, range data from lidars, and positioning data from Global Navigation Satellite Systems (GNSS) receivers. The main challenge faced in sensor-based state estimation is the presence of noisy, erroneous, and even lack of informative data. Some common examples of such situations include wrong feature matching between images or point clouds, false loop-closures due to perceptual aliasing (different places that look similar can confuse the robot), presence of dynamic objects in the environment (odometry algorithms assume a static environment), multipath errors for GNSS (signals for satellites jumping off tall structures like buildings before reaching receivers) and more. This work studies existing and new ways of how standard estimation algorithms like the Kalman filter and factor graphs can be made robust to such adverse conditions without losing performance in ideal outlier-free conditions. The first part of this work demonstrates the importance of robust Kalman filters on wheel-inertial odometry for high-slip terrain. Next, inertial data is integrated into GNSS factor graphs to improve the accuracy and robustness of GNSS factor graphs. Lastly, a combined framework for improving the robustness of non-linear least squares and estimating the inlier noise threshold is proposed and tested with point cloud registration and lidar-inertial odometry algorithms followed by an algorithmic analysis of optimizing generalized robust cost functions with factor graphs for GNSS positioning problem.

Download

Share

COinS