Author ORCID Identifier

https://orcid.org

Semester

Summer

Date of Graduation

2026

Document Type

Dissertation

Degree Type

PhD

College

Statler College of Engineering and Mineral Resources

Department

Mechanical and Aerospace Engineering

Committee Chair

Jason N. Gross

Committee Member

Cagri Kilic

Committee Member

Guilherme Pereira

Committee Member

Yu Gu

Committee Member

Amr El-Wakeel

Abstract

Many LiDAR odometry and SLAM methods are lightweight and effective, but they assume that all LiDAR point returns are equally reliable. In practice, this assumption is often violated in vegetation rich or semi transparent environments, where in a single beam footprint, LiDAR beams may intersect multiple scattering surfaces like leaves, branches or glass. These ambiguities in returns produces probabilistic or biased range measurements that violate the assumptions of most geometric registration frameworks and can lead to drift and inconsistency.

To address this challenge, this dissertation proposes a semantic-weighted LiDAR-camera SLAM framework that integrates class dependent weights into a front-end LiDAR odometry optimization. The front-end odometry framework in this dissertation is built on top of KISS-ICP, a well known lightweight LiDAR odometry framework. Using a calibrated LiDAR-camera pair, LiDAR points are projected onto a semantically segmented image to acquire class labels. During pointcloud registration, the residuals are scaled both by a robust kernel and semantic weights that down weight points from classes prone to range ambiguity. The resulting semantically robust odometry front end outputs RGB-colored and semantic-colored pointclouds to a factor-graph SLAM backend that performs global optimization. This proposed system aims to enhance odometry stability and improve overall map consistency in complex natural and agricultural environments where purely geometric methods struggle.

Download

Included in

Robotics Commons

Share

COinS