Date of Award

Spring 4-25-2024

Degree Type

Thesis

Degree Name

Master of Science in Software Engineering

Department

Department of Software Engineering

Committee Chair/First Advisor

Dr. Sungchul Jung

Second Advisor

Dr. Arthur Choi

Third Advisor

Dr. Michael Franklin

Abstract

Semantic segmentation of point clouds is a basic step for many autonomous systems including automobiles. In autonomous driving systems, LiDAR sensors are frequently used to produce point cloud sequences that allow the system to perceive the environment and navigate safely. Modern machine learning techniques for segmentation have predominately focused on single-scan segmentation, however sequence segmentation has often proven to perform better on common segmentation metrics. Using the popular Semantic KITTI dataset, we show that by providing point cloud sequences to a segmentation pipeline based on Point Transformer v3, we increase the segmentation performance between seven and fifteen percent when compared to the single-scan baseline. Experiments using two segmentation network sizes and several different hyperparameters are performed to show these increases.

Share

COinS