View-Agnostic Point Cloud Generation

Date of Award

2022

Degree Name

Ph.D. in Computer Engineering

Department

Department of Electrical and Computer Engineering

Advisor/Chair

Vijayan Asari

Abstract

Occlusions are one of the primary data challenges when working with lidar. Unfortunately, occlusions are highly dependent on sensor viewpoints, and efforts to mitigate occlusions involve costly data collection strategies like additional overlap or multiple views. This research focuses on reducing occlusions by generating the missing points in a post-processing step. We introduce an entirely new occlusion dataset for aerial lidar called DALES Viewpoints. We also propose two fundamental changes that we can use in conjunction with current point cloud completion networks to provide an appropriate solution for occlusion reduction in aerial lidar. Specifically, we propose a new method of Eigen feature selection for hierarchical downsampling. This method takes into account point features, in addition to spatial location. We also introduce a point correspondence loss that helps build more robust features by ensuring similar network behavior when processing point clouds that depict the same scene with different physical point locations.

Keywords

Artificial Intelligence, Computer Science, Statistics, lidar, aerial lidar, 3d, point clouds, occlusion, autoencoder

Rights Statement

Copyright © 2022, author

Link to Full Text

Share

COinS