Access Type
Open Access Dissertation
Date of Award
January 2023
Degree Type
Dissertation
Degree Name
Ph.D.
Department
Computer Science
First Advisor
Loren Schwiebert
Abstract
High Definition (HD) Mapping for Autonomous Driving creates an easily accessible wealth of information to confirm and enrich feedback from on-vehicle sensors. HD Maps built offline generally use dedicated ground collection vehicles equipped with high-cost, survey-grade LiDAR to accurately map a vehicle's areas of operation. We focus our work on controlled-access, divided highways characterized by wide roadways that require complex permutations to cover all directions of travel and access ramps. We study two common issues when collecting this road class that cause low-fidelity (LF) point clouds for HD Maps. First, aerial collection covers all directions and all ramps in a single pass; however, this method creates sparse point clouds due to the increased range to target. Second, vehicle-based occlusions block required map information and they are unavoidable with ground collection. Instead of incurring costly multi-pass collection, we propose novel methods to solve these problems. In this paper, we discuss a highly accurate 4D LiDAR infill using a novel Bird's-Eye View (BEV) data preparation catered to HD Map building. Our approach utilizes a series of dimensionality reductions to simplify the data representation prior to dense pixel estimation. Further, we propose two techniques that apply realistic sensor degradation onto high-fidelity (HF) collects to generate cost-effective, synthetic supervised training sets to address these common problems. We show our deep generative models complete the 4D data at the desired fidelity to reduce the cost of HD Maps.
Recommended Citation
Bojrab, Micah, "Improving Lidar Fidelity For Hd Mapping Using Pattern Transplanting And Temporal Infill" (2023). Wayne State University Dissertations. 3791.
https://digitalcommons.wayne.edu/oa_dissertations/3791