3D Map Evaluation in LiDAR Point Clouds Using Deep Neural Networks:

Dataset, Framework, DNN Architecture, and Methods

Christopher Jan Plachetka

• Publisher: Shaker
• Year: 2025
• eBook ISBN: 9783819100147
• Edition: 0

This dissertation examines the evaluation of high-definition maps using LiDAR point clouds as sensor data and deep neural networks (DNNs). The developed methods are suitable for both automated driving and geodesy. By utilizing the map as an additional input, the DNN can verify the map even under adverse conditions that degrade the sensor input. This work introduces a novel dataset, a conceptual framework, a specialized DNN architecture, and innovative evaluation methods, contributing significantly to the field.

Table of contents

• Abstract
• Zusammenfassung
• Disclaimer
• Contents
• Introduction
• 3DHD CityScenes: A Novel LiDAR and HD Map Dataset
• A Novel Conceptual Framework for Dependable Maps
• 3DHDNet: A Novel Network Architecture for 3D Object Detection
• Novel Methods for DNN-Based 3D Map Evaluation
• Conclusions
• Notations
• Acronyms
• Procedure for Improving Annotation Completeness
• System Requirements for Dependable Maps
• Augmentation Studies
• Statistical Significance of Performance Results
• Ablation on Downstream Network Strides
• Map Trust: Driving Under Correct Map Assumptions
• Supplementary Results
• First Author Publications
• Bibliography