Chen, Xieyuanli: LiDAR-Based Semantic Perception for Autonomous Vehicles. - Bonn, 2022. - Dissertation, Rheinische Friedrich-Wilhelms-Universität Bonn.
Online-Ausgabe in bonndoc:
author = {{Xieyuanli Chen}},
title = {LiDAR-Based Semantic Perception for Autonomous Vehicles},
school = {Rheinische Friedrich-Wilhelms-Universität Bonn},
year = 2022,
month = sep,

note = {Scene understanding is one of the fundamental building blocks that enable mobile systems to achieve autonomy. It is the process of perceiving, analyzing, and elaborating an interpretation of a 3D dynamic scene ob- served through the onboard sensors equipped on autonomous vehicles. The light detection and ranging sensors, in short LiDAR, are one of the popular sensors for autonomous vehicles to sense their surroundings, because they are robust to light changes and provide high-accurate range measurements. Based on LiDAR sensors, autonomous vehicles can explore environments, understand the locations and types of objects therein, and then make plans and execute actions to fulfill complex tasks. Among them, key capabilities are localization within a given map as well as simultaneous localization and mapping (SLAM), which pro- vide the robots location, the necessary prerequisite for other downstream tasks. Traditional LiDAR-based global localization and SLAM methods can provide ac- curate pose estimates in indoor environments with the static world assumption. However, as the demand for autonomous driving in dynamic outdoor environments grew, using only geometric and appearance information is not enough to provide reliable localization and mapping results for autonomous systems. A high-level understanding of the world, which includes the estimation of semantic information, is required for robust and safe deployments of autonomous vehicles in dynamic and complex real-world scenarios.
The main contributions of this thesis are novel approaches that exploit se- mantic information to improve the performance of LiDAR perception tasks such as SLAM and global localization for autonomous vehicles. This thesis consists of three parts. The first part focuses on how to apply semantic information for SLAM and localization. We present a semantic-based LiDAR SLAM method, which exploits semantic predictions from an off-the-shelf semantic segmentation network to improve the pose estimation accuracy and generate consistent seman- tic maps of the environments. We furthermore propose a novel neural network exploiting both geometric and semantic information to estimate the similarities between pairs of LiDAR scans. Based on these similarity estimates, our network can better find loop closure candidates for SLAM and achieve global localization in outdoor environments across seasons.
The second part investigates which type of semantics are useful for specific tasks. In this context, we propose a novel moving object segmentation method for SLAM. It aims at separating the actually moving objects such as driving cars from static or non-moving objects such as buildings, parked cars, etc. With more specific moving/non-moving semantics, we get a better SLAM performance compared to setups using general semantics. For localization, we propose to use pole-like objects such as tra?ic signs, poles, lamps, etc., due to their local distinctiveness and long-term stability. As a result, we obtain reliable and accurate localization results over comparably long periods of time.
Deep learning-based approaches can provide accurate point-wise semantic pre-dictions. They, however, strongly rely on the diversity and amount of labeled training data that may be costly to obtain. In the third part, we therefore propose approaches that can automatically generate labels for training neural networks. Benefiting from specifying and simplifying the semantics for specific tasks, we turn the comparably challenging multiclass semantic segmentation problem into more manageable binary classification tasks, which makes automatic label generation feasible. Using our proposed automatic labeling approach, we alleviate the reliance on expensive human labeling for supervised training of neural networks and enable our method to work in a self-supervised way. Therefore, our pro- posed task-specific semantic-based methods can be easily transferred to different environments with different LiDAR sensors.
All our proposed approaches presented in this thesis have been published in peer-reviewed conference papers and journal articles. Our proposed OverlapNet for LiDAR-based loop closing and localization was nominated for the Best System Paper at the Robotics: Science and Systems (RSS) conference in 2020. Our proposed moving object segmentation method was selected to be presented at the Robotics Science and Systems Pioneers event in 2021. Additionally, we have made implementations of all our methods presented in this thesis open-source to facilitate further research.},

url = {}

The following license files are associated with this item: