Tesi etd-01222024-104602
Link copiato negli appunti
Tipo di tesi
Dottorato
Autore
D'AMICO, GIANLUCA
URN
etd-01222024-104602
Titolo
Advancing Railway Safety and Automation: Graphic Engine Simulation for Sensory Data Generation in Railway Environments
Settore scientifico disciplinare
ING-INF/05
Corso di studi
Istituto di Tecnologie della Comunicazione, dell'Informazione e della Percezione - PHD IN EMERGING DIGITAL TECHNOLOGIES
Commissione
relatore BUTTAZZO, GIORGIO CARLO
Presidente Prof. SANNA, ANDREA
Membro Prof. IANNIZZOTTO, GIANCARLO
Membro Dott.ssa COLLA, VALENTINA
Presidente Prof. SANNA, ANDREA
Membro Prof. IANNIZZOTTO, GIANCARLO
Membro Dott.ssa COLLA, VALENTINA
Parole chiave
- Railway Environments
- Simulation Frameworks
- Synthetic Dataset Generation
- Graphic Engine
- Perception Algorithms
Data inizio appello
27/05/2024;
Disponibilità
parziale
Riassunto analitico
In the domain of railway transportation, there is a growing interest in developing and implementing innovative solutions that significantly enhance the safety and operational efficiency of critical functions like train localization.
This extends to the automation of various train operations, such as signal recognition, obstacle detection, and track discrimination.
These enhancements are vital for ensuring the overall reliability of railway networks.
These tasks require the artificial perception of the railway environment through the data acquired from different types of sensors, such as cameras, LiDARs, wheel encoders, GNSS receivers, and inertial measurement units.
However, a significant challenge in these perception tasks is the acquisition of extensive, accurately labeled datasets that reflect a wide range of scenarios and operational conditions.
In the railway sector, this is particularly crucial due to the stringent regulatory framework governing railway operations and the practical difficulties associated with accessing railway infrastructures.
The constraints extend to the installation and use of the necessary sensors on actual trains, making the collection of real-world data a complex, time-consuming, and often unfeasible task.
To overcome such difficulties, this thesis introduces two innovative visual simulation frameworks: the first one, based on Unreal Engine 4, was created within an industrial project to generate visual data, such as images and point clouds, with the purpose of integrating visual and LiDAR odometry algorithms into a tightly coupled Kalman filter for train navigation.
The second framework was developed as an independent research project on top of Unreal Engine 5, to improve the realism of the images collected in virtual environments and enhance the creation of high-fidelity environments that reflect the complexity and diversity of real-world railway settings.
Both frameworks allow the generation of diverse labeled datasets from a suite of emulated sensors, including RGB and depth cameras, LiDARs, and inertial measurement units.
These simulated environments and the consequently generated datasets are invaluable for several purposes.
They serve as a test-bed for pioneering algorithms, providing a controlled yet realistic setting for algorithm development and refinement.
Furthermore, they are instrumental in the training and evaluation of deep neural networks.
These networks are employed in various complex tasks, such as detailed image segmentation, accurate object detection, visual odometry for understanding train movement, and the nuanced task of track discrimination.
An additional notable contribution of this dissertation is the presentation of specialized use cases aimed at validating the data generated within the presented frameworks. The experimental results and insights presented in this dissertation underscore the immense potential of visual simulation frameworks.
They stand as novel tools in the advancement of railway safety and the automation of operational processes, offering a path forward for researchers and developers in this specialized field.
The rest of the thesis is organized as follows:
The first chapter provides an introduction to the topic, highlighting the research context, stating the problem, outlining the objectives, and detailing the contributions of this work.
The second chapter offers a comprehensive review of the related work across various application domains, focusing on public datasets, synthetic data generation, and simulation frameworks, presenting the research gap specifically related to the railway domain.
It also introduces the validation processes for simulated data and the rationale behind selecting Unreal Engine as the foundational 3D creation tool.
Chapters three and four present the two simulation frameworks discussed in this thesis, namely TrainSim (under UE4) and RailSim (under UE5), offering insights into their design decisions.
Each of these chapters also presents the experimental studies carried out to validate the effectiveness of the simulated sensors and the potential usage of the frameworks.
Finally, Chapter five summarizes the main contributions of this work and outlines potential directions for future research.
This extends to the automation of various train operations, such as signal recognition, obstacle detection, and track discrimination.
These enhancements are vital for ensuring the overall reliability of railway networks.
These tasks require the artificial perception of the railway environment through the data acquired from different types of sensors, such as cameras, LiDARs, wheel encoders, GNSS receivers, and inertial measurement units.
However, a significant challenge in these perception tasks is the acquisition of extensive, accurately labeled datasets that reflect a wide range of scenarios and operational conditions.
In the railway sector, this is particularly crucial due to the stringent regulatory framework governing railway operations and the practical difficulties associated with accessing railway infrastructures.
The constraints extend to the installation and use of the necessary sensors on actual trains, making the collection of real-world data a complex, time-consuming, and often unfeasible task.
To overcome such difficulties, this thesis introduces two innovative visual simulation frameworks: the first one, based on Unreal Engine 4, was created within an industrial project to generate visual data, such as images and point clouds, with the purpose of integrating visual and LiDAR odometry algorithms into a tightly coupled Kalman filter for train navigation.
The second framework was developed as an independent research project on top of Unreal Engine 5, to improve the realism of the images collected in virtual environments and enhance the creation of high-fidelity environments that reflect the complexity and diversity of real-world railway settings.
Both frameworks allow the generation of diverse labeled datasets from a suite of emulated sensors, including RGB and depth cameras, LiDARs, and inertial measurement units.
These simulated environments and the consequently generated datasets are invaluable for several purposes.
They serve as a test-bed for pioneering algorithms, providing a controlled yet realistic setting for algorithm development and refinement.
Furthermore, they are instrumental in the training and evaluation of deep neural networks.
These networks are employed in various complex tasks, such as detailed image segmentation, accurate object detection, visual odometry for understanding train movement, and the nuanced task of track discrimination.
An additional notable contribution of this dissertation is the presentation of specialized use cases aimed at validating the data generated within the presented frameworks. The experimental results and insights presented in this dissertation underscore the immense potential of visual simulation frameworks.
They stand as novel tools in the advancement of railway safety and the automation of operational processes, offering a path forward for researchers and developers in this specialized field.
The rest of the thesis is organized as follows:
The first chapter provides an introduction to the topic, highlighting the research context, stating the problem, outlining the objectives, and detailing the contributions of this work.
The second chapter offers a comprehensive review of the related work across various application domains, focusing on public datasets, synthetic data generation, and simulation frameworks, presenting the research gap specifically related to the railway domain.
It also introduces the validation processes for simulated data and the rationale behind selecting Unreal Engine as the foundational 3D creation tool.
Chapters three and four present the two simulation frameworks discussed in this thesis, namely TrainSim (under UE4) and RailSim (under UE5), offering insights into their design decisions.
Each of these chapters also presents the experimental studies carried out to validate the effectiveness of the simulated sensors and the potential usage of the frameworks.
Finally, Chapter five summarizes the main contributions of this work and outlines potential directions for future research.
File
| Nome file | Dimensione |
|---|---|
Ci sono 1 file riservati su richiesta dell'autore. |
|