Author: Jean-Emmanuel DESCHAUD

This thesis was realized under the direction of François Goulette and Jean-Emmanuel Deschaud of the Center of Robotics of Mines Paris - PSL.

The defense took place on Monday 10 March 2025 at Mines Paris - PSL.

Jury members:

• NUCHTER Andreas, Professor at Wurtzbourg University - Rapporteur
• CHECCHIN Paul, Professor at Clermont Auvergne University - Rapporteur
• CHAMBON Sylvie, Professor at Institut de Recherche en Informatique de Toulouse (IRIT) - Examinatrice
• GOULETTE François, Professor at ENSTA Paris - IP Paris - Thesis Director
• DESCHAUD Jean-Emmanuel, Associate Professor at Mines Paris - PSL University - Thesis Co-Director

Summary of the PhD thesis:

The development of robust perception systems is fundamental for the safe and efficient operation of autonomous vehicles. These systems perform essential tasks such as 3D object detection, enabling vehicles to identify and localize obstacles, including other vehicles, pedestrians, and various objects within their environment. Accurate detection is critical for effective decision-making and navigation through complex driving scenarios. However, 3D object detection presents significant challenges due to the diverse nature of real-world conditions, which encompass a broad range of sensor setups, geographic environments, and scene complexities.

Firstly, this thesis work identifies and addresses the domain shifts that occur between different LiDAR datasets due to variations in sensor specifications, geographic environments, and dataset-specific attributes. These shifts often lead to significant performance gaps when models are transferred between datasets. To mitigate these issues, a multi-dataset training framework called MDT3D is introduced. MDT3D integrates data from various sources and employs novel augmentation and label harmonization techniques to create models that can generalize effectively across different conditions.

Secondly, the thesis presents ParisLuco3D, a dataset captured in urban areas around the Luxembourg Garden in Paris, designed to test model robustness in complex real-world scenarios. This provides a dedicated dataset to test model generalization, which we benchmark to highlight the poor performance of baseline generalization methods.

Lastly, the last axis of generalization explored is generalizing detection to novel unknown objects. We reframe the detection of unknown objects as an out-of-distribution (OOD) problem, allowing models to differentiate between known and previously unseen objects without compromising accuracy on familiar categories.

Congratulations to Louis !!!

This thesis was realized under the direction of François Goulette and Jean-Emmanuel Deschaud of the Center of Robotics of Mines Paris - PSL.

The defense took place on Tuesday 5 December 2023 at Mines Paris - PSL.

Jury members:

• LANDRIEU Loïc, Chargé de recherche, HDR, École des Ponts ParisTech
• CHECCHIN Paul, Professeur des universités, Université Clermont Auvergne
• CHAINE Raphaëlle, Professeur des universités, Université Claude Bernard Lyon 1, LIRIS
• GOULETTE François, Professeur ENSTA Paris - IP Paris
• DESCHAUD Jean-Emmanuel, Chargé de recherche HDR, Mines Paris - PSL

Summary of the PhD thesis:

LiDAR perception for autonomous vehicles managed to achieve suitable results on the various online benchmarks in the single-domain framework, that is to say when the training domain is the same as the evaluation domain. From there, the fields of research diversified and focused on questions of transferability, robustness and generalization.

This work focuses on generalization issues for LiDAR semantic segmentation. A global overview of the generalization performances, single-source and multi-source, of existing segmentation methods is carried out. To fairly perform these experiments, a dataset, ParisLuco3D, is created specifically to evaluate generalization.

Furthermore, a new single-source domain generalization method, 3DLabelProp, is proposed. This method differs from existing strategies by exploiting the geometry of the data to perform domain alignment rather than learning strategies. Beyond semantic segmentation, this method is also applied to the task of moving object segmentation.

Congratulations to Jules !!!

This thesis was realized in collaboration between Mines Paris - PSL and Kitware under the direction of François Goulette and Jean-Emmanuel Deschaud of the Center of Robotics of Mines Paris - PSL, as well as Raphaël Cazorla from Kitware.

The defense took place on Friday 10 November 2023 at Mines Paris - PSL.

Summary of the PhD thesis: 3D LiDARs have become increasingly popular in the past decade, notably motivated by the safety requirements of autonomous driving requiring new sensor modalities. Contrary to cameras, 3D LiDARs provide direct, and extremely precise 3D measurements of the environment. This has led to the development of many different mapping and Simultaneous Localization And Mapping (SLAM) solutions leveraging this new modality. These algorithms quickly performed much better than their camera-based counterparts, as evidenced by several open-source benchmarks. One critical component of these systems is LiDAR odometry. A LiDAR odometry is an algorithm estimating the trajectory of the sensor, given only the iterative integration of the LiDAR measurements. The focus of this work is on the topic of LiDAR Odometries. More precisely, we aim to push the boundaries of LiDAR odometries, both in terms of precision and performance. To achieve this, we first explore classical LiDAR odometries in depth, and propose two novel LiDAR odometries, in chapter 3. We show the strength, and limitations of such methods. Then, to address to improve them we first investigate Deep Learning for LiDAR odometries in chapter 4, notably focusing on end-to-end odometries. We show again the limitations of such approaches and finally investigate in chapter 5 fusing inertial and LiDAR measurements.

Congratulations to Pierre !!!

This thesis was realized in collaboration between Mines Paris - PSL and Ansys under the direction of François Goulette and Jean-Emmanuel Deschaud of the Center of Robotics of Mines Paris - PSL, as well as Nicolas Dalmasso, Nassim Jibai, and David Ganieux from Ansys.

The defense took place on Thursday 08 December 2022 at Mines Paris - PSL.

Summary of the PhD thesis: development of autonomous vehicles and testing can be achieved by simulating the different sensors in a virtual environment. However, manually handcrafted virtual environments fail to generalize to real-world scenes, due to the over-simplified models used in such environments. In this thesis, we propose to reduce this domain gap by leveraging real-world scans of urban scenes in the form of 3D point clouds acquired using a LiDAR sensor mounted on a mobile mapping system. Toward this end, we propose an automatic simulation pipeline. The pipeline introduces an accurate scene modeling method based on semantic adaptive splatting to create the virtual environment and a real-time LiDAR simulation.

Congratulations to Jean-Pierre !!!