Deep Learning for Multi-Sensor Calibration in Autonomous Driving

Abstract

This thesis covers online extrinsic calibration for autonomous driving: camera–LiDAR, LiDAR–event, and uncertainty-aware methods. It includes PseudoCal (BMVC 2023), UniCal (international patent), MULi-Ev (CVPR 2024 workshop), and conformal-prediction intervals for online calibration (WACV 2025).

Methods are evaluated on autonomous driving datasets including KITTI, DSEC, and nuScenes. Defended in April 2025 at the Université de technologie de Compiègne, under the supervision of Julien Moreau and Franck Davoine at the Heudiasyc Laboratory.

Key Contributions

• Uncertainty-aware calibration: first approach integrating conformal prediction into online extrinsic calibration with guaranteed coverage intervals
• Initialisation-free calibration: PseudoCal — first deep learning method for camera-LiDAR calibration without manual initialization, operating directly in 3D space
• LiDAR-event camera calibration: MULi-Ev — first online framework for extrinsic calibration of event cameras with LiDAR
• Transformer-based architecture: UniCal — efficient single-branch model for simultaneous calibration and validation, leading to an international patent
• Validation: evaluated on KITTI, DSEC, and nuScenes

Publications from this Thesis

• WACV 2025 — Uncertainty-Aware Online Extrinsic Calibration: A Conformal Prediction Approach — Read more →
• CVPR 2024 — MULi-Ev: Maintaining Unperturbed LiDAR-Event Calibration — Read more →
• BMVC 2023 — PseudoCal: Towards Initialisation-Free Camera-LiDAR Self-Calibration — Read more →
• Patent — UniCal: A Single-Branch Transformer-Based Model for Camera-to-LiDAR Calibration — Read more →

Citation

@phdthesis{cocheteux2025thesis,
    author = {Cocheteux, Mathieu},
    title  = {Deep Learning for Multi-Sensor Calibration
              in Autonomous Driving},
    school = {Université de technologie de Compiègne},
    year   = {2025},
    type   = {PhD Thesis},
    url    = {https://theses.hal.science/tel-05268827v1}
}