Deep Learning for Autonomous Driving: Calibration of Localization Uncertainty in 3D Object Detectors

Institut

Lehrstuhl für Fahrzeugtechnik

Typ

Masterarbeit /

Inhalt

experimentell / theoretisch /  

Beschreibung

Help autonomous cars plan safer, more reliable trajectories by making their object detectors honest about what they don’t know!

Even top-tier detectors won’t be perfect. Downstream modules—tracking, prediction, and trajectory planning—work best when they can trust the detector’s confidence and localization uncertainty. Today’s networks can output per-object uncertainties, but these are often miscalibrated and don’t match empirical error. Calibrated uncertainties are the missing link to robust decision-making.

Your mission

Compare existing uncertainty calibration techniques and develop a method that works exceptionally well for our 3D object detector—ideally robust enough to run on our research vehicle EDGAR.

What you’ll do

1. Survey the field: Review uncertainty calibration for (3D) object detection.
2. Measure the gap: Quantify miscalibration in state-of-the-art 3D detectors.
3. Build baselines: Implement promising confidence/uncertainty calibration methods.
4. Calibrate & evaluate: Apply and rigorously evaluate calibration on 3D detectors.
5. Develop a calibration strategy that combines what you have learned from literature and analysis

What you’ll learn

• Practical uncertainty estimation & calibration (e.g., reliability diagrams, ECE/NLL).
• 3D detection evaluation
• Hands-on deep learning engineering in Python (PyTorch or similar), experiment design, and reproducible research.

Voraussetzungen

You are a great fit if you

• Have a strong interest in autonomous driving and object detection.
• Bring an engaged, independent working style.
• Have solid Python skills (experience with deep learning frameworks is a plus).

Should you be interested in this thesis project or any other project in the context of perception for autonomous driving, please send a CV and a transcript of records to cornelius.schroeder@tum.de.

Tags

FTM Studienarbeit, FTM AV, FTM Schroeder, FTM Informatik

Möglicher Beginn

sofort

Kontakt

Cornelius Schröder, M.Sc.
cornelius.schroedertum.de

Ausschreibung