Uncertainty-Aware 3D Object Detection for Autonomous Driving (Transformer-Based)

Institut

Lehrstuhl für Fahrzeugtechnik

Typ

Masterarbeit /

Inhalt

experimentell / theoretisch /  

Beschreibung

Autonomous vehicles live or die by their perception stack. Transformer-based 3D detectors are the new state of the art—but great mAP alone isn’t enough on the road. We also need well-calibrated uncertainty so downstream modules (tracking, motion prediction, planning) can make safer, smarter decisions.

In this thesis, you’ll bring epistemic and aleatoric uncertainty estimation to modern transformer-based 3D detectors, and investigate how the attention mechanism itself can be leveraged to improve calibration. Your work helps pave the way for deploying a more capable, uncertainty-aware detector on our research vehicle EDGAR (deployment is a mid-term goal and outside this thesis scope).

What you’ll do

• Survey the field: Review and compare leading transformer-based 3D object detectors.
• Build & evaluate: Implement epistemic and aleatoric uncertainty estimation by transferring our existing techniques to a selected transformer architecture.
• Leverage attention: Explore how transformer attention can be exploited for better-calibrated uncertainty (and rigorously evaluate the effect).

 

Why this is exciting

• Impactful topic: Uncertainty estimation is a fast-growing, high-impact area in autonomous driving.
• Real-world relevance: We already use aleatoric uncertainty in our research vehicle; your work extends this to cutting-edge transformer models.
• Publishable potential: Strong results can lead to a publication.
• Supportive environment: Access to prior codebases for uncertainty estimation and mentoring from an experienced research team.

Voraussetzungen

What you bring

• Curiosity for autonomous driving and machine learning.
• Hands-on skills with Python and PyTorch (experience helps).
• Drive and ownership to work independently and push ideas from concept to evaluation.
• attitude

Tags

FTM Studienarbeit, FTM AV, FTM Schroeder, FTM Informatik

Möglicher Beginn

sofort

Kontakt

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

Ausschreibung