Mining and Training Pipeline for Tail Light and Turn Signal Detection in Autonomous Driving

Institut

Lehrstuhl für Fahrzeugtechnik (TUM-ED)

Typ

Semesterarbeit / Masterarbeit /

Inhalt

 

Beschreibung

The aim of this project is to develop a scenario mining pipeline to extract interesting scenes for tail light detection, and to integrate and evaluate it using a state-of-the-art detection approach. 

MOTIVATION

Brake lights and turn indicators are critical communication signals for predicting the future intent of traffic participants. While proprietary autonomous driving stacks heavily rely on these cues, open-source alternatives like Autoware.universe currently lack the robust capability to detect and interpret them.

While training a basic 2D detector is straightforward, the primary bottleneck is scenario mining—finding rare, high-quality data candidates showing active signaling behavior within massive datasets. This thesis focuses on building an automated data mining and integrating it into an existing active learning pipeline to iteratively maximize detector performance, utilizing both massive public datasets and real-world data collected by our EDGAR research platform.

Voraussetzungen

YOUR ROLE

The work packages of this thesis contain:

• Scenario Mining Pipeline: Design and implement a pipeline to extract interesting data candidates (vehicles signaling, braking, changing lanes) from large public datasets (e.g., NuPlan, Zensact, Truckscenes, EDGARScenes).
• EDGAR Data Ingestion: Integrate raw sensor data collected from the chair's EDGAR research vehicle into the mining framework.
• Active Learning Loop: Develop an active learning strategy to select the most informative samples for retraining, minimizing manual labeling effort.
• Model Training & Benchmarking: Implement and benchmark state-of-the-art 2D object detectors (e.g., using MMDetection/PyTorch) tailored to tail light state classification.
• Ablation & Performance Analysis: Systematically analyze how the iteratively mined data affects the model's precision, recall, and overall robustness.

WHAT YOU SHOULD BRING ALONG

• High motivation and strong interest in autonomous driving and deep learning pipelines.
• Solid programming skills in Python.
• Hands-on experience with PyTorch; familiarity with MMDetection or similar object detection frameworks is a major plus.
• Structured, independent working style and excellent team player skills.
 

Joint Work Note: This project is highly collaborative and will be executed in close cooperation with Efficient ROS 2 Integration and Planning Logic for Vehicle Intent Detection, who will port your trained models into a real-time ROS 2 node running directly on the EDGAR vehicle.
 

If you are interested in joining this project, feel free to submit your application along with your CV and transcript of records. We look forward to receiving your application!
 

This work is also available as IDP and can be done in English or German! 

EMail: loic.stratil<script>document.write('@');</script>

<noscript>(at)</noscript>tum.de, niklas.krauss<script>document.write('@');</script><noscript>(at)</noscript>tum.de 

 

Verwendete Technologien

PyTorch, Python, Detection, Active Learning, TUM EDGAR, Autonomous Driving,

Tags

FTM Studienarbeit, FTM Krauss, FTM AV, FTM AV Safe Operation, FTM AV Perception, FTM Informatik, FTM Detection

Möglicher Beginn

sofort

Kontakt

Niklas Krauß
Raum: 3507
Tel.: +49172 1736882
niklas.krausstum.de