Data Synthesis for Autonomous Driving Perception

Institut

Lehrstuhl für Fahrzeugtechnik

Typ

Semesterarbeit / Masterarbeit /

Inhalt

experimentell / konstruktiv /  

Beschreibung

Deep-Learning based perception algorithms are commonly used in autonomous driving to detect and track obstacles on the road. To train these neural networks large quantities of training data are required for good results, however the process of labeling the training data is error-prone, time-consuming and expensive. In addition to this, recording real-world data is not always possible if recording with the testing vehicle is too expensive and not all scenarios can be evaluated in the real-world.

There are multiple approaches to solving this problem:

• Recording of training data with comprehensive simulation of sensors and scenarios
• Synthesis of training data using neural networks
• Augmentation of existing real-world data to create new scenarios
• Self-supervised learning using automatic labeling processes

The goal of the thesis:

• Literature research on solutions for data annotation
• Evaluating different approaches and choosing one of the possible paths
• Design and implemention of algorithms for data generation
• Evaluation of detections algorithms with generated training data and real-world validation data
• Document your process and code

The overall goal is the development of a new approach to generate large amounts of training data for the TUM Autonomous Motorsport team to accelerate development of detections algorithms and improve their performance through better data.

Voraussetzungen

• Good knowledge of Python and/or C++
• Basic knowledge of Unity(Game Engine)/C# (or experience with Unreal/C++, Carla)
• Experience with ROS 2 advantageous

If you are interested in this project, please send your resume and transcript to:

dominic.ebnertum.de

Tags

FTM Studienarbeit, FTM AV, FTM Informatik, FTM Ebner

Möglicher Beginn

sofort

Kontakt

Dominic Ebner, M.Sc.
Raum: MW3510
Tel.: +49.89.289.15871
dominic.ebnertum.de