Die nachfolgende Aufgabenstellung ist in englischer Sprache verfasst. Die Bearbeitung der Themen kann, je nach Wunsch des Studenten, wahlweise in Englisch oder Deutsch erfolgen.

Description

One of the oldest and still major challenges of computer vision is to enable segmentation and classification of images into meaningful parts, i.e. assigning labels to different regions in the image. The motivation is to allow unsupervised systems to learn to see has an many applications in robotics, computer applications, manufacturing processes, etc.

The goal of this thesis is the development of a framework that combines possibly multimodal data from visual and laser data to segment and classify urban scenes into meaningful parts. The system should enable an urban robot to safely navigate on the sidewalk as well as to perceive other significant regions in the scene, such as buildings, road, sky, etc. Emphasis will be put on making the algorithms usable for online application.

Tasks

• Literature review of state of the art image segmentation and classification.

• Set up a stereo camera system and tilting laser for joint calibration. Data recording.
• Feature extraction from multiple data sensors and integration in the ROS framework .
• Implementation of a probabilistic model the classification of regions in the scene.
• Experimental evaluation on local urban scenes and against labeled image databases.

Literature

• A generative framework for fast urban labeling using spatial and temporal context. Ingmar Postner et al.
• Multiscale conditional random fields for image labeling - He, X., Zemel et al.
• Segmentation-based urban traffic scene understanding - Ess, A.,Muller, T. et al.

Requirements for students

• Programming experience C++/Matlab.

• Considerable reading and understanding of machine/statistical learning is involved. If you want to do this thesis, you should like this. Not for math-allergic students.

 

Are you interested? Please contact:

Dipl-Ing. Roderick de Nijs

rsdenijs@lsr.ei.tum.de