Supervisors: Jan-Nico Zaech, Dr. Ajad Chhatkuli, Prof. Dr. Luc van Gool
The RoboCup Standard Platform League is a tournament in which autonomous robots compete in soccer matches. Among the numerous tasks of the robot, image segmentation and object detection is of great importance for understanding the robot's environment. In this project, we propose a deep learning based, two stage approach that is capable of running on the limited hardware available in embedded systems and annotate an improved version of the ETH Robocup Team’s dataset. First in the segmentation pipeline, image regions potentially containing objects of interest are extracted using an algorithm based on the Hough transform. Subsequently, the proposed regions are classified with a convolutional neural network. We show that the optimized network architecture outperforms previous random forest based methods while keeping computational time at the same magnitude, thus meeting the real-time constraint on the given NAO robot platform.