Supervisors: Jan-Nico Zaech, Dr. Ajad Chhatkuli and Prof. Luc Van Gool
RoboCup Standard Platform League is an international competition where teams compete against each other, by letting Nao robots play 5 versus 5 soccer matches. This thesis addresses the challenge of understanding of the environment around the robot by creating a dataset and using it to train a robot detection and pose estimation network. Thereby, it addresses key challenges of the RoboCup competition. Based on the perception, decisions on where to move and higher-level strategies can be made. Information about the positions of other players is used to ensure a more fluent game play and to avoid collisions. A random forest network is currently used in the NomdaZ team for player detection. The latest Nao version 6 has significantly more computational power and will be able to run a small and efficient convolutional neural network (CNN) which could enable a more robust player detection. To train such a CNN, a good dataset is needed; hence a dataset was created which covers real game situations and different lighting condition. It consists of roughly 1’100 images with COCO style bounding box, segmentation mask and keyoint annotations. Tools to extend the dataset either by hand or by outsourcing it to Amazon Mechanical Turk were adapted. The dataset was used to fine tune a Mask R-CNN Human-pose network pretrained on the COCO keypoint dataset. The network detects robots with an average precision score of 93.3% and average precision score of 94.4%. The network is too slow to run on the Nao robots, but will be used at the RoboCup 2019 in Sydney to label ground truth data for random forest training. Additionaly keypoints are detected and could be used in future work to obtain an opponent’s location and its orientation, or even to create a three dimensional model of the detected robots pose. All things considered; this thesis lays the base for future work to develop a stable robot detection via a CNN that is running on the Nao robots.