Real-Time Object Detection On Humanoid Robots For The RoboCup Soccer SPL Using Cascaded Classifiers

摘要

As robots autonomously compete to win soccer RoboCup competition annually, more challenging rules are introduced to prepare robot teams for the ultimate goal of winning against human soccer world champions by the year 2050. Recent rules in the RoboCup Standard Platform League (SPL) cleared all color-coded field information such as replacing the red ball with the classic black and white soccer ball and replacing the colored goal posts with white posts which made the field perfectly symmetric around the mid-field. In this thesis, we try to improve the robots vision using rapid object detection architecture developed by Viola-Jones. The algorithm was used to detect the new soccer ball, to detect other robots, to detect external field features and to detect robots’ feet. The detectors’ performance was then evaluated based on accuracy, speed and detection range. The detectors’ quality was tested using 10-fold cross validation. Each object detector demonstrated different performance which is due to the different natural textures of the objects and the variation parameters used to optimize the detector. The University of Miami poster detector showed the best results in terms of accuracy and detection range while the NAO robot feet detection was the fastest.