Improving Field and Ball Detector for Humanoid Robot Soccer EROS Platform

Humanoid robot soccer perceives environment mostly through cameras. The performance decrement in our humanoid soccer platform (EROS) is primarily due to the visual perception that is less robust to the RoboCup new rule which specifically reducing color coding in the field. Notable works favorably employ simple color segmentation, image morphology, and blob detector due to simplicity in the implementation and run in real-time for most embedded hardware, while some employ a more advanced supervised learning running in sophisticated hardware to boost detection accuracy. In this paper, a visual perception system consisting of field and ball detection is developed in our platform EROS to address the RoboCup new rule. Color segmentation and image morphology are stacked with a more advanced supervised learning cascade classifier. In this way, the favorable color segmentation and image morphology help to reduce the number of object candidates while the cascade classifier helps to boost the accuracy of detection. Experiments show encouraging result for detecting field and ball position. Our approach has successfully been implemented in practice and achieves remarkably result in Indonesian humanoid robot soccer competition.