Efficient Objects Identification for a KidSize Humanoid Soccer Robot: Using Feedforward Neural Networks with Image Segmentation

Abstract

One of the biggest problems faced in robot soccer tournaments is related to object identification in the soccer field. The ball, field marks, crossbars of the goal, the opponent players or teammates are considered objects and should be identified by the robot soccer player in the category of KidSize RoboCup Humanoid League. The main target of this paper is to present a robotic architecture composed of the Software and Hardware Modules responsible for identifying these objects. The Software Module was composed of a DataBase (DB) with images of the objects to be identified and a Feedforward Neural Network (FNN) trained by the Neural Network Toolbox (NNT) of the MATrix LABoratory (MATLAB). In order to integrate the identification process with the Hardware Module, it was necessary to develop the NeuralNet library. This library was implemented in JAVA, making it compatible with multiple platforms. The purpose of this library was to transfer the FNN already trained by the NNT to the Raspberry Pi 3B. The Raspberry Pi 3B was responsible for processing the images captured by the Vision System of the robotic player. Also, all the objects in the field were identified through the trained FNN and the Open Source Computer Vision (OpenCV) library. The results showed an efficiency of about 82% in ball identification, 92% in field marks, 81% in crossbars of the goal, and 93% in opponent players or teammates.