CNN-Based Self Localization Using Visual Modelling of a Gyrocompass Line Mark and Omni-Vision Image for a Wheeled Soccer Robot Application

Abstract

The Convolutional Neural Network (CNN) is an object classification method that has been widely used in recent research. In this paper, we propose CNN for use in the self-localization of wheeled soccer robots on a soccer field. If the soccer field is divided into equally sized quadrants with imaginary vertical and horizontal lines intersecting in the middle of the field, then the soccer field has an identical shape for each quadrant. Every quadrant is a reflection of the other quadrants. Superficially similar images appearing in different positions may result in positioning mistakes. This paper proposes a solution to this problem by using a visual modelling of the gyrocompass line mark and omnivision image for the CNN-based self-localization system. A gyrocompass is used to obtain the angle of the robot on the soccer field. A 360° omni-vision camera is used to capture images that cover all parts of the soccer field wherever the robot is located. The angle of the robot is added to the omni-vision image using the visual modelling method. The implementation of self-localization without visual modelling gives accuracy rates of 0.3262, and this result is increased to 0.6827 with the proposed methods. The experiment was carried out in the robotics laboratory of the Institut Teknologi Sepuluh Nopember (ITS) with the ITS Robot with Intelligent System (IRIS) robot.