Ball Detection System for a Soccer on Wheeled Robot Using the MobileNetV2 SSD Method

Abstract

This paper discusses the research on the use of Artificial Intelligence in autonomous robot object identification. The specific focus of this research is on a wheeled soccer playing robot. The goal is to recognize a ball as an object using the Single Shot MultiBox Detector MobileNetV2 model. This system has multi-vision inputs such as distance measurements and angle values ​​for object detection. This methodology is based on deep learning with the TensorFlow Object Detection API with the MobileNetV2 SSD model. This model is trained with a dataset of 3707 ball images over 617 thousand steps on Google Collaboratory. It was found that the average measurement error of the ball object is 6.58% for the distance when viewed through the robot's front camera. In addition, the omnidirectional camera is able to detect the ball object and angle values ​​from the front of the robot. What makes this research different is the use of distance and angle measurements for detection and the omnidirectional camera for system performance in dynamic environments. This research aims to address the improvement of AI-based object detection systems for autonomous robotics in the context of real-world use cases.