Ball Distance Estimation Using Machine Learning on an Omnidirectional Camera for a Middle Size Soccer Robot

Abstract

The RoboCup Middle Size League (MSL) is a robotic soccer competition where fully autonomous robots play soccer using onboard perception and decision-making systems. Accurate ball perception, especially distance estimation, is crucial for effective navigation and control. This paper presents a ball distance estimation system for MSL robot soccer using an omnidirectional camera combined with machine learning models. Bounding box parameters from object detection, along with the width and height of the omnidirectional camera frame, are used as inputs for regression models to predict the ball's distance, which is then converted into local X and Y coordinates. Four regression models—K-Nearest Neighbor (KNN), Support Vector Regression (SVR), Random Forest Regression (RFR), and Multilayer Perceptron (MLP)—were evaluated in ground and airborne ball conditions. RFR and MLP achieved the highest accuracy, with MLP outperforming all other models based on experimental results. These findings demonstrate the feasibility of the proposed approach, with MLP emerging as the most reliable model for robust and accurate ball distance estimation in MSL robot soccer.