Ball Trajectory Estimation on Wheeled Goalkeeper Robot Based on Robot Operating System (ROS2) Using Stereo Vision

Abstract

Wheeled goalkeeper robots require a perception system capable of rapidly detecting and estimating the direction of ball movement under dynamic match conditions. This study aims to implement a ball trajectory estimation system using Stereo Vision within the Robot Operating System (ROS2) architecture. Camera images are processed using You Only Look Once (YOLOv5) to detect the ball and obtain its center coordinates on the image plane, while depth data is used to acquire the ball's distance relative to the camera. The ball position data is then processed using a Kalman Filter to generate state estimations consisting of position and position changes on the image plane. A Constant Velocity model is subsequently used to predict the next position and determine the directional trend of the ball's trajectory. The experimental results show that YOLOv5 is capable of detecting the ball across all test distance variations from 100 to 800 centimeters with a 100 percent success rate. Depth data acquisition yields an average error of 1.37 percent. The integration results of trajectory estimation and prediction demonstrate directional alignment in 9 out of 10 tests, achieving a 90 percent success rate. These results indicate that the integration of Stereo Vision, Kalman Filter, and the Constant Velocity model on ROS2 is capable of generating ball trajectory estimations and predictions to support the perception system of wheeled goalkeeper robots.