Prediction of Ball Position Using CNN Methods With Zed Camera on Goalkeeper Robot Application

摘要

EEPIS Robot Soccer On Wheeled (ERSOW) is a wheeled soccer robot team competing in the Indonesian Wheeled Soccer Robot Contest. The ERSOW goalkeeper robot currently faces a challenge in responding to incoming balls due to delays in image processing, leading to slow movement responses. This delay can be mitigated by implementing a prediction system. By predicting the ball’s trajectory, the robot can make timely movements to block shots effectively. This research designs a prediction system using the ZED stereo camera to enhance the goalkeeper robot’s responsiveness in preventing goals. The research identifies the ball’s position during airborne and ground states using the ZED camera. RGB-D readings are processed with Deep Learning methods to train on ball data and extract 3D object information from the dataset. The ball detection system is tested in three scenarios: right, center, and left positions relative to the robot. Accuracy evaluation is conducted three times for ball coordinates on the X, Y, and Z axes, resulting in X-, Y-, and Z-axis accuracy values of 90%, 80.77%, and 100%, respectively, with a <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\le 5$ </tex-math></inline-formula> cm error tolerance. These results allow accurate determination of the ball’s X, Y, and Z positions. The ball position data is then trained using Convolutional Neural Networks (CNN) to predict the ball’s arrival position for the goalkeeper robot. Evaluation results show an overall accuracy of 70%, with an F1-score of 76% for the left position, 39% for the center position, and 84% for the right position.