Real-Time Ball Position Measurement for Football Games Based on Ball's Appearance and Motion Features

Abstract

A novel ball tracking system that can be used for actual football games is developed. It measures the positions of a ball from multiple video sequences in real time. The positions can be used for calculating various statistical data and making visual effects for a broadcast program. Although a number of methods for visual object tracking have been proposed, tracking a small ball is still a challenging task. We focused on the motion features of a ball object, adding to its appearance features. We created two discriminate models for tracking a ball in the manner of supervised machine learning. The appearance classifier was trained with visual features of ball images, and the motion classifier was trained with ball trajectories. Tracking results from multiple cameras are integrated in consideration of their reliabilities, and the final ball position is robustly measured. The system was evaluated with two professional football games. The results showed that the proposed system is effective for practical use at football games.