Automated ball tracking in tennis videos

Sports video analysis tools are gaining enormous popularity as they enable enhanced visualization and analysis of the game. An intriguing problem is localizing and tracking the ball in a game of tennis. However, achieving a practically suitable tradeoff between high detection accuracy and speed of the ball is a challenging problem in automatic ball tracking algorithms. In this paper, we propose a machine learning based automated ball tracking algorithm in tennis match videos acquired by a camera mounted on a quadcopter. We begin with applying a video stabilization technique followed by random forest segmentation for detecting tennis ball candidates in the video frames. The yellow color plane intensity and Phase Quaternion Fourier Transform (PQFT) saliency are used as features for random forest segmentation. The effectiveness of the proposed algorithm is demonstrated on datasets from real-world tennis videos. The experimental results show that our algorithm is robust and achieves 94% accuracy in best case and successfully detects the ball even in cases of partial occlusion by net or racket.