Research Sports Video Analysis

Automatic Match Analysis in Sports Videos Using Bayesian Network and Random Forest

There are many automatic match analysis methods that use video information. Our goal is to improve the precision of previous match analysis methods. We address three match analysis tasks: playtime separation, play classification, and ball trajectory estimation. We focus on American football videos. First, we separate playtime using the strength of players’ motions and Bayesian Network. Second, we classify plays using the information about each team player’s position, global motion features and Random Forest. To improve the classification precision, we calculate the play start and end positions, which involve many heavily occluded scenes. This is one of the main contributions of our method. Third, we estimate ball trajectories using the results of the playtime separation and play classification. Our method can estimate the trajectory without detecting the ball.