Multi-view tracking of soccer players with dynamic cameras

Abstract

Challenges such as player occlusion, fast player motion, small size of players relative to the background make it difficult to track soccer players accurately and consistently throughout a game. To solve these challenges, in this work we present a multi-view approach to tracking soccer players. Here, we formulate tracking as the problem of assigning a label to each pixel in every frame of each camera view, where the label is either the background or one of the players. As a preprocessing step, we utilize the information from the soccer field for camera trajectory estimation and background modeling. Tracking is first carried out independently for each camera view with a layered tracker. Then we integrate the results of layered trackers from multiple views through MCMC inference over tracklet-to- player association. We show that through camera calibration, common background and shared states of the players, inference across multiple camera views significantly alleviates the problem of player occlusion and loss of tracks in some view. As a result, we are able to produce accurate and long tracks for players, enabling further analysis of the game.