Multiple soccer players tracking

Abstract

This paper, describes a solution for tracking multiple soccer players, simultaneously, in soccer ground. It adapts Kalman filter for tracking multiple players. Adapting Kalman filter is divided to four main tasks. The first task is defining the state vector for multiple object tracking. The second task is determining a motion model for estimating the position of soccer players in the next frame. The third task is defining an observation method for detecting soccer players in each frame. Finally, the fourth task is tuning the measurement noise covariance and estimating noise covariance. In the third task, a novel observation method for detecting soccer players is proposed. This method divides the player body into three parts and calculates the histogram of each part, separately. Also, an algorithm for updating the reference object patch is introduced in observation method. Each task is discussed in detail and the promising performance of the proposed method for tracking soccer players when run on the Azadi dataset is shown.