Prediction of the Ball Location on the 2D Plane in Football Using Optical Tracking Data

Abstract

Tracking the ball location is essential for automated game analysis in complex ball-centered team sports such as football. However, it has always been a challenge for image processing-based techniques because the players and other factors often occlude the view of the ball. This study proposes an automated machine learning-based method for predicting the ball location from players' behavior on the pitch. The model has been built by processing spatial information of players acquired from optical tracking data. Optical tracking data include samples from 300 matches of the 2017-2018 season of the Turkish Football Federation's Super League. We use neural networks to predict the ball location in 2D axes. The average coefficient of determination of the ball tracking model on the test set both for the x-axis and the y-axis is accordingly 79% and 92%, where the mean absolute error is 7.56 meters for the x-axis and 5.01 meters for the y-axis.

Ball Trajectory Inference from Multi-Agent Sports Contexts Using Set Transformer and Hierarchical Bi-LSTM

“…Though ball tracking from sports videos is a topic of interest in computer vision, only a few studies [1,23] tried to estimate ball trajectories not relying on optical tracking but only using players' movement data. Amirli et al [1] aggregated players' locations and speeds to make handcrafted input features and constructed a neural network regressor to estimate ball locations. However, they did not employ sophisticated architectures to encode the sequential nature or permutation-invariance of multi-agent trajectories.…”

Player Tracking Data in Sports

“…At the same time that some researchers used tracking data to enrich prediction models of traditional outcomes, other researchers turned their focus to the prediction of novel spatial outcomes. Several studies investigated the prediction of 2D locations in sports, including ball location in tennis (Wei et al 2016) and soccer (Amirli & Alemdar 2022) as well as rebound locations in basketball (Masheswaran et al 2014). Full spatial trajectories have also garnered increasing interest for sports prediction models; thus far, neural network approaches have been the most commonly used methodology.…”