Continuous football player tracking from discrete broadcast data

Matthew J. Penn, Christl A. Donnelly, Samir Bhatt

Abstract

Football analytics is currently split between widely available event data and expensive, high-fidelity tracking data. This paper introduces a parsimonious method to bridge this gap by estimating continuous, full-pitch player trajectories from discrete, incomplete broadcast data. The approach utilizes an ARMAX forecasting model integrated with a novel velocity-corrected interpolation technique. Evaluation against ground-truth tracking data demonstrates a mean error of 7.33 meters for off-camera players, while successfully preserving the team's tactical structure. The system is computationally efficient, enabling the generation of tracking-level insights for large-scale broadcast datasets.

Introduction

In football analytics, a "data divide" exists between event data—manual logs of actions like passes and shots—and tracking data, which provides high-frequency coordinates for all players. While tracking data is the "gold standard," it typically requires specialized camera systems. Broadcast footage is more accessible but suffers from the "off-camera" problem: players frequently disappear from view as the camera follows the ball. This research addresses how to accurately estimate player positions during these occlusions to generate smooth, continuous trajectories.

Methodology

The proposed pipeline consists of three primary stages:

  1. Player Position Forecasting: An ARMAX (Autoregressive Moving Average with Exogenous inputs) model is employed. The ball's position serves as the exogenous variable, leveraging the high correlation between player movement and ball location in football.
  2. Trajectory Assignment: The Hungarian algorithm is used to match observed player positions in new frames to predicted trajectories, maximizing the log-likelihood of assignments and maintaining identity consistency.
  3. Continuous Path Interpolation: To avoid the unrealistic appearance of linear interpolation, a velocity correction is applied. The model calculates the average velocity of visible players to "nudge" the estimated paths of off-camera players, ensuring their movement aligns with the overall team dynamic (e.g., a collective sprint during a counter-attack).

Key Results

The model was validated using the Metrica Sports Open Dataset. For players who are "off-camera," the system achieves a median error of 5.37 meters and a mean error of 7.33 meters.

Distribution of tracking errors (m) for off-camera players.

A critical finding is the model's temporal robustness. Even when players are off-camera for extended periods, the error remains manageable. For instance, after 20 seconds of occlusion, the mean error stays below 10 meters.

Error growth over time as a function of seconds from the nearest observation.

Beyond individual accuracy, the model excels at preserving the team's tactical structure. Even with several meters of individual error, the defensive line and midfield width remain recognizable, making the data viable for high-level tactical analysis. Computationally, the system is highly efficient, processing a standard 90-minute match in 20–25 minutes on a standard laptop.

Discussion and Limitations

While the method significantly improves upon event-only models (reducing mean error from ~7.86m to 3.45m overall), several limitations remain. The current model tracks team members rather than specific individuals, leading to potential identity swaps during chaotic events like corner kicks. Additionally, the error margin precludes high-precision applications such as automated offside detection. Future work involves training position-specific models (e.g., separate models for defenders and strikers) on larger datasets to further refine accuracy.

Summary updated

BibTeX
@inproceedings{penn2025continuous,
  author = {Matthew J. Penn and Christl A. Donnelly and Samir Bhatt},
  title = {{Continuous football player tracking from discrete broadcast data}},
  booktitle = {Royal},
  year = {2025},
  doi = {10.1098/rsos.251175},
}
Figure