TL;DR
This paper introduces a novel method to reconstruct continuous player tracking data in football matches from sparse, discrete broadcast information. By leveraging an autoregressive model and a clever trajectory assignment strategy, they can estimate player positions even when they are off-camera, making high-fidelity tracking data more accessible to a wider range of clubs.
What Problem Does This Paper Address?
High-quality player tracking data, which provides the precise location of every player on the pitch at high frame rates, is incredibly valuable for football analytics. However, generating this data is currently a bottleneck for many professional teams. The gold standard for tracking data requires specialized camera systems and sophisticated computer vision, which are often too expensive or the broadcast footage is not high enough quality for automated analysis. While "event data" (recording key actions like passes and shots) is more accessible, it lacks the crucial positional context. This paper aims to bridge this gap by creating continuous tracking data from the more readily available, but sparse, discrete data derived from broadcast footage.
Methods
The authors propose a two-stage approach. First, they build a predictive model for individual player positions using an autoregressive moving average with exogenous inputs (ARMAX) model. This model uses a player's past positions and the ball's position as inputs to predict future locations. Crucially, they train this model on a limited amount of open-source tracking data, making it accessible.
The second stage involves assigning these predictions to actual player trajectories and then interpolating between observed and predicted points to create continuous paths. They use a Hungarian algorithm for trajectory assignment, ensuring that predicted positions are matched to the most likely player. For interpolation, they employ a novel continuous path interpolation method that considers estimated player velocities, aiming for more realistic movement patterns. A key tunable parameter, $\alpha$, controls the weight given to velocity adjustments in smoothing trajectories.
Datasets
The primary dataset used for training and testing is the Metrica Sports open dataset. They specifically use Match 1 for testing and Match 2 for training their ARMAX model. They also mention adapting their code to process Statsbomb's "360 data," which provides discrete broadcast data for a large set of over 200 games, creating a new benchmark dataset for this type of input.
Key Insights and Ideas
- Bridging the Data Gap: The core idea is that even sparse, discrete broadcast data can be leveraged to reconstruct rich, continuous tracking data, democratizing access to this valuable information.
- ARMAX for Prediction: Using an ARMAX model with ball position as an exogenous variable is an effective way to predict player movement based on past observations.
- Trajectory Assignment is Key: The Hungarian algorithm plays a vital role in correctly assigning predicted positions to individual player trajectories, preventing "swaps" and maintaining team structure.
- Velocity Smoothing Matters: The continuous path interpolation method, which incorporates velocity adjustments, significantly improves accuracy compared to simpler methods like static position assumptions or ignoring velocity.
- Off-Camera Prediction is Feasible: The model demonstrates a surprising ability to predict player positions accurately even when they are not visible in the frame for extended periods.
Key Results
The model achieves a mean error of 7.33 m for off-camera players on their test dataset. This error is reduced to 6.97 m for frames close to recorded events. When compared to benchmarks, their model shows a substantial improvement of over 70%. The trajectory smoothing component (using $\alpha \approx 0.6$) is highlighted as crucial for this performance. The authors also note that their model's error is comparable to or better than existing methods that use similar discrete input data.
Limitations and Future Work
The authors acknowledge that the model is not designed to track individual player trajectories perfectly, and events like corners can lead to "swaps" in assignments. They suggest that incorporating prior information about team formations could help. The model also struggles after long stoppages in play, where predictable structure is lost. Future work could involve using ensemble predictions from forward and reverse models to address this. While the error level is sufficient for analyzing team formations, it might be too high for detailed analysis of individual player positioning or off-camera runs. The authors also suggest that incorporating features from other models could further improve predictions for players who have been off-camera for extended periods.
Summary updated