Fatigue monitoring is central in association football due to its links with injury risk and tactical performance. However, objective fatigue-related indicators are commonly derived from subjective self-reported metrics, biomarkers derived from laboratory tests, or, more recently, intrusive sensors such as heart monitors or GPS tracking data. This paper studies whether monocular broadcast videos can provide spatio-temporal signals of sufficient quality to support fatigue-oriented analysis. Building on state-of-the-art Game State Reconstruction methods, we extract player trajectories in pitch coordinates and propose a novel kinematics processing algorithm to obtain temporally consistent speed and acceleration estimates from reconstructed tracks. We then construct acceleration--speed (A-S) profiles from these signals and analyze their behavior as fatigue-related performance indicators. We evaluate the full pipeline on the public SoccerNet-GSR benchmark, considering both 30-second clips and a complete 45-minute half to examine short-term reliability and longer-term temporal consistency. Our results indicate that monocular GSR can recover kinematic patterns that are compatible with A-S profiling while also revealing sensitivity to trajectory noise, calibration errors, and temporal discontinuities inherent to broadcast footage. These findings support monocular broadcast video as a low-cost basis for fatigue analysis and delineate the methodological challenges for future research.
Figures & tables
Figure 1: Overview of the proposed method for athlete fatigue assessment from monocular broadcast football videos. Left: Game state reconstruction enables the recovery of metric player trajectories by projecting detections into pitch coordinates, resulting in time-evolving (x,y) positional signals for each athlete. Right (top): From these trajectories, temporally consistent speed and acceleration signals are estimated in the metric coordinate system. Right (bottom): Acceleration–Speed (A–S) profiles are constructed by modeling the relationship between instantaneous acceleration and running speed. The slope and intercept parameters of the A–S diagram are analyzed as fatigue-related kinematic indicators, allowing the study of their evolution over match time.
Figure 2: Qualitative examples. Qualitative examples of predicted athlete trajectories (red) and their respective ground-truth trajectories (blue) on the SoccerNet Game State Reconstruction (GSR) test set. (a) Depicts the trajectory of player 17 in the SNGS-200 sequence, and (b) portrays the one of player 3 in the SNGS-199 sequence. As observed, although the predicted trajectories exhibit minor localization noise, they closely follow the ground-truth trajectories.
ln
Filter
Speed
Acceleration
MAE ↓
RMSE ↓
r ↑
max-lag r ↑
MAE ↓
RMSE ↓
r ↑
max-lag r ↑
None
0.776
1.364
0.732
0.767
1.857
2.548
0.301
0.370
5
Kalman
0.761
1.350
0.739
0.773
1.788
2.470
0.320
0.390
Savgol
0.768
1.389
0.724
0.761
1.758
2.439
0.327
0.398
None
0.610
1.196
0.814
0.841
0.934
1.493
0.523
0.580
10
Kalman
0.607
1.192
0.816
0.842
0.926
1.486
0.531
0.589
Table 1: Ablation study on the effect of the filtering method on predicted athlete speed and acceleration. Results are reported on the test set of the SoccerNet Game State Reconstruction (GSR) benchmark. We show MAE, RMSE, Pearson correlation ( r ), and max-lag r at 25 fps. Arrows indicate whether higher values correspond to better (↑) or worse (↓) performance.
ln
Temporal
Speed
Acceleration
window (s)
MAE ↓
RMSE ↓
r ↑
max-lag r ↑
MAE ↓
RMSE ↓
r ↑
max-lag r ↑
2
0.20
1.067 ± 1.361
1.729 ± 1.361
0.616 ± 0.380
0.652 ± 0.373
3.442 ± 2.358
4.172 ± 2.358
0.094 ± 0.150
0.195 ± 0.131
5
0.44
0.761 ± 1.115
1.350 ± 1.115
0.739 ± 0.495
0.773 ± 0.481
1.788 ± 1.704
2.470 ± 1.704
0.320 ± 0.233
0.390 ± 0.219
10
0.80
0.607 ± 1.026
1.192 ± 1.026
0.816 ± 0.613
0.842 ± 0.581
0.926 ± 1.163
1.486 ± 1.163
0.531 ± 0.389
0.589 ± 0.384
15
1.24
0.529 ± 1.000
1.131 ± 1.000
0.858 ± 0.696
0.876 ± 0.665
0.601 ± 0.920
1.099 ± 0.920
0.646 ± 0.511
0.687 ± 0.501
20
1.64
0.485 ± 0.991
1.103 ± 0.991
0.879 ± 0.771
0.896 ± 0.739
0.451 ± 0.815
0.932 ± 0.815
0.716 ± 0.597
0.749 ± 0.586
Table 2: Ablation study on the effect of the local neighborhood ln on predicted athlete speed and acceleration. Results are reported on the test set of the SoccerNet GSR benchmark. We report MAE, RMSE, Pearson correlation ( r ), and max-lag r at 25 fps. Results are shown as mean ± standard deviation. Arrows indicate whether higher values correspond to better (↑) or worse (↓) performance.
Figure 3: Qualitative examples of reliable and non-reliable speed and acceleration predictions. The first two columns depict two athletes with high reliability scores showing strong agreement between predictions and ground truth for both speed and acceleration. The last column shows a low-reliability case where the athlete was not detected for most of the sequence, leading to large error.
Metric
Top 30% Reliable athletes
Bottom 30% Reliable athletes
Avg. visible frames
537 (21.48 s)
351 (14.04 s)
Speed
Pearson r
0.975 ± 0.021
0.489 ± 0.327
MAE
0.234 ± 0.055
1.078 ± 0.751
RMSE
0.132 ± 0.073
4.545 ± 5.554
Acceleration
Table 3: Reliability analysis of athlete motion dynamics prediction. on the test set of the SoccerNet Game State Reconstruction (GSR) benchmark. Values are reported as mean ± standard deviation. Max-lag r is not defined for some low-reliability athletes.
Figure 4: Comparison between predicted and ground truth A–S profiles for sequence SNGS-199, team left, jersey 3. As can be seen, predicted A–S curve follows the ground truth signal, despite clear fluctuations due to noise and accumulated errors. Green points are the ones selected to regress the linear equation.
Figure 5: Comparison of early (left) and late (right) acceleration–speed (A–S) profiles over a 45-minute broadcast sequence. The early profile shows a gradual decrease in acceleration as speed increases, whereas the late profile indicates a substantially faster decline.
Soccer analytics draws on two kinds of information: spatio-temporal data describing where players and the ball are, and event data describing what they do. Public datasets offer them apart, or together only on broadcast footage that leaves players outside the frame unobserved. SoccerTrack v2 combines continuous full-pitch video, long player trajectories and actor-linked events in one resource: ten university-level matches, 932 minutes of fixed-camera 4K panoramic video, annotated per frame with metric pitch coordinates, jersey numbers and persistent identities, roles and team sides for all players, and with ball action events in twelve classes, linked to the acting players through the same identifiers used in the trajectories. We fix a match-level split and report baselines for two tasks. For game state reconstruction, we run a full pipeline over all twenty halves and find that GS-HOTA scores degrade as sequence length increases. For ball action spotting, we train a model on the player trajectories, with and without the ball track. The data, the split and the evaluation tooling are released so that both tasks can be developed and compared at match length on the same footage.
Minute-resolution athlete monitoring is increasingly common, while injury annotation may exist only at the athlete-session level and omit within-session onset time. Replicating a positive session label across every recorded minute would therefore create unsupported minute-level supervision. We address this label-resolution mismatch using fixed elapsed-time landmarks at 10, 20, 30, 40, 50, and 60 min, constructing one representation per athlete-session from information available up to each landmark while keeping the target as a same-day injury-associated session indicator. Using 2020 SoccerMon data from elite women's football, the modelling cohort contains 2,259 Team A athlete-sessions from 27 athletes, including all 22 positive sessions from five athletes. Evaluation is athlete-disjoint. We compare contextual, cumulative, and dynamic representations; Logistic Regression, Random Forest, XGBoost, and TabPFN; and training-only NONE, SMOTE, and CTGAN conditions. Robustness is assessed using athlete-cluster bootstrap, a fixed common cohort, alternative fold allocations, leave-one-positive-athlete-out analysis, and equal-athlete weighting. Discrimination is landmark-dependent and non-monotonic. TabPFN improves later-landmark discrimination relative to Logistic Regression but does not consistently outperform Random Forest. Synthetic augmentation provides condition-specific rather than universal benefit. The contribution is a unit-aligned framework for using minute-resolution predictors with session-level supervision, not minute-specific injury prediction.
Evangelos Chatzidimitriou, Konstantinos Tserpes
National Technical University of Athens, Athens, Greece
Spatial football metrics such as pitch control assume access to the positions of all 22 players, yet the most widely available source of positional data -- the broadcast main camera -- shows only 10-16 of them at any moment. We quantify the resulting distortion with an open, reproducible benchmark: a simulated broadcast viewport applied to open full-pitch tracking data (Metrica Sports; three matches, one held out from method development). Ignoring off-screen players -- the visible-only baseline implied whenever a video-based game-state-reconstruction (GSR) pipeline adds no imputation layer -- inflates hidden-zone pitch-control error to 25.1-26.9 percentage points and a mean absolute control-share error of 11.1-13.4 points across the three matches. We then evaluate a ladder of training-free, online imputation baselines that use only observations from the match being analysed. The best overall on these decision-relevant metrics, role-anchored centroid voting (each visible player votes for the full-team centroid by subtracting its running role offset, attenuating the viewport-induced subset bias), roughly halves hidden-zone error (to 12.2-13.8 points) and cuts control-share error to 28-48% of the ignore policy at every viewport width from 36 m to 60 m in all three matches. For occlusions <=9.6 s -- the regime of the closest learned prior work -- it reaches binwise median position errors of 3.3-8.9 m; but 50-57% of hidden-player observations lie beyond that regime. Integrated end-to-end into a broadcast-video GSR pipeline, imputation moves a downstream possession-quality score (Space-Creation Index) by 15.6 and 17.2 points on two real World Cup broadcast windows, flipping the verdict class in one.