Source-linked AI summary
ELASTIC: Trajectory-Based Synchronization of Event and Tracking Data in Soccer
Hyunsung Kim, Hoyoung Choi, Kunhee Lee, Sangwoo Seo, Tom Boomstra, Jinsung Yoon, Chanyoung Park
TL;DR
Human-annotated event timestamps and locations can misalign soccer event and tracking data, while existing methods miss ball receptions. ELASTIC synchronizes event starts and ends from player and ball trajectories using candidate frames and extended sequence alignment, and its public-benchmark evaluation reports state-of-the-art synchronization with downstream gains.
Problem
Human annotations can misplace event timing and location, while existing synchronizers miss ball receptions needed for possession and destination-conditioned analyses.
Method
ELASTIC extracts physically plausible ball-touch candidate frames from trajectories, inserts virtual termination events, and applies extended Needleman–Wunsch alignment without annotated event locations.
Results
ELASTIC achieves the most accurate synchronization to date on a re-annotated public benchmark and improves downstream soccer analytics across evaluated tasks.
Takeaways & Limitations
Synchronization quality materially affects the reliability and performance of downstream soccer analytics, making alignment more than a preprocessing detail.
Takeaways & Limitations
Candidate selection caps accuracy: if no candidate lies within two frames of a true event timestamp, synchronizers using that candidate set cannot align it correctly.
Abstract
from arXiv · showhide
Combining event and tracking data is fundamental to modern soccer analytics, yet the two sources are rarely well aligned: event timestamps recorded by human annotators often miss the true moment of the action, distorting the spatiotemporal context that downstream models rely on. Existing synchronization methods depend on noisy human-annotated event locations and fail to detect ball receptions, obscuring when each player gains ball possession. To address these limitations, we propose ELASTIC (Event-Location-AgnoSTIC synchronizer), a framework that infers the start and end timestamps of events solely from player and ball trajectories, without relying on annotated event locations. To recover ball receptions, ELASTIC enriches the event sequence by inserting virtual termination events between consecutive events, so that the end of each event is detected jointly with its start. It then extracts a sparse set of candidate frames where ball touches are physically plausible, and aligns the termination-inserted event sequence with the candidate-frame sequence using an extended Needleman-Wunsch algorithm. For reproducible evaluation, we construct a publicly available benchmark by annotating ground-truth timestamps on the Sportec Open DFL Dataset, on which ELASTIC substantially outperforms existing methods. Through downstream task evaluation, we further show that improved synchronization translates into measurable gains in soccer analytics. The source code and benchmark are available at https://github.com/hyunsungkim-ds/elastic.git.
1 Introduction
ELASTIC addresses inaccurate event–tracking alignment caused by noisy human timestamps and locations, while also recovering ball receptions that existing synchronizers miss. It uses player and ball trajectories, candidate-frame extraction, and sequence alignment, supported by a reproducible benchmark and downstream-task evaluation.
- Motivation: Event data records annotated on-ball actions, while tracking data records player and ball positions continuously; combining them supports downstream soccer analytics.Examples include estimating shot scoring probability and other event-context analyses.
- Motivation: Human event timestamps can miss the true action moment, producing inaccurate player–ball configurations and distorted spatiotemporal context for downstream models.The resulting alignment problem pairs sparse semantic records with dense temporal signals.
- Prior limitations: Distance-based synchronizers can select a reception instead of a pass when the noisy annotated event location lies closer to the reception point.In the illustrated example, player #12 receives the ball at frame 38 and passes at frame 79, but the annotation favors the earlier reception.
- Prior limitations: Existing approaches either rely on noisy annotated locations, require labels and coarse windows, or align events globally while covering only selected event types.Tracking-only event detectors likewise recover only event types defined by their detection rules, whereas this work synchronizes existing fine-grained records.
- Prior limitations: Existing synchronizers do not detect ball receptions, leaving possession intervals and destination-conditioned pass analyses without accurate reception timing.Coverage ranges from shots or passes to several recorded event types, but reception moments remain unsynchronized.
- ELASTIC: ELASTIC detects event starts and ends from player and ball trajectories, inserts virtual termination events to recover receptions, and aligns sequences using candidate frames and extended Needleman–Wunsch.Candidate frames use cues such as player–ball distance and ball acceleration within in-play segments.
- Evaluation and contributions: The paper releases a ground-truth timestamp benchmark on the Sportec Open DFL Dataset and evaluates synchronization effects on next-action and pass-success prediction.The authors report substantial improvements over existing baselines and practically meaningful downstream gains.
2 Proposed Framework
ELASTIC synchronizes soccer event and tracking sequences by identifying physically plausible ball-touch frames and aligning them with an enriched event sequence. Its repeat-augmented Needleman-Wunsch procedure jointly determines event starts and ends, including receptions and events sharing one moment.
- Framework objective: ELASTIC infers true start and end timestamps from event and tracking sequences while preserving event order.Tracking snapshots contain player and ball positions, and event records specify action types and involved players.
- Framework pipeline: The framework processes matches episode by episode through preprocessing, candidate extraction, compatibility scoring, sequence alignment, and postprocessing.Episode batching reduces the alignment cost from O(MT) to O(KM′T′), with approximately 100 episodes per match.
- Event enrichment: Virtual termination events are inserted between selected adjacent events so one alignment jointly identifies every original event’s initiation and termination.Inserted events represent goals, out-of-play transitions, or receptions when consecutive same-episode events involve different players.
- Candidate extraction: ELASTIC extracts candidate frames from distance minima, pitch-boundary proximity minima, and ball-acceleration maxima, retaining only physically feasible player-ball pairs.Candidates are discarded when player-ball distance exceeds 3 m or ball height exceeds 4 m, then grouped by frame to handle simultaneous candidates.
- Pairwise scoring: Event-candidate compatibility combines trajectory-derived feature scores without human-annotated event locations, with category-dependent components and a hard acting-player constraint.The scoring functions map feature values to [0,1] and include ball acceleration, player-ball distance, and category-specific terms.
- Repeat-augmented alignment: A down-match extension lets two consecutive events share one candidate frame, handling contests and one-touch control-to-pass or control-to-shot sequences.The repeated match incurs a constant penalty r≤0 and assigns the candidate to both events.
3 Main Experiments
The experiments evaluate ELASTIC on a re-annotated Sportec Open DFL benchmark, testing synchronization accuracy, candidate coverage, feature sensitivity, and runtime. ELASTIC-NW achieves high timestamp accuracy, while ablations show benefits from global alignment, complementary candidate cues, and balanced features.
- 3.1 Ground-Truth Benchmark Construction: The benchmark re-annotates event timestamps in three Sportec Open DFL matches and adds missing termination events for evaluation.The annotation tool supports frame-level timestamp correction alongside match animation and event records.
- 3.1 Ground-Truth Benchmark Construction: 95.0% of events received identical timestamps from all annotators, with a 0.521-frame mean pairwise difference and 99.3% having two labels within two frames.These agreement statistics support using the median of the three timestamps as a reliable reference.
- 3.3 Synchronization Accuracy: 96.5% W2 accuracy for event starts and 93.6% for event ends give ELASTIC-NW 95.2% accuracy within two frames, with a 1.697-frame mean difference.ELASTIC-NW also exceeds ETSY, Biermann et al., and DataBallPy on event starts, including outgoing open-play and set-piece events.
- 3.3 Synchronization Accuracy: Global Needleman-Wunsch alignment raises event-start accuracy from 84.1% with ELASTIC-Greedy to 96.5% with ELASTIC-NW.The improvement is largest for minor events, increasing from 62.3% to 87.9%, because global alignment avoids sequential error propagation.
- 3.4 Candidate Frame Coverage: Candidate selection limits achievable synchronization because ELASTIC-NW cannot correctly align an event when no nearby candidate exists.W2 coverage therefore upper-bounds any synchronizer operating on the same candidate set.
- 3.4 Candidate Frame Coverage: Default candidate extraction covers 98.4% of ground-truth timestamps, whereas distance-only and acceleration-only conditions reduce coverage to 77.2% and 88.5%.The two cue types are complementary: distance conditions miss touches without local minima, while acceleration misses soft touches.
- 3.5 Hyperparameter Sensitivity: Removing any feature degrades its corresponding categories, while removing PBDS reduces set-piece accuracy to 38.7% and equal feature weights provide the most consistent performance.The kick-distance feature has the broadest impact across event categories.
- 3.5 Hyperparameter Sensitivity: Default clipping bounds are best or tied for best, while accuracy forms a wide plateau around default alignment penalties; ELASTIC-NW runs in under 40 seconds per match on average.The event-gap penalty is insensitive, whereas candidate-gap and repeat penalties can degrade accuracy in either direction.
4 Downstream Task Evaluation
The evaluation tests whether synchronized player-ball configurations improve downstream soccer prediction tasks, using graph-based models for next-action and pass-success prediction. Synchronization consistently improves performance, with ELASTIC-NW outperforming ELASTIC-Greedy across all reported metrics.
- Tasks and setup: The evaluation uses graph-based models that predict the ball possessor’s next action and whether passes to teammates will succeed.Both tasks represent each game state as a fully connected graph of players and two goals, processed with Graph Attention Network layers.
- Tasks and setup: Next-action prediction selects among teammate passes, dribbles, and shots using accuracy, crossentropy, and mean reciprocal rank.
- Tasks and setup: Pass-success prediction independently estimates each teammate pass’s outcome using F1 score, AUC, and Brier score.
- Tasks and setup: The models are trained and evaluated on unsynchronized data and data synchronized by ELASTIC-Greedy or ELASTIC-NW.The larger proprietary Dutch Eredivisie tracking dataset contains 200 training matches, 50 validation matches, and 157 test matches.
- Results: Data synchronization consistently improves downstream performance, and ELASTIC-NW further outperforms ELASTIC-Greedy across all metrics.For next-action accuracy, ELASTIC-Greedy increases performance from 0.5796 to 0.6759 relative to unsynchronized data.
5 Conclusion
The paper addresses noisy event locations and missing ball-reception timing by synchronizing event endings and starts from player and ball trajectories. Its benchmark and downstream evaluation show that alignment quality matters for the reliability of subsequent soccer analytics.
- ELASTIC removes dependence on noisy annotated event locations and recovers ball receptions that mark when events end.
- The re-annotated public benchmark supports reproducible evaluation of synchronization and its downstream analytical effects.
- Downstream evaluation shows that synchronization quality shapes the reliability of analytics built from aligned event and tracking data.
GenAI Usage Disclosure
The authors report limited use of generative AI tools for manuscript clarity, preprocessing, visualization, and debugging, while stating that the core research was developed and validated by the authors.
- Claude was used for phrasing and grammar refinement, while Claude Code supported preprocessing, visualization, and debugging.
- The authors state that the core contributions, methodology, experiments, and analytical insights were developed and validated entirely by them.