PathCRF: Ball-Free Soccer Event Detection via Possession Path Inference from Player Trajectories

Technical Summary: PathCRF

Problem Statement
Professional soccer analytics relies on two primary data modalities: event data (manual annotations of on-ball actions like passes and shots) and tracking data (high-frequency coordinates of players and the ball). While player tracking has become increasingly automated and affordable, ball tracking remains costly, prone to occlusion errors, and difficult to scale beyond top-tier competitions. Consequently, comprehensive data collection is limited, and the broader adoption of data-driven analysis is hindered. Existing automatic event detection methods typically require accurate ball tracking, which is a bottleneck. Furthermore, prior attempts to infer ball states solely from player trajectories often fail to enforce physical constraints, resulting in implausible ball movements or illogical possession changes. The core problem addressed is how to detect on-ball soccer events and infer possession paths accurately using only player tracking data, without relying on expensive ball-tracking hardware or labor-intensive manual annotation.

Methodology
The authors propose PathCRF, a framework that reframes soccer event detection as a sequential edge selection problem over a fully connected dynamic graph.

1. Problem Formulation:

   • Player trajectories are modeled as a fully connected directed graph where nodes represent players and four "outside" nodes (representing the ball leaving the pitch).
   • The ball possession state at any time step is defined as exactly one edge: a self-loop $(u, u)$ indicates player $u$ is dribbling, while a directed edge $(u, v)$ with $u \neq v$ represents a pass or kick from $u$ to $v$.
   • The goal is to infer the sequence of edges (the possession path) that best explains the observed player trajectories.

2. Neural Backbone:

   • The framework employs a socio-temporal encoder inspired by ROLAND and TranSPORTmer.
   • Social Module: Uses a combination of Partially Permutation-Equivariant (PPE) and Fully Permutation-Equivariant (FPE) networks based on Induced Set Attention Blocks (ISAB). PPE captures intra-team dynamics, while FPE captures global inter-team contexts.
   • Temporal Module: Aggregates node representations over time using an attention-based architecture (Positional Encoding + Stacked Set Attention Blocks) to capture sequential dependencies.
   • Edge Embeddings: Refined node embeddings are concatenated and projected via an MLP to form edge embeddings, which serve as features for the CRF.

3. Dynamic Masked Conditional Random Field (CRF):

   • To ensure logical consistency, a CRF is applied over the sequence of edges.
   • Dynamic Scoring: Unlike standard CRFs with static transition matrices, PathCRF dynamically computes emission and transition scores from the edge embeddings. This allows the model to adapt transition probabilities to the evolving spatiotemporal context (e.g., the likelihood of a pass being completed changes as the ball approaches the receiver).
   • Hard Constraints (Masking): The CRF explicitly masks impossible transitions (e.g., a player cannot receive a pass from themselves, or a ball cannot "teleport" between non-adjacent players without a pass edge). Illegal transitions are assigned a large negative score, effectively pruning them from the search space.
   • Inference: The most probable edge sequence is obtained via Viterbi decoding. Events (controls, kicks, out-of-play) are detected at time steps where the selected edge changes.

4. Training:

   • The model is trained end-to-end by minimizing the negative log-likelihood of the ground-truth edge sequence.
   • Auxiliary losses (intermediate classification and emission classification) are added to stabilize training and encourage early learning of possession semantics.

Key Contributions

• Formulation: The paper formulates soccer event detection as a sequential edge selection problem over a dynamic graph, inferring possession paths directly from player trajectories without estimating ball positions.
• Neural CRF in Sports: It introduces neural CRF-based sequence inference, utilizing dynamic transition scoring and hard masking, to enforce domain-specific physical constraints in sports analytics.
• Performance without Ball Tracking: It demonstrates that key event-based analytics (spatial tendencies, possession statistics, pass networks) can be reliably approximated using only player tracking data, eliminating the need for manual event annotation or ball-tracking hardware.
• Reproducibility: The source code and experiments are based on publicly available tracking data (Sportec Open DFL Dataset), facilitating easy reproduction.

Experimental Results
Experiments were conducted on the Sportec Open DFL Dataset (German Bundesliga).

• Edge Accuracy: PathCRF achieved an edge accuracy of 69.64% in selecting the correct possession edge among 676 candidates.
• Event Detection: The framework achieved an F1-score of 75.69% for event detection.
• Logical Consistency: The Dynamic Masked CRF variant achieved a 0.00% violation rate, successfully eliminating physically impossible transitions that plagued non-CRF baselines (which had a 2.80% violation rate).
• Comparison: PathCRF significantly outperformed baselines that relied on ball trajectory regression or unconstrained edge selection. It also outperformed static CRF variants, highlighting the benefit of dynamic transition scoring.
• Downstream Utility: The detected events closely approximated ground-truth spatial heatmaps, team possession statistics (predicting 62.64% vs. ground truth 62.13%), and pass network structures (low Jensen-Shannon divergence and spectral distance).

Significance and Claims
The paper claims that PathCRF offers a scalable solution for soccer analytics by removing the dependency on expensive ball-tracking infrastructure and labor-intensive manual annotation. By producing logically consistent possession paths, the framework enables reliable downstream analyses such as tactical heatmaps, possession metrics, and pass network construction.

The authors position the work as a step toward democratizing automated data collection for lower-tier and youth competitions where ball tracking is unavailable. They also note that while the framework is designed for soccer, its formulation is domain-agnostic and could potentially extend to other multi-agent systems (e.g., autonomous driving, multi-robot coordination) by modifying the transition constraints. The paper modestly suggests that future work could integrate visual features from broadcast videos or human guidance to further refine accuracy and categorize events into more granular types.