Master thesis : Understanding the phases of play leading to a goal in soccer

Abstract

``` Title of the Thesis: Understanding the phases of play leading to a goal in soccer Author: MPARIRWA Julien Section: Applied sciences, master in computer sciences with a focus in intelligent systems Academic Year: 2024-2025 Supervisors: Pr. VAN DROOGENBROECK Marc & Dr. CIOPPA Anthony

Summary

This master's thesis tackles the challenge of analyzing short soccer play phases to retrieve tactically similar actions and predict goal outcomes. Unlike existing methods that focus on isolated events, our goal was to capture tactical dynamics in the 15-second buildups preceding goals, enabling both retrieval of comparable plays and probabilistic outcome prediction.

Our methodology consisted of three pillars: dataset curation, multi-modal model design, and comprehensive evaluation. We built a dedicated dataset from SoccerNet-v2 with 4,000 clips covering background play (25%), pre-goal sequences (28%), free kicks (2.5%), penalties (2.8%), and shots not resulting in goals (41.7%). From this collection, we processed 1,600 clips through parallel feature extraction: tracking-based crowd features using ByteTrack and visual embeddings from a pretrained SlowFast model.

We developed two crowd feature representations to assess the impact of spatial granularity. The baseline (94D) computes global statistical descriptors, while the enhanced representation (95D) incorporates localized 3×3 grid decomposition, position covariance, and temporal statistics designed to capture zone-specific occupancy and temporal evolution of team shape.

We propose a multi-modal architecture that fuses visual and tracking features via an attention mechanism into a 256-dimensional embedding. Two task-specific heads operate on this space: a clustering head using pseudo-labels and a prediction head estimating goal probability, trained jointly with balanced loss weighting.

The enhanced 95D model achieved 67.2% goal prediction accuracy with balanced precision (50.8%) and recall (48.0%), representing a 6 percentage point improvement over the baseline 94D model (61.2%). Clustering remained weak (silhouette scores ≤0.171), suggesting soccer tactics exist along a continuous spectrum rather than discrete categories.

The attention mechanism exhibited different behavior between models: the baseline strongly preferred crowd features (63.6%) over visual (36.4%), while the enhanced model achieved near-perfect balance (50.1% visual, 49.9% tracking). This shift indicates genuine multimodal synergy where enhanced tracking features complement visual information more effectively.

Ablation studies revealed that visual features are essential: the tracking-only model achieved 66.5% accuracy but suffered catastrophic recall failure (12.5% vs. 48.0%), demonstrating that crowd features alone cannot reliably identify goal-scoring sequences.

In conclusion, this thesis contributes (1) a curated dataset with dual feature representations, (2) a multi-modal architecture demonstrating the value of localized spatial-temporal features, (3) ablation studies quantifying the necessity of multimodal fusion, and (4) protocols revealing soccer's tactical complexity as continuous rather than discrete. The enhanced model's 67.2% accuracy captures a substantial portion of predictable goal-scoring dynamics, with remaining variance likely attributable to stochastic factors beyond visual and spatial features. ```