Characterisation of Goalkeeper Actions from Skeletal Data
- published
Modern player-tracking systems provide detailed 3D skeletal trajectories, yet technical analysis in soccer remains constrained by limited labeled datasets and reliance on predefined action taxonomies. We present a fully unsupervised framework for discovering goalkeeper behaviors directly from large-scale skeletal data collected across all 51 matches of UEFA Euro 2024 using Hawk-Eye SkeleTRACK. Our preprocessing synchronizes ball and skeleton data streams and extracts continuous goalkeeper trajectories when the ball is in the defensive final third. Goalkeeper trajectories are further filtered to retain high-energy segments, which serve as a proxy for interesting or salient movements, using an adaptive motion-energy signal. These sequences are then represented in a canonical goal-centered coordinate frame to enable analysis that is invariant to the goalkeeper’s absolute position on the pitch
To learn motion representations without annotations, we use the CrosSCLR contrastive learning framework with an ST-GCN backbone and multi-view encoders operating on joint, motion, and bone representations. The model is trained with domain-specific augmentations tailored to goalkeeping dynamics. The resulting 128-dimensional embeddings are projected using UMAP and clustered with HDBSCAN to identify recurring behaviors in a fully data-driven manner.
Training the model from scratch outperforms fine-tuning from NTU RGB+D, producing coherent and semantically interpretable clusters corresponding to ready stances, lateral shuffling, ball distribution, and diverse save techniques. Further subclustering of the save category isolates fine-grained mechanics such as lateral dives, low saves, forward smothers, and jumping punches. We assess cluster quality by visually inspecting representative clips and analyzing the kinematic characteristics of each cluster. Future work should focus on enriching representations with additional contextual features, refining clip segmentation, and systematically evaluating how preprocessing choices—such as spatial normalization and dimensionality reduction—impact clustering stability and the quality of resulting clusters.