Scholarlink

<p>Recognizing tactical formations is a critical task in professional football performance analytics. However, existing methods often fail to capture the dynamic and continuously evolving nature of modern tactical systems within conventional classification frameworks. Current deep-learning models typically rely on snapshot-based recognition and struggle to capture continuous adaptations, role-switching, and multi-scale temporal dependencies characteristic of elite-level play. To address these limitations, this paper proposes the Hierarchical Dual stream Spatiotemporal Graph Transformer (HDS-SGT) as a new deep-learning architecture specifically designed to recognize dynamic tactical formations. The proposed model is a combination of a spatiotemporal graph attention network stream (equivalent to player-level relational dynamics) and a parallel temporal transformer stream (equivalent to hierarchical phase-transition patterns at varying temporal resolutions). A key innovation is the Dynamic Role Assignment Module, which uses learnable role embeddings along with cross-attention mechanisms and thus enables the recognition of fluid positional replacement during attacking and defensive transitions. Wide-ranging experiments on three professional-level datasets comprising 847 matches showed that HDS-SGT achieves a formation classification accuracy of 94.7% and a tactical transition detection F1-score of 0.912, which is 8.3 per cent and 11.6 per cent better than the state-of-the-art baselines, respectively. Importantly, real-time inference is sustained at 23.4 frames per second (including preprocessing and sliding-window buffering) on GPU-enabled hardware (NVIDIA RTX 3090), enabling deployment in live match-analysis applications. Ablation studies verify that the synergistic action of dual-stream fusion and multi-scale temporal modelling contributes to the overall performance improvements. These results represent a significant advance in understanding tactical dynamics in the team sports that has direct implications to coaching decision support systems, opponent analysis systems, and broadcast augmentation systems. However, the study relies on elite-level professional tracking data, which may limit generalizability to other contexts, and real-time performance was validated on specific GPU-enabled hardware configurations.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Deep learning for dynamic tactical formation recognition in professional football

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月17日

查看全文
  • YuDong Wang,
  • QianCheng ShangGuan,
  • GuoQiang Chen,
  • YuChen Chen

摘要

Recognizing tactical formations is a critical task in professional football performance analytics. However, existing methods often fail to capture the dynamic and continuously evolving nature of modern tactical systems within conventional classification frameworks. Current deep-learning models typically rely on snapshot-based recognition and struggle to capture continuous adaptations, role-switching, and multi-scale temporal dependencies characteristic of elite-level play. To address these limitations, this paper proposes the Hierarchical Dual stream Spatiotemporal Graph Transformer (HDS-SGT) as a new deep-learning architecture specifically designed to recognize dynamic tactical formations. The proposed model is a combination of a spatiotemporal graph attention network stream (equivalent to player-level relational dynamics) and a parallel temporal transformer stream (equivalent to hierarchical phase-transition patterns at varying temporal resolutions). A key innovation is the Dynamic Role Assignment Module, which uses learnable role embeddings along with cross-attention mechanisms and thus enables the recognition of fluid positional replacement during attacking and defensive transitions. Wide-ranging experiments on three professional-level datasets comprising 847 matches showed that HDS-SGT achieves a formation classification accuracy of 94.7% and a tactical transition detection F1-score of 0.912, which is 8.3 per cent and 11.6 per cent better than the state-of-the-art baselines, respectively. Importantly, real-time inference is sustained at 23.4 frames per second (including preprocessing and sliding-window buffering) on GPU-enabled hardware (NVIDIA RTX 3090), enabling deployment in live match-analysis applications. Ablation studies verify that the synergistic action of dual-stream fusion and multi-scale temporal modelling contributes to the overall performance improvements. These results represent a significant advance in understanding tactical dynamics in the team sports that has direct implications to coaching decision support systems, opponent analysis systems, and broadcast augmentation systems. However, the study relies on elite-level professional tracking data, which may limit generalizability to other contexts, and real-time performance was validated on specific GPU-enabled hardware configurations.