Scholarlink

<p>Accurate and context-aware action recognition in soccer remains a challenging task due to the sport’s dynamic nature, overlapping actions, and complex causal relationships among events. Traditional video understanding models often rely on dense optical flow and static frame-based processing, which struggle in scenarios involving occlusion, background clutter, or subtle gameplay nuances. Furthermore, many existing systems fail to incorporate temporal reasoning or causal inference, leading to suboptimal recognition of interdependent actions like tackles, fouls, or passes. To address these limitations. This research aims to develop a cognitively inspired framework that overcomes the limitations of traditional video understanding models, particularly their reliance on dense optical flow and inability to model temporal causality and context effectively. We propose SPEAR-Net (Spiking-Perception and Episodic Abstraction for Recognition Network), a neuro-inspired action recognition system that emulates brain-like perception and reasoning. It introduces modules such as the Neuro-Perception Encoding Module (NPEM) for sparse event encoding, the Causal Relational Decoder (CRD) for DAG-based causal reasoning, the Context-Aware Environment Encoder (CAEE) for semantic context embedding, and the Micro–Macro Event Alignment (MMEA) for hierarchical temporal abstraction. Reliability is further enhanced through the Episodic Memory Replay Module (EMRM) and Self-Critic Refinement Agent (SCRA) for memory-based adaptation and confidence calibration. Empirical results on the SoccerNet-v2 benchmark demonstrate that SPEAR-Net outperforms state-of-the-art models with a Top-1 accuracy of 89.6%, robustness under occlusion (only 4.5% accuracy drop), and significantly reduced misclassification and uncertainty post-refinement, making it highly suitable for real-time, high-stakes sports analytics.</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

SPEAR-net: a neuro-inspired causal perception and episodic memory framework for fine-grained and context-aware soccer action recognition

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年10月28日

查看全文
  • Abdulrahman Alqarafi

摘要

Accurate and context-aware action recognition in soccer remains a challenging task due to the sport’s dynamic nature, overlapping actions, and complex causal relationships among events. Traditional video understanding models often rely on dense optical flow and static frame-based processing, which struggle in scenarios involving occlusion, background clutter, or subtle gameplay nuances. Furthermore, many existing systems fail to incorporate temporal reasoning or causal inference, leading to suboptimal recognition of interdependent actions like tackles, fouls, or passes. To address these limitations. This research aims to develop a cognitively inspired framework that overcomes the limitations of traditional video understanding models, particularly their reliance on dense optical flow and inability to model temporal causality and context effectively. We propose SPEAR-Net (Spiking-Perception and Episodic Abstraction for Recognition Network), a neuro-inspired action recognition system that emulates brain-like perception and reasoning. It introduces modules such as the Neuro-Perception Encoding Module (NPEM) for sparse event encoding, the Causal Relational Decoder (CRD) for DAG-based causal reasoning, the Context-Aware Environment Encoder (CAEE) for semantic context embedding, and the Micro–Macro Event Alignment (MMEA) for hierarchical temporal abstraction. Reliability is further enhanced through the Episodic Memory Replay Module (EMRM) and Self-Critic Refinement Agent (SCRA) for memory-based adaptation and confidence calibration. Empirical results on the SoccerNet-v2 benchmark demonstrate that SPEAR-Net outperforms state-of-the-art models with a Top-1 accuracy of 89.6%, robustness under occlusion (only 4.5% accuracy drop), and significantly reduced misclassification and uncertainty post-refinement, making it highly suitable for real-time, high-stakes sports analytics.