Scholarlink

Human Action Recognition (HAR) aims to automate behavior analysis via video, facing challenges in complex backgrounds, multimodal fusion efficiency, and computational costs. We propose NeuroSync, a neuroscience-inspired framework with three innovations: (1) a dual-path architecture (ventral for static, dorsal for dynamic features) enhanced by Dynamic Spatial Attention Gating (DSAG) to suppress noise; (2) a Causal Debiasing Module (CSD) to mitigate scene bias via counterfactual reasoning; (3) Hierarchical Spatio-Temporal Compression and Multi-Granularity Fusion (MGF) to reduce complexity while preserving critical features. NeuroSync achieves 86.1% (Kinetics-400) and 95.6% (UCF101) Top-1 accuracy, outperforming state-of-the-art (SOTA) methods while balancing accuracy and efficiency for real-world HAR applications.

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

NeuroSync: A Dual-Path Dynamically Modulated Framework with Spatiotemporal Compression for Human Action Recognition

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Xingquan Cai,
  • Haoyu Song,
  • Yupeng Zhang,
  • Jiatong Li,
  • Haiyan Sun

摘要

Human Action Recognition (HAR) aims to automate behavior analysis via video, facing challenges in complex backgrounds, multimodal fusion efficiency, and computational costs. We propose NeuroSync, a neuroscience-inspired framework with three innovations: (1) a dual-path architecture (ventral for static, dorsal for dynamic features) enhanced by Dynamic Spatial Attention Gating (DSAG) to suppress noise; (2) a Causal Debiasing Module (CSD) to mitigate scene bias via counterfactual reasoning; (3) Hierarchical Spatio-Temporal Compression and Multi-Granularity Fusion (MGF) to reduce complexity while preserving critical features. NeuroSync achieves 86.1% (Kinetics-400) and 95.6% (UCF101) Top-1 accuracy, outperforming state-of-the-art (SOTA) methods while balancing accuracy and efficiency for real-world HAR applications.