Scholarlink

<p>We present a real-time deep learning system for human activity recognition (HAR) in human–robot interaction (HRI) settings. The approach fuses MoViNet and R(2+1)D via late (score-level) fusion and a lightweight meta-learner, enabling accurate, streaming-friendly recognition on a 19-class HRI dataset. Under a unified protocol (224 resolution, 16-frame clips at 12&#xa0;fps, Adam with cosine learning rate), our method achieves 83.9% Top-1, 97.6% Top-5, and 83.5% macro F<sub>1</sub>, outperforming the best single backbone by +2.7 Top-1. The implementation sustains 17.9&#xa0;ms per frame (~56&#xa0;FPS) on an NVIDIA T4 (batch=1, streaming), demonstrating suitability for real-time deployment. An ablation with TV-L1 optical flow shows modest but consistent gains from RGB+Flow late fusion over RGB-only. These results indicate that efficient score-level fusion of complementary video models offers a practical accuracy–latency trade-off for collaborative HRI applications.</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

Act-Reco: Development of activity recognition framework based on deep learning in the real-time environment

  • 期刊论文

  • 发表日期：

    2025年12月22日

查看全文
  • Aaditya Garg,
  • Ashmit Calja,
  • Priya Kansal,
  • Sachin Kansal

摘要

We present a real-time deep learning system for human activity recognition (HAR) in human–robot interaction (HRI) settings. The approach fuses MoViNet and R(2+1)D via late (score-level) fusion and a lightweight meta-learner, enabling accurate, streaming-friendly recognition on a 19-class HRI dataset. Under a unified protocol (224 resolution, 16-frame clips at 12 fps, Adam with cosine learning rate), our method achieves 83.9% Top-1, 97.6% Top-5, and 83.5% macro F1, outperforming the best single backbone by +2.7 Top-1. The implementation sustains 17.9 ms per frame (~56 FPS) on an NVIDIA T4 (batch=1, streaming), demonstrating suitability for real-time deployment. An ablation with TV-L1 optical flow shows modest but consistent gains from RGB+Flow late fusion over RGB-only. These results indicate that efficient score-level fusion of complementary video models offers a practical accuracy–latency trade-off for collaborative HRI applications.