Scholarlink

<p>Real-time human activity recognition (HAR) must balance accuracy and latency under device constraints. We present SERA-HAR, a residual temporal convolutional framework augmented by two lightweight attentions. A staged early-exit attention reads the dispersion of temporal attention to decide when to stop observing and classify, while an inter-channel correlation attention (ICCA) captures cross-sensor dependencies to complement temporal encoding. On UCI-HAR, WISDM, and KU-HAR, SERA-HAR attains 94.6–96.7% accuracy and up to 95.95% macro-F1 within a unified time-domain pipeline. On a workstation, mean inference latency is 1.79&#xa0;ms with throughput of 760 samples/s; the early-exit policy reduces expected observation time by as much as 43% relative to fixed-window processing. These results indicate a favorable accuracy–latency trade-off for wearables, healthcare monitoring, and pervasive sensing.</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

SERA-HAR: A real-time human activity recognition framework via staged early-exit and inter-channel attention

  • 期刊论文

  • 发表日期：

    2025年11月10日

查看全文
  • Zhengjie Li,
  • Ming Fu,
  • Baole Fu,
  • Yinhua Liu

摘要

Real-time human activity recognition (HAR) must balance accuracy and latency under device constraints. We present SERA-HAR, a residual temporal convolutional framework augmented by two lightweight attentions. A staged early-exit attention reads the dispersion of temporal attention to decide when to stop observing and classify, while an inter-channel correlation attention (ICCA) captures cross-sensor dependencies to complement temporal encoding. On UCI-HAR, WISDM, and KU-HAR, SERA-HAR attains 94.6–96.7% accuracy and up to 95.95% macro-F1 within a unified time-domain pipeline. On a workstation, mean inference latency is 1.79 ms with throughput of 760 samples/s; the early-exit policy reduces expected observation time by as much as 43% relative to fixed-window processing. These results indicate a favorable accuracy–latency trade-off for wearables, healthcare monitoring, and pervasive sensing.