Scholarlink

<p>Recently, convolutional neural networks have emerged as a dominant approach in sensor-based human activity recognition (HAR) due to the powerful feature representation capabilities for multimodal time series data, which has become a research hotspot in the field of mobile computing. The attention mechanism is well-known for its potential to enhance activity recognition performance by emphasizing critical temporal sequences and noteworthy sensor modalities. However, existing research primarily focuses on developing deeper or more intricate convolutional attention networks, which inevitably incurs a significant increase in computational complexity. To address these concerns above, we propose a lightweight multi-scale Segmentation Attention called MSA method tailored for sensor-based HAR scenarios. MSA leverages a segmentation structure to extract a channel vector that encapsulates multi-scale information and employs one-dimensional convolution to facilitate cross-channel interaction while substantially reducing model complexity. We evaluate the proposed model on several publicly available HAR datasets, including WISDM, UCI–HAR, PAMAP2, and UniMib–SHAR. The experimental results demonstrate that our method significantly reduces the number of parameters compared to baseline models, while outperforming recently published results under identical configurations. Furthermore, extensive ablation studies are conducted to explore the impact of different group sizes on classification performance. The codes and models will be released at <a href="https://github.com/qwm1/decon">https://github.com/qwm1/decon</a>.</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

MSA–HAR: multi-scale segmented attention networks for human activity recognition using sensor signals

  • 期刊论文

  • 发表日期：

    2025年7月9日

查看全文
  • Weiming Quan,
  • Yin Tang,
  • Wei Luo,
  • Lei Zhang

摘要

Recently, convolutional neural networks have emerged as a dominant approach in sensor-based human activity recognition (HAR) due to the powerful feature representation capabilities for multimodal time series data, which has become a research hotspot in the field of mobile computing. The attention mechanism is well-known for its potential to enhance activity recognition performance by emphasizing critical temporal sequences and noteworthy sensor modalities. However, existing research primarily focuses on developing deeper or more intricate convolutional attention networks, which inevitably incurs a significant increase in computational complexity. To address these concerns above, we propose a lightweight multi-scale Segmentation Attention called MSA method tailored for sensor-based HAR scenarios. MSA leverages a segmentation structure to extract a channel vector that encapsulates multi-scale information and employs one-dimensional convolution to facilitate cross-channel interaction while substantially reducing model complexity. We evaluate the proposed model on several publicly available HAR datasets, including WISDM, UCI–HAR, PAMAP2, and UniMib–SHAR. The experimental results demonstrate that our method significantly reduces the number of parameters compared to baseline models, while outperforming recently published results under identical configurations. Furthermore, extensive ablation studies are conducted to explore the impact of different group sizes on classification performance. The codes and models will be released at https://github.com/qwm1/decon.