Scholarlink

<p>With the proliferation of wearable sensors and the Internet of Things (IoT), sensor-based human activity recognition (HAR) has become indispensable in various domains such as healthcare, sports tracking, and environmental automation. Traditional models like recurrent neural networks (RNNs) and convolutional neural networks (CNNs) face trade-offs between receptive field size and computational efficiency, limiting their ability to capture multi-scale temporal features. Transformer-based methods, while powerful, often require substantial data or resource-intensive optimizations, impractical in data-constrained HAR scenarios. To address these challenges, we propose AdaMaskNet, a novel deep learning framework that dynamically generates receptive fields based on the temporal length of sensor data. AdaMaskNet incorporates three key modules: receptive field scaling, multi-scale enhancement, and fine-grained refinement. These modules enable coarse-to-fine feature extraction, improving both recognition accuracy and computational efficiency. Evaluated on four popular HAR datasets, AdaMaskNet achieves state-of-the-art accuracy: 97.42% on UCI-HAR, 92.53% on PAMAP2, 76.02% on UNIMIB-SHAR, and 99% on WISDM. AdaMaskNet demonstrates its superiority in recognizing human activities across varying time scales and its practicality for real-world applications. Our code and models are released at <a href="https://github.com/Catherine618/AdaMaskNet">https://github.com/Catherine618/AdaMaskNet</a> for reproduction.</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

AdaMaskNet: adaptive multi-scale masked kernels for enhanced sensor-based human activity recognition

  • 期刊论文

  • 发表日期：

    2025年5月27日

查看全文
  • Jiamei Tang,
  • Chao Dong,
  • MengKun Li

摘要

With the proliferation of wearable sensors and the Internet of Things (IoT), sensor-based human activity recognition (HAR) has become indispensable in various domains such as healthcare, sports tracking, and environmental automation. Traditional models like recurrent neural networks (RNNs) and convolutional neural networks (CNNs) face trade-offs between receptive field size and computational efficiency, limiting their ability to capture multi-scale temporal features. Transformer-based methods, while powerful, often require substantial data or resource-intensive optimizations, impractical in data-constrained HAR scenarios. To address these challenges, we propose AdaMaskNet, a novel deep learning framework that dynamically generates receptive fields based on the temporal length of sensor data. AdaMaskNet incorporates three key modules: receptive field scaling, multi-scale enhancement, and fine-grained refinement. These modules enable coarse-to-fine feature extraction, improving both recognition accuracy and computational efficiency. Evaluated on four popular HAR datasets, AdaMaskNet achieves state-of-the-art accuracy: 97.42% on UCI-HAR, 92.53% on PAMAP2, 76.02% on UNIMIB-SHAR, and 99% on WISDM. AdaMaskNet demonstrates its superiority in recognizing human activities across varying time scales and its practicality for real-world applications. Our code and models are released at https://github.com/Catherine618/AdaMaskNet for reproduction.