Scholarlink

<p>This research aims to enhance activity recognition within indoor environments by proposing an efficient and accurate SConv-LSTM model. The motivation stems from the growing demand to improve intelligent monitoring systems in smart homes, where multiple complex and compound activities are performed daily. Existing models often struggle to capture the full spatiotemporal dynamics of such environments, leading to suboptimal recognition performance. To address this, the proposed framework combines separable convolutional layers, which reduce computational complexity while effectively extracting spatial features, with long short-term memory (LSTM) networks, which excel at modeling temporal dependencies in sensor data. The main goal is to enable real-time activity recognition with high accuracy and computational efficiency, making the system practical for deployment in resource-constrained environments. The model is rigorously evaluated using three benchmark multimodal activity datasets: HWU/USP, PAMAP2, and MHEALTH. It achieves high accuracy rates of 97%, 95%, and 96%, respectively, outperforming existing state-of-the-art methods. These results demonstrate the model’s strong potential for improving human activity recognition in domestic settings.</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

Multimodal activity recognition using separable convolutional long short-term memory

  • 期刊论文

  • 发表日期：

    2026年3月30日

查看全文
  • Gunjan Pareek,
  • Rajiv Singh,
  • Swati Nigam

摘要

This research aims to enhance activity recognition within indoor environments by proposing an efficient and accurate SConv-LSTM model. The motivation stems from the growing demand to improve intelligent monitoring systems in smart homes, where multiple complex and compound activities are performed daily. Existing models often struggle to capture the full spatiotemporal dynamics of such environments, leading to suboptimal recognition performance. To address this, the proposed framework combines separable convolutional layers, which reduce computational complexity while effectively extracting spatial features, with long short-term memory (LSTM) networks, which excel at modeling temporal dependencies in sensor data. The main goal is to enable real-time activity recognition with high accuracy and computational efficiency, making the system practical for deployment in resource-constrained environments. The model is rigorously evaluated using three benchmark multimodal activity datasets: HWU/USP, PAMAP2, and MHEALTH. It achieves high accuracy rates of 97%, 95%, and 96%, respectively, outperforming existing state-of-the-art methods. These results demonstrate the model’s strong potential for improving human activity recognition in domestic settings.