Scholarlink

<p>The aging population increases healthcare demands, driving the need for Human Activity Recognition (HAR) using multimodal sensor data. Yet HAR faces challenges due to high-dimensional, noisy inputs and limited accuracy on complex datasets. This paper introduces two solutions: (1) a Correlation-Centroid Reduction (CCR) method to reduce feature complexity and noise, and (2) ARCL-Net, a hybrid deep learning model integrating convolutional, residual, LSTM, and attention layers for improved HAR performance. To enhance adaptability, ARCL Net incorporates multiple activation functions within its design, while the Cuckoo Search algorithm primarily optimizes network hyperparameters and architectural components, with activation functions included as part of the hyperparameter search space, to determine the most suitable configuration for each dataset. We evaluate the proposed approach on four benchmark HAR datasets (PAMAP2, UCI HAR, UniMiB SHAR, and multiple variants of the Opportunity dataset) and demonstrate consistently strong performance across all of them. ARCL-Net achieved F1 scores of 0.9926 on PAMAP2, 0.9792 on UCI HAR, and 0.9787 on UniMiB SHAR. In Opportunity dataset, we analyze Locomotion and High-Level activities using standard and label-based windowing. While standard windowing yields solid performance (0.9328 and 0.9885, respectively), label-based segmentation further improves results (0.9605 and 0.9919). The proposed work advances HAR towards more accurate, adaptive, and deployable systems for practical 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

An adaptive attention residual convolutional LSTM network for robust human activity recognition in healthy aging

  • 期刊论文

  • 发表日期：

    2026年7月20日

查看全文
  • Mohamed Abderrahmen Boulahia,
  • Saber Benharzallah,
  • Abdelkader Laouid,
  • Faiza Titouna,
  • Tahar Dilekh

摘要

The aging population increases healthcare demands, driving the need for Human Activity Recognition (HAR) using multimodal sensor data. Yet HAR faces challenges due to high-dimensional, noisy inputs and limited accuracy on complex datasets. This paper introduces two solutions: (1) a Correlation-Centroid Reduction (CCR) method to reduce feature complexity and noise, and (2) ARCL-Net, a hybrid deep learning model integrating convolutional, residual, LSTM, and attention layers for improved HAR performance. To enhance adaptability, ARCL Net incorporates multiple activation functions within its design, while the Cuckoo Search algorithm primarily optimizes network hyperparameters and architectural components, with activation functions included as part of the hyperparameter search space, to determine the most suitable configuration for each dataset. We evaluate the proposed approach on four benchmark HAR datasets (PAMAP2, UCI HAR, UniMiB SHAR, and multiple variants of the Opportunity dataset) and demonstrate consistently strong performance across all of them. ARCL-Net achieved F1 scores of 0.9926 on PAMAP2, 0.9792 on UCI HAR, and 0.9787 on UniMiB SHAR. In Opportunity dataset, we analyze Locomotion and High-Level activities using standard and label-based windowing. While standard windowing yields solid performance (0.9328 and 0.9885, respectively), label-based segmentation further improves results (0.9605 and 0.9919). The proposed work advances HAR towards more accurate, adaptive, and deployable systems for practical applications.