Scholarlink

<p>Human Activity Recognition (HAR) has emerged as a critical research area in the domains of Artificial Intelligence (AI) due to its extensive application in healthcare, intelligent monitoring, and human–computer interaction. The development of robust HAR models capable of accurately identifying human activities is growing in demand. This study aims to advance the field by proposing a novel hybrid framework that integrates Convolutional Neural Networks (CNN), Attention mechanisms and bidirectional LSTMs are used to extract spatial and temporal representations jointly. An ordered attention strategy is adopted, in which channel attention is applied first, followed by spatial attention, which enables refined feature selection before temporal modeling within a CNN–BiLSTM pipeline. The architecture achieves competitive performance with accuracies of 93% on UCI-HAR and 94% on MotionSense, while maintaining real-time efficiency with an inference latency of 65&#xa0;ms per instance on non-GPU hardware. We also include explainability through Gradient-weighted Class Activation Mapping (Grad-CAM), which highlights the sensor signals and time periods that are primarily responsible for recognizing each activity. To ensure the reliability and validity of our findings, we employed rigorous validation techniques, including cross-validation, detailed classification reports, and ablation studies. The use of attention-based modeling, combined with Grad-CAM visualization, makes our framework accurate, reliable, and suitable for practical human activity recognition 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

Hybrid deep learning architecture for efficient human activity recognition: a CNN-attention-BiLSTM framework

  • 期刊论文

  • 发表日期：

    2026年4月29日

查看全文
  • Sudipta Saha,
  • Purba Mukhopadhyay,
  • Krishnendu Mahata,
  • Saikat Basu,
  • Koushik Majumder

摘要

Human Activity Recognition (HAR) has emerged as a critical research area in the domains of Artificial Intelligence (AI) due to its extensive application in healthcare, intelligent monitoring, and human–computer interaction. The development of robust HAR models capable of accurately identifying human activities is growing in demand. This study aims to advance the field by proposing a novel hybrid framework that integrates Convolutional Neural Networks (CNN), Attention mechanisms and bidirectional LSTMs are used to extract spatial and temporal representations jointly. An ordered attention strategy is adopted, in which channel attention is applied first, followed by spatial attention, which enables refined feature selection before temporal modeling within a CNN–BiLSTM pipeline. The architecture achieves competitive performance with accuracies of 93% on UCI-HAR and 94% on MotionSense, while maintaining real-time efficiency with an inference latency of 65 ms per instance on non-GPU hardware. We also include explainability through Gradient-weighted Class Activation Mapping (Grad-CAM), which highlights the sensor signals and time periods that are primarily responsible for recognizing each activity. To ensure the reliability and validity of our findings, we employed rigorous validation techniques, including cross-validation, detailed classification reports, and ablation studies. The use of attention-based modeling, combined with Grad-CAM visualization, makes our framework accurate, reliable, and suitable for practical human activity recognition applications.