Scholarlink

<p>Human Activity Recognition (HAR) is crucial in multiple fields. Existing HAR techniques include manual feature extraction, codebook-based methods, and deep learning, each with limitations. This paper presents DCAM-Net (DeepConvAttentionMLPNet), a novel deep neural network model without relying on pre-trained model weights. It integrates CNN and MLP with an attention mechanism. Experiments using data from 30 participants’ smartphone sensors (acceleration and gyroscope) show that after preprocessing and sampling, the model takes 561-dimensional feature vectors as input. With multi-scale feature extraction, residual and skip connections, and dual attention mechanisms, along with a series of optimization strategies like dropout, batch normalization, and AdamW optimizer, the model achieves an average accuracy of 99.03% in five-fold cross-validation. It outperforms other models and has good generalization ability. However, future work could involve using more diverse datasets, improving computational efficiency for real-time applications, enhancing activity transition recognition, and fusing other sensor data to better meet practical needs.</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

A human activity recognition model based on deep neural network integrating attention mechanism

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年7月2日

查看全文
  • Feng Xu,
  • Xuchen Gao,
  • Weigang Wang

摘要

Human Activity Recognition (HAR) is crucial in multiple fields. Existing HAR techniques include manual feature extraction, codebook-based methods, and deep learning, each with limitations. This paper presents DCAM-Net (DeepConvAttentionMLPNet), a novel deep neural network model without relying on pre-trained model weights. It integrates CNN and MLP with an attention mechanism. Experiments using data from 30 participants’ smartphone sensors (acceleration and gyroscope) show that after preprocessing and sampling, the model takes 561-dimensional feature vectors as input. With multi-scale feature extraction, residual and skip connections, and dual attention mechanisms, along with a series of optimization strategies like dropout, batch normalization, and AdamW optimizer, the model achieves an average accuracy of 99.03% in five-fold cross-validation. It outperforms other models and has good generalization ability. However, future work could involve using more diverse datasets, improving computational efficiency for real-time applications, enhancing activity transition recognition, and fusing other sensor data to better meet practical needs.