Scholarlink

<p>Accurate recognition of human activities in complex environments is vital for numerous applications, including healthcare, smart environments, and human-machine interaction. This research presents a novel hybrid approach to enhance the robustness and accuracy of human activity recognition (HAR) by leveraging multimodal data fusion techniques. Depth data, processed through Convolutional Neural Networks (CNNs), and inertial data, analyzed using Long Short-Term Memory (LSTM) networks, are effectively integrated to exploit complementary spatial and temporal features. Additionally, Principal Component Analysis (PCA) is employed alongside autoencoders for efficient dimensionality reduction and feature enhancement. The proposed PCA-Enhanced Hybrid Pipeline is evaluated using the UTD Multimodal Human Action Dataset, showcasing its ability to integrate depth-derived spatial insights with inertial sensor-based temporal dynamics. Experimental results highlight the superiority of the proposed method, achieving a weighted accuracy of up to 97.56% with RMSProp optimizer, surpassing traditional autoencoder-based pipelines. This approach demonstrates exceptional robustness and accuracy in recognizing human activities across diverse scenarios. The findings underscore the potential of multimodal fusion techniques for advancing HAR systems, offering a promising solution for applications in healthcare monitoring, smart environments, and beyond.</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

FusionNet: a PCA-enhanced hybrid framework for robust multimodal human activity recognition

  • 期刊论文

  • 发表日期：

    2026年5月7日

查看全文
  • Basamma Umesh Patil,
  • D V Ashoka

摘要

Accurate recognition of human activities in complex environments is vital for numerous applications, including healthcare, smart environments, and human-machine interaction. This research presents a novel hybrid approach to enhance the robustness and accuracy of human activity recognition (HAR) by leveraging multimodal data fusion techniques. Depth data, processed through Convolutional Neural Networks (CNNs), and inertial data, analyzed using Long Short-Term Memory (LSTM) networks, are effectively integrated to exploit complementary spatial and temporal features. Additionally, Principal Component Analysis (PCA) is employed alongside autoencoders for efficient dimensionality reduction and feature enhancement. The proposed PCA-Enhanced Hybrid Pipeline is evaluated using the UTD Multimodal Human Action Dataset, showcasing its ability to integrate depth-derived spatial insights with inertial sensor-based temporal dynamics. Experimental results highlight the superiority of the proposed method, achieving a weighted accuracy of up to 97.56% with RMSProp optimizer, surpassing traditional autoencoder-based pipelines. This approach demonstrates exceptional robustness and accuracy in recognizing human activities across diverse scenarios. The findings underscore the potential of multimodal fusion techniques for advancing HAR systems, offering a promising solution for applications in healthcare monitoring, smart environments, and beyond.