Scholarlink

<p>Human Activity Recognition (HAR) plays a pivotal role in computer vision, enabling the automatic detection and classification of human actions using video or sensor-based data. Its applications span diverse domains such as surveillance, healthcare, human-computer interaction, and smart environments. However, traditional HAR models face significant performance degradation in dark or low-light conditions due to poor visibility, low contrast, and increased noise. These limitations are particularly critical in night-time surveillance and security applications. To address these challenges, this study presents a lightweight and efficient HAR framework based on the MobileNetV1 architecture, which is specifically fine-tuned for low-illumination scenarios. The model leverages the Action Recognition in the Dark (ARID) v1.5 dataset—tailored for dim environments—and the UCF101 dataset for general action recognition. To ensure robustness and generalization, the architecture is enhanced with Global Average Pooling (GAP), dense layers with L2 regularization, Batch Normalization, and Dropout mechanisms. Extensive experiments demonstrate that the proposed approach achieves 97.69% accuracy on UCF101 and 75.14% on ARID v1.5, outperforming several state-of-the-art models in both performance and efficiency. These results validate the model’s robustness in low-light conditions and suitability for real-time applications on mobile and edge devices. Future work will explore multi-modal sensor fusion and advanced pre-processing techniques, including adaptive contrast enhancement and noise reduction, to further improve recognition accuracy in challenging environments.</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

Lightweight Deep Learning for Activity Recognition in Dark Environments

  • 期刊论文

  • 发表日期：

    2026年3月6日

查看全文
  • Ahmed Alkhayyat,
  • Gadug Sudhamsu,
  • Rami Oweis,
  • Mandeep Kaur Chohan,
  • Ankita Aggarwal,
  • Gupteswar Sahu,
  • Sarbeswara Hota,
  • Anita Gehlot

摘要

Human Activity Recognition (HAR) plays a pivotal role in computer vision, enabling the automatic detection and classification of human actions using video or sensor-based data. Its applications span diverse domains such as surveillance, healthcare, human-computer interaction, and smart environments. However, traditional HAR models face significant performance degradation in dark or low-light conditions due to poor visibility, low contrast, and increased noise. These limitations are particularly critical in night-time surveillance and security applications. To address these challenges, this study presents a lightweight and efficient HAR framework based on the MobileNetV1 architecture, which is specifically fine-tuned for low-illumination scenarios. The model leverages the Action Recognition in the Dark (ARID) v1.5 dataset—tailored for dim environments—and the UCF101 dataset for general action recognition. To ensure robustness and generalization, the architecture is enhanced with Global Average Pooling (GAP), dense layers with L2 regularization, Batch Normalization, and Dropout mechanisms. Extensive experiments demonstrate that the proposed approach achieves 97.69% accuracy on UCF101 and 75.14% on ARID v1.5, outperforming several state-of-the-art models in both performance and efficiency. These results validate the model’s robustness in low-light conditions and suitability for real-time applications on mobile and edge devices. Future work will explore multi-modal sensor fusion and advanced pre-processing techniques, including adaptive contrast enhancement and noise reduction, to further improve recognition accuracy in challenging environments.