Scholarlink

<p>Human Action Recognition (HAR) is a significant research area in computer vision with applications in surveillance, healthcare, and human–computer interaction. This study proposes a hybrid deep learning framework that integrates the lightweight MobileNetV1 architecture with the scalable EfficientNetB3 network to achieve a balance between computational efficiency and recognition accuracy. The hybrid design leverages the complementary strengths of MobileNetV1 for efficient low-level feature extraction and EfficientNetB3 for enhanced high-level feature representation, enabling improved discriminative capability while maintaining reduced computational complexity. The model is evaluated on the UCF101 dataset, comprising 13,320 video clips across 101 action classes. A frame-based classification strategy with uniform sampling is adopted to reduce computational overhead, and data augmentation techniques such as rotation, shifting, shearing, and brightness adjustment are applied to improve generalization. Experimental results demonstrate that the proposed model achieves an accuracy of 89.18%, precision of 89.38%, and F1-score of 89.16%, with an AUC of 96.30%. In addition to accuracy, the model contains approximately 17.8&#xa0;million parameters, requires around 3.6 GFLOPs, and has an estimated model size of 71&#xa0;MB, highlighting its computational efficiency compared to more complex spatio-temporal architectures. The model achieves an average inference time of approximately 12–18 ms per frame on a standard GPU, indicating its potential for near real-time performance. While the current evaluation is limited to a single dataset and frame-level modeling, the results demonstrate that the proposed hybrid framework provides a favorable trade-off between accuracy and efficiency. This makes it a promising candidate for deployment in resource-constrained environments, with future work focusing on temporal modeling and edge-device optimization.</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

Efficient Human Action Recognition Using MobileNetV1 and EfficientNetB3 Based Hybrid Network

  • 期刊论文

  • 发表日期：

    2026年5月12日

查看全文
  • Muhammad Irsyad Abdullah,
  • Ahmed Alkhayyat,
  • Gadug Sudhamsu,
  • Pooja Rani,
  • Aman Shankhyan,
  • Jasgurpreet Singh Chohan,
  • M. Janaki Ramudu,
  • Devendra Singh

摘要

Human Action Recognition (HAR) is a significant research area in computer vision with applications in surveillance, healthcare, and human–computer interaction. This study proposes a hybrid deep learning framework that integrates the lightweight MobileNetV1 architecture with the scalable EfficientNetB3 network to achieve a balance between computational efficiency and recognition accuracy. The hybrid design leverages the complementary strengths of MobileNetV1 for efficient low-level feature extraction and EfficientNetB3 for enhanced high-level feature representation, enabling improved discriminative capability while maintaining reduced computational complexity. The model is evaluated on the UCF101 dataset, comprising 13,320 video clips across 101 action classes. A frame-based classification strategy with uniform sampling is adopted to reduce computational overhead, and data augmentation techniques such as rotation, shifting, shearing, and brightness adjustment are applied to improve generalization. Experimental results demonstrate that the proposed model achieves an accuracy of 89.18%, precision of 89.38%, and F1-score of 89.16%, with an AUC of 96.30%. In addition to accuracy, the model contains approximately 17.8 million parameters, requires around 3.6 GFLOPs, and has an estimated model size of 71 MB, highlighting its computational efficiency compared to more complex spatio-temporal architectures. The model achieves an average inference time of approximately 12–18 ms per frame on a standard GPU, indicating its potential for near real-time performance. While the current evaluation is limited to a single dataset and frame-level modeling, the results demonstrate that the proposed hybrid framework provides a favorable trade-off between accuracy and efficiency. This makes it a promising candidate for deployment in resource-constrained environments, with future work focusing on temporal modeling and edge-device optimization.