Scholarlink

<p>The advancements in computer vision are crucial for development of numerous artificial intelligence applications, including self-driving vehicles, smart home systems, facial recognition systems, monitoring, automated medical diagnosis, entertainment, surveillance, security, etc. Human Activity Recognition (HAR) is stands out one of its emerging applications. However, accurately identifying human activity from a sequence of frames is challenging because of background clutter, varying viewpoints, low resolution, and partial occlusion. This study presents an advanced method for identifying human actions or activities through video data. For this purpose, Bidirectional Long Short-Term Memory network, a Convolutional Neural Network (CNN), and an attention mechanism were employed to enhance classification accuracy. Firstly, data augmentation techniques were utilized to expand the dataset and further deep features were extracted significantly from sequences of human activity videos using pre-trained CNNs, VGG16, InceptionV3, and MobileNet. Bi-LSTM and LSTM architectures were employed to capture the temporal dynamics within these sequences. These architectures processed the prioritized extracted deep features from the video frames, enabling to learn temporal patterns bidirectionally across each time step. Afterwords, an attention mechanism was employed to enhance the model’s ability of recognition. Publicly available datasets, YouTube11, UCF50, and a new dataset, HAR Video Dataset, were utilized to access the model’s performance. The proposed model achieved remarkable performance of 93.39%, 97%, and 99.2% on YouTube11, UCF50, and HAR Video Dataset, respectively, using MobileNet. Finally, the performance of the proposed model was compared and validated against previous models.</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

Hybridization of attention mechanism based CNN bi-directional LSTM model for enhancing HAR

  • 期刊论文

  • 发表日期：

    2025年9月5日

查看全文
  • Divya Yadav,
  • Deepika Rani,
  • Om Prakash Verma

摘要

The advancements in computer vision are crucial for development of numerous artificial intelligence applications, including self-driving vehicles, smart home systems, facial recognition systems, monitoring, automated medical diagnosis, entertainment, surveillance, security, etc. Human Activity Recognition (HAR) is stands out one of its emerging applications. However, accurately identifying human activity from a sequence of frames is challenging because of background clutter, varying viewpoints, low resolution, and partial occlusion. This study presents an advanced method for identifying human actions or activities through video data. For this purpose, Bidirectional Long Short-Term Memory network, a Convolutional Neural Network (CNN), and an attention mechanism were employed to enhance classification accuracy. Firstly, data augmentation techniques were utilized to expand the dataset and further deep features were extracted significantly from sequences of human activity videos using pre-trained CNNs, VGG16, InceptionV3, and MobileNet. Bi-LSTM and LSTM architectures were employed to capture the temporal dynamics within these sequences. These architectures processed the prioritized extracted deep features from the video frames, enabling to learn temporal patterns bidirectionally across each time step. Afterwords, an attention mechanism was employed to enhance the model’s ability of recognition. Publicly available datasets, YouTube11, UCF50, and a new dataset, HAR Video Dataset, were utilized to access the model’s performance. The proposed model achieved remarkable performance of 93.39%, 97%, and 99.2% on YouTube11, UCF50, and HAR Video Dataset, respectively, using MobileNet. Finally, the performance of the proposed model was compared and validated against previous models.