Scholarlink

<p>This study introduces a novel method for recognizing human actions through the semantic segmentation of images. The aim is to enhance action-motion dynamics by directing attention toward regions crucial for action recognition. The proposed approach utilizes a SegNet model with an incorporated attention mechanism and modified bidirectional gated recurrent unit (BiGRU) backbone. The process begins with the generation of binary masks for each frame in a video dataset, achieved through a combination of grayscale conversion, Gaussian blurring, and adaptive thresholding. The emphasis on crucial regions for action recognition and capturing temporal variations is heightened through the application of the frame-ranking method. In our experiments, we observed that the proposed method significantly enhances the dynamics of the action-motion representation. The SegNet architecture was designed for semantic segmentation tasks and features an encoder-decoder architecture. In this structure, the model performs hierarchical feature extraction from the input image via the encoder, whereas the decoder focuses on reconstructing the segmented output. Attention is paid to the encoded feature maps, augmenting the model's capability to capture dependencies over extensive spatial ranges. A bidirectional GRU layer is employed to capture the sequential dependencies in the concatenated feature maps. The integration of the SegNet model with the attention mechanism and a BiGRU backbone, featuring an encoder-decoder architecture for feature extraction, classification, and segmentation, demonstrated superior performance in capturing nuanced spatiotemporal features. The proposed method demonstrated an accuracy of 98.52% for UCF101 and 84.25% for HMDB51. The findings reveal that the model achieves state-of-the-art results in human action recognition tasks, outperforming the existing methods in terms of accuracy. The combination of semantic segmentation and BiGRU-based temporal modeling proved effective in discerning intricate patterns of human motion, showcasing its potential for real-world applications in video analysis and surveillance systems.</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 novel human actions recognition and classification using semantic segmentation with deep learning techniques

  • 期刊论文

  • 发表日期：

    2025年2月8日

查看全文
  • M. Jayamohan,
  • S. Yuvaraj

摘要

This study introduces a novel method for recognizing human actions through the semantic segmentation of images. The aim is to enhance action-motion dynamics by directing attention toward regions crucial for action recognition. The proposed approach utilizes a SegNet model with an incorporated attention mechanism and modified bidirectional gated recurrent unit (BiGRU) backbone. The process begins with the generation of binary masks for each frame in a video dataset, achieved through a combination of grayscale conversion, Gaussian blurring, and adaptive thresholding. The emphasis on crucial regions for action recognition and capturing temporal variations is heightened through the application of the frame-ranking method. In our experiments, we observed that the proposed method significantly enhances the dynamics of the action-motion representation. The SegNet architecture was designed for semantic segmentation tasks and features an encoder-decoder architecture. In this structure, the model performs hierarchical feature extraction from the input image via the encoder, whereas the decoder focuses on reconstructing the segmented output. Attention is paid to the encoded feature maps, augmenting the model's capability to capture dependencies over extensive spatial ranges. A bidirectional GRU layer is employed to capture the sequential dependencies in the concatenated feature maps. The integration of the SegNet model with the attention mechanism and a BiGRU backbone, featuring an encoder-decoder architecture for feature extraction, classification, and segmentation, demonstrated superior performance in capturing nuanced spatiotemporal features. The proposed method demonstrated an accuracy of 98.52% for UCF101 and 84.25% for HMDB51. The findings reveal that the model achieves state-of-the-art results in human action recognition tasks, outperforming the existing methods in terms of accuracy. The combination of semantic segmentation and BiGRU-based temporal modeling proved effective in discerning intricate patterns of human motion, showcasing its potential for real-world applications in video analysis and surveillance systems.