Scholarlink

Human action and activity recognition (HRA) using frame images or videos is currently one of the interesting domains in the field of digital image processing (DIP), computer vision (CV), artificial intelligence (AI), and related areas due to its potential applications in computer science and information technology. This paper proposed a convolutional neural network (CNN) long short-term memory (LSTM) architecture for frame images or videos to understand HRA. A pretrained VGG16 architecture based on a CNN extracts spatial data from the input frames of images for us. Subsequently, an LSTM network is used for temporal feature sequences for predicting and classifying specific actions and activities within the video. We also analyzed the impacts of varying numbers of frames passed to the LSTM units on the performance of the system. To evaluate the proposed method, the author used the benchmark datasets KTH, UCF11, HMDB-51, and UCF50 for the training and testing phase. The model achieved classification accuracies of 93%, 93.20%, 72%, and 84% on these datasets, respectively. This improvement in results demonstrates the effectiveness of CNN-LSTM architectures for the video-based action recognition tasks, which represents the primary contribution of this work.

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

Analysis Human Action and Activity (HRA) for Video Using CNN and LSTM Approach

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Rakesh Y. Gedam,
  • Ameya R. Ramteke,
  • Rakesh J. Ramteke

摘要

Human action and activity recognition (HRA) using frame images or videos is currently one of the interesting domains in the field of digital image processing (DIP), computer vision (CV), artificial intelligence (AI), and related areas due to its potential applications in computer science and information technology. This paper proposed a convolutional neural network (CNN) long short-term memory (LSTM) architecture for frame images or videos to understand HRA. A pretrained VGG16 architecture based on a CNN extracts spatial data from the input frames of images for us. Subsequently, an LSTM network is used for temporal feature sequences for predicting and classifying specific actions and activities within the video. We also analyzed the impacts of varying numbers of frames passed to the LSTM units on the performance of the system. To evaluate the proposed method, the author used the benchmark datasets KTH, UCF11, HMDB-51, and UCF50 for the training and testing phase. The model achieved classification accuracies of 93%, 93.20%, 72%, and 84% on these datasets, respectively. This improvement in results demonstrates the effectiveness of CNN-LSTM architectures for the video-based action recognition tasks, which represents the primary contribution of this work.