Scholarlink

<p>Early detection of human activity is essential in domains including robotics, entertainment, surveillance, and healthcare. Early detection that is accurate enables prompt decision-making, enhancing system responsiveness and overall effectiveness. Conventional action recognition techniques can’t handle sequential and incomplete data well since they are usually built for offline analysis and concentrate on detecting entire actions. Early detection necessitates real-time result prediction and incomplete activity identification, which are difficult for many current models to do. In order to enhance early detection and prediction of human behaviors, this study proposes a unique method utilizing a Bi-Directional Convolutional Long Short-Term Memory (Bi-ConvLSTM) network. By incorporating both spatial and temporal connections, the model processes sequential data and makes it possible to identify activity initiation and continuing activities with greater accuracy. By examining the temporal sequence of input frames, the Bi-ConvLSTM network is intended to identify the beginning of an activity and forecast its course. The proposed approach utilizes a segment-based strategy in which the input sequence is broken down into smaller intervals, allowing the model to focus on specific temporal segments. This improves the network’s capacity to recognize tiny motion patterns and contextual signals indicating the start of an activity. The model is tested on a real-world dataset that includes a variety of human behaviors recorded in complicated contexts. Experimental findings show that the proposed Bi-ConvLSTM model outperforms current models such as CNN, InceptionV3, VGG19, and regular ConvLSTM networks, with an average accuracy of 89.54%. The findings show that the Bi-ConvLSTM model efficiently balances early detection accuracy with decision-making speed, making it appropriate for real-time applications. This study demonstrates the ability of Bi-ConvLSTM networks to improve early detection and prediction of human behaviors, opening the door for more responsive and intelligent 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

Bi-directional ConvLSTM networks for early recognition of human activities and action prediction

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年11月6日

查看全文
  • M. Ashwin Shenoy,
  • N. Thillaiarasu,
  • S. Santhosh,
  • S. Sandeep Kumar

摘要

Early detection of human activity is essential in domains including robotics, entertainment, surveillance, and healthcare. Early detection that is accurate enables prompt decision-making, enhancing system responsiveness and overall effectiveness. Conventional action recognition techniques can’t handle sequential and incomplete data well since they are usually built for offline analysis and concentrate on detecting entire actions. Early detection necessitates real-time result prediction and incomplete activity identification, which are difficult for many current models to do. In order to enhance early detection and prediction of human behaviors, this study proposes a unique method utilizing a Bi-Directional Convolutional Long Short-Term Memory (Bi-ConvLSTM) network. By incorporating both spatial and temporal connections, the model processes sequential data and makes it possible to identify activity initiation and continuing activities with greater accuracy. By examining the temporal sequence of input frames, the Bi-ConvLSTM network is intended to identify the beginning of an activity and forecast its course. The proposed approach utilizes a segment-based strategy in which the input sequence is broken down into smaller intervals, allowing the model to focus on specific temporal segments. This improves the network’s capacity to recognize tiny motion patterns and contextual signals indicating the start of an activity. The model is tested on a real-world dataset that includes a variety of human behaviors recorded in complicated contexts. Experimental findings show that the proposed Bi-ConvLSTM model outperforms current models such as CNN, InceptionV3, VGG19, and regular ConvLSTM networks, with an average accuracy of 89.54%. The findings show that the Bi-ConvLSTM model efficiently balances early detection accuracy with decision-making speed, making it appropriate for real-time applications. This study demonstrates the ability of Bi-ConvLSTM networks to improve early detection and prediction of human behaviors, opening the door for more responsive and intelligent systems.