Scholarlink

<p>This paper presents a novel approach to enhance real-time surveillance systems by overcoming two key limitations in traditional models: slow tracking and low-resolution output. A dual-stream processing architecture is proposed, integrating a dedicated facial metadata stream to improve responsiveness and tracking accuracy. By isolating facial features and excluding background data, the system achieves greater computational and storage efficiency. Leveraging AI-based computer vision and facial recognition techniques, the model significantly increases the frame processing rate, enabling precise coordinate plotting and faster face tracking. The system’s performance was evaluated in four areas: disk I/O speed, tracking accuracy, GPU performance, and frame rate. Disk tests showed peak write and read speeds of 13.542&#xa0;MB/s and 2.742&#xa0;MB/s, respectively, sufficient under peak load, though occasional write drops point to possible buffer overflow risks. Accuracy testing revealed results with less than 1% error compared to reference values. GPU analysis showed dynamic power scaling from 5.564 to 30.857 W, with clock speeds up to 1785&#xa0;MHz, demonstrating adaptive resource use. Frame rate testing confirmed that the new model consistently captured up to 60 frames, a significant improvement over the 21-frame average in legacy systems. This efficiency in resource utilization not only boosts tracking precision and system reliability but also supports privacy-aware surveillance by avoiding unnecessary data collection. The proposed method proves effective in delivering smart, high-performance, and sustainable security solutions, paving the way for next-generation surveillance technologies aligned with modern privacy standards.</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

Object detection and classification using computer vision and deep learning on edge devices

  • 期刊论文

  • 发表日期：

    2025年11月26日

查看全文
  • S. Mary Joans,
  • J. S. Leena Jasmine,
  • N. Gomathi,
  • R. Gayathri

摘要

This paper presents a novel approach to enhance real-time surveillance systems by overcoming two key limitations in traditional models: slow tracking and low-resolution output. A dual-stream processing architecture is proposed, integrating a dedicated facial metadata stream to improve responsiveness and tracking accuracy. By isolating facial features and excluding background data, the system achieves greater computational and storage efficiency. Leveraging AI-based computer vision and facial recognition techniques, the model significantly increases the frame processing rate, enabling precise coordinate plotting and faster face tracking. The system’s performance was evaluated in four areas: disk I/O speed, tracking accuracy, GPU performance, and frame rate. Disk tests showed peak write and read speeds of 13.542 MB/s and 2.742 MB/s, respectively, sufficient under peak load, though occasional write drops point to possible buffer overflow risks. Accuracy testing revealed results with less than 1% error compared to reference values. GPU analysis showed dynamic power scaling from 5.564 to 30.857 W, with clock speeds up to 1785 MHz, demonstrating adaptive resource use. Frame rate testing confirmed that the new model consistently captured up to 60 frames, a significant improvement over the 21-frame average in legacy systems. This efficiency in resource utilization not only boosts tracking precision and system reliability but also supports privacy-aware surveillance by avoiding unnecessary data collection. The proposed method proves effective in delivering smart, high-performance, and sustainable security solutions, paving the way for next-generation surveillance technologies aligned with modern privacy standards.