Scholarlink

<p>Edge computing emerges as the optimal choice for fire detection under extreme conditions with the advent of nascent visual technologies. This paper proposes an efficient deployment of the YOLOv7 model for real-time fire detection on a custom multi-core edge computing platform. The system utilizes 36 ARM cores, organized into clusters, and integrates FPGA-based optimization for low-latency, high-throughput image processing. YOLOv7 is trained on a fire image dataset, featuring diverse fire scenarios, and is partitioned across multiple cores to balance computational load and enhance energy efficiency. The deployment strategy incorporates modular architecture, including feature extraction, detection, and output modules, to optimize system performance. Experimental results show that the system achieves an inference speed of 11.5 FPS with an average latency of 86.29&#xa0;ms, meeting the real-time processing demands of fire detection in edge environments. The platform demonstrates a balanced trade-off between computational throughput, resource utilization, and energy consumption, with dynamic power consumption averaging 3.902 W. These findings highlight the system’s suitability for large-scale fire monitoring applications in resource-constrained environments, and the study provides a foundation for future optimizations to improve adaptability and extend application domains.</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

Multi-core Edge Computing with Deep Neural Networks for Real-Time Fire Detection

  • 期刊论文

  • 发表日期：

    2026年4月21日

查看全文
  • Gang Xue,
  • Jiang Wang,
  • Bo Gong

摘要

Edge computing emerges as the optimal choice for fire detection under extreme conditions with the advent of nascent visual technologies. This paper proposes an efficient deployment of the YOLOv7 model for real-time fire detection on a custom multi-core edge computing platform. The system utilizes 36 ARM cores, organized into clusters, and integrates FPGA-based optimization for low-latency, high-throughput image processing. YOLOv7 is trained on a fire image dataset, featuring diverse fire scenarios, and is partitioned across multiple cores to balance computational load and enhance energy efficiency. The deployment strategy incorporates modular architecture, including feature extraction, detection, and output modules, to optimize system performance. Experimental results show that the system achieves an inference speed of 11.5 FPS with an average latency of 86.29 ms, meeting the real-time processing demands of fire detection in edge environments. The platform demonstrates a balanced trade-off between computational throughput, resource utilization, and energy consumption, with dynamic power consumption averaging 3.902 W. These findings highlight the system’s suitability for large-scale fire monitoring applications in resource-constrained environments, and the study provides a foundation for future optimizations to improve adaptability and extend application domains.