FCMI-YOLO: An efficient deep learning-based algorithm for real-time fire detection on edge devices.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Internet of Things

Background:

• Deep learning and Internet of Things (IoT) enable vision-based fire detection on edge devices.
• Hardware resource constraints create a trade-off between accuracy and inference speed for these algorithms.

Purpose of the Study:

• To propose FCMI-YOLO, a real-time fire detection algorithm optimized for edge devices.
• To address the accuracy-speed trade-off in resource-limited environments.

Main Methods:

• Introduced the FasterNext module for reduced computational cost and enhanced precision.
• Incorporated Cross-Scale Feature Fusion Module (CCFM) and Mixed Local Channel Attention (MLCA) for improved small fire target detection and reduced resource use.
• Utilized Inner-DIoU loss function for optimized bounding box regression.

Main Results:

• FCMI-YOLO achieved a 1.5% increase in mAP@50.
• Reduced model parameters by 40% and GFLOPs to 28.9% compared to YOLOv5s.
• Demonstrated practical value for real-time fire detection on edge devices.

Conclusions:

• FCMI-YOLO offers a practical solution for real-time fire detection on edge devices.
• The algorithm effectively balances detection accuracy and inference speed under hardware constraints.
• The proposed methods contribute to efficient deep learning model deployment in IoT applications.