Real-time fire detection algorithm on low-power endpoint device

Timely detection of fires can significantly reduce the damage caused by them. Most contemporary fire detection algorithms are designed for high-performance computing devices, leading to slow execution on endpoint devices. To address this issue, we explore an enhanced method based on YOLOv8 for rapid...

Full description

Saved in:
Bibliographic Details
Published in:Journal of real-time image processing 2025, Vol.22 (1), p.29
Main Authors: Peng, Ruoyu, Cui, Chaoyuan, Wu, Yun
Format: Article
Language:English
Subjects:
Accuracy
Algorithms
Computation
Computer Graphics
Computer Science
Cost reduction
Damage detection
Deep learning
Fire damage
Fire detection
Floating point arithmetic
Image Processing and Computer Vision
Image resolution
Machine learning
Multimedia Information Systems
Pattern Recognition
Sensors
Signal,Image and Speech Processing
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Timely detection of fires can significantly reduce the damage caused by them. Most contemporary fire detection algorithms are designed for high-performance computing devices, leading to slow execution on endpoint devices. To address this issue, we explore an enhanced method based on YOLOv8 for rapid and efficient fire detection on endpoint chips equipped with Neural Processing Units (NPUs). Our algorithm is designed with consideration for factors not accounted for by indicator Floating Point Operations (FLOPs), as well as the unique characteristics of endpoint chips. We explore how different activation functions impact model execution speed and utilize quantization-friendly activation functions. We introduce a reparameterization module to address FLOPs limitations and optimize execution speed by aligning the input image resolution with sensor-captured images. Additionally, we use transfer learning techniques to improve model accuracy. Experimental results on the D-fire dataset indicate that our designed models offer advantages in terms of fast execution speed and competitive accuracy. Tests conducted on the endpoint chip reveal that the single-frame detection latency of our proposed two sizing models is 46 and 21 ms, respectively, which is significantly lower than that of YOLOv8, thereby placing minimal strain on the computing system. Furthermore, the accuracy is only slightly reduced compared to YOLOv8.
ISSN:1861-8200
1861-8219
DOI:10.1007/s11554-024-01605-7