An Evaluation and Embedded Hardware Implementation of YOLO for Real-Time Wildfire Detection

Abstract

With the constant threat of wildfires, the need for immediate and efficient detection methods is ever-increasing. Current wildfire detection methods from agencies such as FIRESafe Marin and CalFire use human operators to monitor many camera feeds constantly, which can lead to fatigue and inaccuracy. Machine learning, which can be more accurate at predicting outcomes, allows for more flexibility than image processing-based methods, and a better scalability when deployed on embedded devices. This work explores the performance of YOLOv5 (You Only Look Once, version 5) on embedded systems such as Raspberry Pi 4 for real-time wildfire detection. YOLOv3 and YOLOv3-tiny are also implemented on embedded devices for a performance comparison. Experiments show that our system has high detection accuracy and excellent battery life, which make the design suitable for real-world wildfire detection applications.