Edge Vision for Wildfires: A Comprehensive Investigation

Abstract

Wildfire monitoring needs fast, reliable, and scalable systems that can work in different terrains and lighting conditions. Traditional methods such as satellites, towers, and patrols cover large areas but often have delays, errors, and problems with smoke or obstacles. This review introduces recent computer vision progress for early wildfire detection and presents it as a practical pipeline focused on real-time use, stability, and easy operation. The first part reviews image-based screening and localization. Lightweight CNN models can quickly filter fire-related scenes and are improved through continual learning. The Learning-Without-Forgetting (LwF) method keeps new models consistent with old ones without saving past data, preventing forgetting. This helps maintain accuracy for small fires and allows frequent updates that follow seasonal or sensor changes. For localization, anchor-free detectors such as YOLO are efficient, combining good accuracy and speed. Training uses data augmentation and hard-negative samples like fog or clouds, and quantized models run on UAVs or tower devices. Temporal models using short video windows help reduce false alarms by learning fire motion patterns. The second part focuses on pixel-level mapping and system design. Multimodal fusion of RGB and infrared data improves segmentation under smoke or low light. Efficiency is achieved through model pruning, quantization, and adaptive inference. Evaluation includes not only accuracy but also detection speed, alert stability, and energy use, connecting perception with fire prediction and emergency planning.