Computer Vision-Based Wildfire Detection in Video: Deep Learning Using Motion Estimation

This study addresses the issue of false alarms in computer vision-based wildfire detection systems by exploring the use of motion estimation to enhance accuracy. A significant challenge with using a single neural network for wildfire detection is the misclassification of fog or clouds as smoke due to their visual similarities. By incorporating Gunnar Farneback optical flow for motion estimation following the initial wildfire detection, the study achieved promising results on non-wildfire videos, reducing the average false alarm rate from 5.69% to 0.40% while maintaining an acceptable true detection rate for wildfire videos. This integration of motion estimation, specifically through the Gunnar-Farneback optical flow method, proves to be an effective strategy for reducing false alarms in non-wildfire videos while sustaining an acceptable true detection rate for wildfire videos.