CNN-Based Video Surveillance for Fire and Localization Detection

Abstract

Convolutional neural networks (CNNs) have shown success in picture classification and other computer vision issues. The use of CNN in the fire recognition technique will greatly enhance detection accuracy, resulting in fewer fire tragedies and societal and environmental consequences. However, as inference requires a lot of memory and computing power, a significant problem is implementing CNN-based fire sensing devices in an actual video network. We describe a new, energy-efficient, Fire detection, location, and semantic understanding using a computationally efficient CNN architecture scenario based on the Squeeze Net design. It applies compact convolutional kernels and avoids huge, completely linked layers to save computational load. Despite its minimal processing requirements, results of the experiment show that our suggested approach achieves accuracy = 99.7%, F1-score = 98.49%, Precision = 98.99%, and Recall = 98.00%. Furthermore, by considering the specific characteristics of the circumstance at hand as well as the variety of fire data, the study shows how the efficiency and accuracy of the fire detection model.