A Robust Attention-based 3D CNN Model for Fire and Smoke Detection from Videos

Abstract

The unforeseen occurrence of a fire eruption has significant and adverse impacts on all aspects of human existence, including human life, land, industry, forests, and animal populations. The emergence of smoke before the apparent signs of fire makes smoke detection very advantageous in terms of promptly identifying fires. This early detection capability holds significant potential for preserving numerous lives and mitigating the occurrence of catastrophic events and associated losses. The problem of detecting smoke and fire using video-based methods presents significant challenges due to the considerable variability in their color, brightness, and shape. Most prior researchers conducted individual studies focused on the categorization of fire or smoke. Furthermore, video classification studies focused on a 2D approach, which was inadequate due to its inability to capture temporal information, so it was not able to learn the difference between frames. The convolutional neural network (CNN) has demonstrated exceptional performance in several domains, hence establishing itself as the leading approach. This study presents a robust classifier based on a three-dimensional convolutional neural network (3D-CNN) specifically developed to classify events into three distinct classes: fire, smoke, and neutral. The classifier is meant to effectively process video data. 3D models have the capability to acquire spatial and temporal characteristics from video sequences, rendering them appropriate for undertaking such tasks. The proposed 3D-CNN model had a notable test accuracy of 90% and showcased outstanding performance with an accuracy of 96% when subject to cross data [1] that have seven smoke videos. To enhance its capabilities, the model integrates the CBAM (Convolutional Block Attention Module) attention mechanism. The utilization of this approach allows the model to concentrate on significant regions of interest (ROI) throughout the entirety of the frame, hence enhancing its precision in categorizing complex attributes inside video frames. By integrating the attention mechanism, the model attained a remarkable gain of 6% in test accuracy with an accuracy rate of 96% and an F1-score of 98%.