Smoke Detection through Motion-Aware Attention-Driven

Fire is regarded as one of the most probable disasters, with significant natural and social consequences. Therefore, it poses a significant risk to human life, wildlife, and forest ecosystems. To effectively mitigate the consequences of fires, early smoke detection is particularly important. Smoke, as an early indicator of fire, exhibits various shapes, colors, and textures that can assist in fire prediction. Thus, this chapter introduces an innovative smoke detection method that integrates optical flow features with a deep learning framework. The classification of smoke is achieved by leveraging its motion-based, or dynamic features. The motion features of smoke are extracted through optical flow. The motivation for deploying optical flow color maps instead of images is to accurately determine the growth rate and its extent. The estimated optical flow color maps are fed into the proposed model for classification purposes. The modified DenseNet-121 architecture incorporates attention modules, including channel attention and spatial attention, which enhance the network’s overall performance. The presented framework also employs the Grad-CAM as well as the Guided Grad-CAM techniques for localization and visualization of smoke regions in the images. Different evaluation metrics are used for comparative analysis.