Smoke detection in foggy surveillance environment using parallel vision transformer network

Recent years have seen an unprecedented increase in fire incidents, resulting in severe damage to forest regions, loss of human and animal lives, and unwarranted displacement of people. Owing to these issues, artificial intelligence-based fire detection systems have been developed. These systems mostly work with camera-based inputs for fire detection and alarm generation. As fire-smoke can be seen from a distance, methods for smoke detection have also been deployed for identifying fire incidents. But detecting smoke in an outdoor environment can be challenging due to weather conditions such as haze, fog, or clouds. These challenges have led to advancements in the field, and various deep learning architectures have been suggested over the years for fire-smoke detection in normal and foggy or cloudy weather conditions. Many existing methods are either computationally expensive and demand high memory, or if they are lightweight, they suffer from lower smoke detection accuracy and high false alarms. In the present paper, a dual-channel vision transformer model ‘SmokeViT’ is proposed for detecting fire-smoke in an outdoor environment. These channels are interlaced with a convolutional neural network to improve the feature learning capability of the model. ‘SmokeViT’ shows a remarkable accuracy of more than 99% and false alarm rate below 0.20% on two publicly available smoke datasets, outperforming six state-of-the-art methods. The model architecture has only 0.4 million parameters and 0.17 giga floating-point operations, making it suitable for deployment on resource-constrained surveillance devices.