Fire and smoke detection is a critical safety issue in industrial settings, but common sensor-based approaches—such as infrared thermal and smoke detectors—do not work well outdoors. A new paper on arXiv argues that this limitation makes vision-based methods necessary, and frames the problem as inherently spatiotemporal: fire and smoke change over both space and time.
The authors propose a detection approach that combines spatial features with temporal features. The system is built from three main parts, each designed to reduce the error of the previous stage, so the final model performs robustly. Transformer modules are also incorporated to boost accuracy, a choice that reflects the growing use of attention mechanisms in computer vision.
The paper reports that the proposed method performs well on the fire and smoke detection task. The authors suggest it could be used to improve workplace safety, particularly in outdoor environments where traditional sensors fall short. The work was published in a WASET conference and is available as arXiv:2610.20141.