ES-ATF: Early Smoke Detection based on Attention-aggregated Temporal Feature Extraction^*

Smoke detection is an area where strong research interest was put since its practical meaning in extinguishing fire in the early stage. Detecting smoke not only reduces the life or property loss but also has environment-protection value as inefficient combustion can be restricted. In the literature, there are bunches of investigations on single-frame smoke detection via YOLO-based methods, however, these methods are always attenuated by the transparency and imperceptibility nature of smoke. Therefore, more useful information that can work as an extra prompt to the model is expected. To strengthen the information density of inputs, one possible solution is to explore inter-frame correlation within a smoke snippet. With this added temporal information, a boost in model performance can be expected. This paper adopts an attention-aggregated temporal feature extraction method, by which, the inter-image feature can be better exploited, and more frames can be detected at one time. Consequently, the smoke detection process is speeding up while simultaneously, the model's capability is enhanced in terms of alarming a potential smoke before it evolves to the late stage.