An Efficient Real-Time Fire Detection Method Using Computer Vision and Neural Network-Based Video Analysis

Automatic fire detection in live video shots is an active research topic as the demand for a fast and accurate response to fire accidents is high. The fire detection task is still a challenging problem by leaving enough room for further improvement. Various fire features, different weather conditions, and complicated backgrounds are some of the complications involved in the fire detection task. This paper presents an accurate and fast fire alarm system for fire detection in live video, using a framework that consists of five stages, namely (i) video preprocessing that includes two steps; the first step is dynamic sampling for selectively processing the frames likely to contain a fire and then compressing the selected frames, (ii) GoDec framework for static background subtraction to extract the moving parts, (iii) color-scale-based filtering for finding the most probable fire region, (iv) YOLOv3 framework for fire localization, and (v) fire alarming system. The effectiveness of the proposed model was tested and confirmed on various datasets and evaluated in terms of accuracy. The proposed model can be implemented in stationary CCTV cameras for fire detection with low computational cost.