A combined real-time intelligent fire detection and forecasting approach through cameras based on computer vision method

Fire is one of the most common hazards in the process industry. Until today, most fire alarms have had very limited functionality. Normally, only a simple alarm is triggered without any specific information about the fire circumstances provided, not to mention fire forecasting. In this paper, a combined real-time intelligent fire detection and forecasting approach through cameras is discussed with extracting and predicting fire development characteristics. Three parameters (fire spread position, fire spread speed and flame width) are used to characterize the fire development. Two neural networks are established, i.e., the Region-Convolutional Neural Network (RCNN) for fire characteristic extraction through fire detection and the Residual Network (ResNet) for fire forecasting. By designing 12 sets of cable fire experiments with different fire developing conditions, the accuracies of fire parameters extraction and forecasting are evaluated. Results show that the mean relative error (MRE) of extraction by RCNN for the three parameters are around 4-13%, 6-20% and 11-37%, respectively. Meanwhile, the MRE of forecasting by ResNet for the three parameters are around 4-13%, 11-33% and 12-48%, respectively. It confirms that the proposed approach can provide a feasible solution for quantifying fire development and improve industrial fire safety, e.g., forecasting the fire development trends, assessing the severity of accidents, estimating the accident losses in real time and guiding the fire fighting and rescue tactics.

Automated summary

This paper discusses a method of real-time intelligent fire detection and forecasting through cameras, using two neural networks for feature extraction and prediction, and evaluates its accuracy through experiments.