Real-time smoke detection with Faster R-CNN

Abstract

Fires must be put out as soon as possible, because many economic losses and lead valuable lives may be caused. In particular, smoke detection, which is often earlier in a fire, is of great importance. There are many difficulties in finding a smoke detection algorithm on basis of images. This study introduces a novel smoke detection algorithm in real time on basis of fast R-CNN surveillance cameras installed in the factory, which reduces the false positive detection due to the irregular form of smoke. First, we calculate the similarity of the overall structure and the average square error (MSE) for the detection of the smoke movement of the entry surveillance camera. The regions that are candidates for the delay were extracted by a deep learning algorithm (faster R-CNN). Third, the space-time characteristics are used as a smoke region to finally determine the region that is candidates for the extraction. This study presents a new algorithm that uses spatial and temporal characteristics of global and local frames, which are good target materials to decrease false positives on basis of deep learning technologies. According to experimental results, the algorithm has good detection performance while maintaining smoke detection performance and reduces false alarm frequency with 99.0%.