Real-time Video Fire Detection via Convolutional Neural Networks

Existing fire alarm mainly relies on traditional smoke and heat-based systems, yet their monitoring range is limited, which leads to serious security risks. Therefore, this paper proposes a novel method for video fire detection, in which the suspected areas are firstly segmented via a cascaded judgement by using dynamic segmentation, color model and Gaussian model, and then the obtained suspected regions are input into a transfer learning-based CNN model for fine judgement whether it is a true flame. Since pre-processing methods can filter out large non-flame area, and which can greatly optimize calculation of CNN model. We finally migrated the proposed approach into embedded IMX.6 platform and test on the collected dataset. The experimental results show that, the algorithm can achieve a detection of 15fps and more than 90% recognition rate with an input of 960*720.