Dynamic Smoke Detection Using Cascaded Convolutional Neural Network for Surveillance Videos

The extraction of stable smoke features in complex scenes is a challenging task for video based smoke detection. For this issue, a convolutional neural network (CNN) framework which employs both static and dynamic features of the smoke is proposed. On the basis of analyzing the static features of individual frame, we further explore the dynamic features in spatial-temporal domain to reduce the influence of the noise from environment. Experimental results show that the proposed cascaded convolutional neural network framework performs well in real-time video based smoke detection for complex scenes.