Prediction of Fire Signatures Based on Fractional Order Optical Flow and Convolution Neural Network

Abstract

As we know that thousands of indoor and outdoor fires break out every day in different parts of the world, which result in a large number of serious causalities and threat to property safety. Therefore, it becomes of extreme importance to detect the fire in its very early stage, because once it is spread it becomes disastrous and difficult to control. The early detection of fire is associated with smoke, which is small at the beginning and have different colors, shape and textures. The initial stage of smoke can be seen easily through digital cameras installed in many locations. This paper proposed a smoke detection algorithm based on dynamical features of smoke using convolutional neural network (CNN). The dynamical features are considered in the form of optical flow color map. The estimation of optical flow is performed based on a fractional order variational model, which is capable in preserving dynamical discontinuities in the optical flow. Optical flow helps to find the active region of the images (video). The estimated color map is further dissected into its RGB channels and the channel with more sensitivity towards smoke motion is segmented with the help of binary mask. Finally, the segmented optical flow color maps are fed into a random forest based CNN architecture and a proposed ensemble learning based CNN architecture. Different accuracy metrics are considered for performance evaluation and comparison with other techniques. A variety of datasets consisting of 10 smoke (4576 frames) and 10 non-smoke (3219 frames) videos are considered for experiments.