ABSTRACT
Hundreds of fires break out every day around the world, resulting in thousands of causalities, serious threat to property safety and large areas of forest vegetation damage. Therefore, the real-time monitoring of fire plays an important role in protecting any building, forest, office or workplace etc. The early detection of fire is the major key to reducing the loss caused by fire because, once the fire spreads, it becomes extremely difficult to be controlled. The early detection of fire is performed based on smoke, which is small at the beginning, has different colors, shape and textures and can be seen easily from long distances by surveillance cameras.
This work introduces a convolutional neural-network-based smoke detection framework for fire prediction. The smoke detection is performed with the help of the dynamic and static features of smoke. The dynamic features are considered in the form of optical flow, while static features are generated from smoke color characteristics. The motivation of this work to use fractional order optical flow instead of images is to provide the precise location and rate of growth. The estimation of optical flow is carried out by formulating a fractional order variational model, which is capable of preserving dynamic discontinuities in the optical flow. Optical flow helps to find the active region of the images. The estimated optical flow color map is further dissected into its red, green and blue channels, and the channel with more sensitivity towards smoke motion is segmented with the help of a binary mask. Moreover, a novel convolutional neural network architecture based on global average pooling and dropout layer is also designed for the classification of segmented regions. Different accuracy metrics are considered for performance evaluation and comparison with other techniques.
