Fire Detection Using Level Set Segmentation Based Fractional Order Optical Flow and 4D Fire Features with Mixed Data CNN-LSTM Model

As we are aware that the world witnesses a huge number of fire breakouts everyday, which results in high numbers of hazardous events and severe losses to property and forest vegetation. Therefore, early stage fire detection is of vital importance, for once it spreads it becomes unmanageable and disastrous. The early detection of fire can be performed with the help of vision based deep learning techniques. The novelty of the work lies in performing the fire detection using the static and dynamic features of fire. The static fire features are taken as shape, texture, and color, while the dynamic feature accounts for its flickering motion. For this purpose, the fire motion is estimated in terms of optical flow from videos (image sequences) by using a motion edge preserving level set segmentation based fractional order variational model. Level sets provide nicely segmented flow fields, while fractional order derivatives are capable to deal with discontinuities in the motion field. The estimated optical flow field is used to derive four fire features, which are constituted as 4D vectors. These 4D vectors reduce the data dimensionality and mitigates over-fitting problem. Finally, the fire detection is carried out by implementing a mixed data CNN-LSTM model. The mixed data presented in the work is composed of a reference image frame and the corresponding 4D vector sequence. Also, the significance of the model is manifested through an ablation study. The model performance validation is performed thorough a comparison study conducted with several existing models.