Fire Detection in Video Sequences using Optical Flow Estimation

Vision-based flame detection can be applied in large open spaces for fire detection, which has been achieved with camera surveillance systems. The motion estimator exploits some motion features such as color, shape and texture that are used to recognize fire and non-fire region. Classical systems applied optical flow estimation methods for flame detection which are based on the assumptions, for example intensity constancy and flow smoothness that are not met by fire motion and also they classify only the instances of dynamic textures, not between presence or absence of an event. So two optical flow estimators specifically designed for the fire detection task; optical mass transport exploits the dynamic texture of flames. Second non-smooth data optical flow models saturated flames. Then, characteristic features related to the flow magnitudes and directions are extracted from these two flow fields to discriminate between fire and non-fire motion. To classify the fire and non-fire region neural network is used as classifier that is just one of many possible choices for learning the class separation boundaries. This paper proposes support vector machine (SVM) which is a binary classifier that produce more accurate results than the previous classifiers. KeywordsOptical flow, optimal mass transport, dynamic texture, rigid motion, flame saturation.