Night-Fire Detection using Randomness Test

초록

A novel method is presented to detect the night-fire based on robust features and randomness test in this paper. Firstly, color probability is estimated based on the brightness since the fire pixels provide high brightness at night. After that, motion probability is obtained by adopting the background image dynamically updated with approximate median method. Furthermore, the color and motion probabilities are integrated to gain candidate fire, from which a feature vector is extracted for each frame. The successive feature vectors are then applied to randomness test for obtaining the prior fire probability. Finally, convolution is defined to update the prior probability for improving system reliability, and the posterior probability is employed to alarm. The presented method was successfully adopted to real-environment intelligent surveillance systems, and it is proved to be effective, robust, adaptive and efficient invariant to environment situation and camera quality.