Fire Surveillance Method Based on Quaternionic Wavelet Features

Abstract

Color cues are important for recognizing flames in fire surveillance. Accordingly, the rational selection of color space should be considered as a nontrivial issue in classification of fire elements. In this paper, quantitative measures are established using learning-based classifiers to evaluate fire recognition accuracy within different color spaces. Rather than dealing with color channels separately, the color pixels are encoded as quaternions so as to be clustered as whole units in color spaces. Then, a set of quaternion Gabor wavelets are constructed to establish an comprehensive analysis tool for local spectral, spatial and temporal characteristics of fire regions. Their quaternion Gaussian kernels are used to represent the spectral distribution of fire pixel clusters. In addition, a 2D band-pass filter kernel contained in the quaternion Gabor extracts spatial contours of fire regions. Another 1D temporal filter kernel is enforced to capture random flickering behavior in the fire regions, greatly reducing the false alarms from the regular moving objects. For early alerts and high detection rate of fire events, smoke region is also recognized from its dynamic textures in the proposed fire surveillance system. Experimental results under a variety of conditions show the proposed vision-based surveillance method is capable of detecting flame and smoke reliably.