Non-Rigid Feature Extraction Methods in Real Time Forest Fire Detection Algorithms

In this paper a smoke detection algorithm for real time forest fire detection is proposed. The scene complexity in open-air environment and the non-rigid nature of the smoke cause high false positive alarms in many detection algorithms. To increase the efficiency of the algorithm, a multi features smoke approach is presented in this work. To segment the possible smoke regions, a change detection method is applied to the image. Afterwards, static and dynamic features of smoke are analyzed. Merging the extracted smoke features and applying morphological processes, region(s) with the highest probability of having smoke pixels are extracted. Two complementary texture features, Gabor filter and Local Binary Pattern (LBP), are applied to the input images. The input image sequence are first characterized by bank of Gabor filters covering the spatial- frequency domain. As multichannel filtering approach, Gabor filters extract features at different orientations and scales. By segmenting the energy image, smoke candidates are extracted and examined using An eXtended Center-Symmetric Local Binary Pattern (XCS-LBP). The image is converted to an array of integer labels as feature vectors for further analysis on smoke and non-smoke classification. The smoke area shows a blurred and smooth texture characteristic compared to the non-smoke areas. This criterion is examined using the histogram of LBP. A number of 5000 labelled smoke blocks are applied to the XCS-LBP operator and the average histogram is calculated and normalized as a priori variable. Applying spectral analysis, a fuzzy logic decision process is implemented on a chromatic analysis enhanced in the HSI (Hue-Saturation-Intensity) color mode. To define the fuzzy rules, empirical analysis is applied on a set of image data. A trial and error method is then used to reduce the failures. The algorithm is tested on a natural scene forest fire data set, collected from three different sits in Germany. Experimental results show high performance accuracy in smoke classification