Patch-wise periodical correlation analysis of histograms for real-time video smoke detection

In this paper, an approach for video smoke detection is proposed. The basic idea is smoke has a highly varying chrominance/luminance texture in long periods. Since smoke has no shape, it also creates high shape changes in long periods. In this paper, two kinds of histogram are employed to observe change in luminance/chrominance texture and shape. Linearly interpolated chrominance/luminance subtraction image is used as input image for periodical analysis after thresholding. Intensity histogram which consists of 256 bins and oriented gradients histogram with 8 bins are employed for this purpose. Smoke generally creates transparent textures in which histogram bins create high variations. By considering the algorithmic cost and nature of smoke, periodical normalized cross-correlation analysis is performed in histogram bins instead of two-dimensional image context which makes algorithm more speedy and efficient for smoke detection. Experiments with a large number of smoke and non-smoke video sequences give promising results.