Forest fire visual tracking with mean shift method and gaussian mixture model

Abstract

Forest fire region from surveillance video is non-rigid object with varying size and shape. It is complex and thus difficult to be tracked automatically. In this paper, a new tracking algorithm is proposed by mean shift method and Gaussian mixture model. Based on moment features of mean shift algorithm, the size adaptive tracking window is employed to reflect size and shape changes of object in real time. Meanwhile, the Gaussian mixture model is utilized to obtain the probability value of belonging to fire region for each pixel. Then, the probability value is used to update the weighting parameter of each pixel in mean shift algorithm, which reduces the weight of non-fire pixels and increases the weight of fire pixels. With these techniques, the mean shift algorithm can converge to forest fire region faster and accurately. The presented algorithm has been tested on real monitoring video clips, and the experimental results prove the efficiency of our new method.