Identification Of Forest Fire Video Based On Spatio - Temporal Features

Forest is a kind of valuable resource for human. Therefore, it is of great importance to protect forest from fires as they can cause devastating consequences in a short time. In the past, many detection systems of fire and alarm are mainly based on the physical and chemical properties, but they have some drawbacks such as poor reliability and obvious time delay. In terms of the prevention of large areas, the forest fire alarm system based on video images has become a heated topic of forest fire prevention over the past years. And its main advantages are no restrictions of space and distance and fast response. However, smoke produced by forest fires is visible much before the flames. So this paper will focus on how to detect some accurately and quickly and send out an alarm in complex environment.Firstly, Pre-treatment of smoke images are proceeded, because the quality of image is directly related to the subsequent smoke region segmentation. By comparing it with the traditional histogram equalization, neighborhood average, frequency enha ncement, this paper presents a fast fuzzy logical algorithms of smoke image. At first RGB smoke image will be converted into HSV space. Then this paper selects the appropriate fuzzy operator and the iteration number to complete S and V fuzzy iterative enha ncements respectively. Finally, the whole enhancement process of image is achieved by converting images from HSV to RGB space.Secondly, suspected smoke area is segmented. As we know, the traditional algorithm of inter frame difference has some shortages, such as hollow phenomenon. So the paper will model real time background and compute differences between current frame and the real-time background. After that, suspended smoke regions are segmented based on RGB and HSV color space. Finally, mathematical morphology and connected domain analysis are used to refine the suspected area.Then, four flutter features including flutter direction, growth rate of smoke areas, complexity and background ambiguity are extracted. With these four features, the accuracy of smoke recognition.At last, this paper design a smoke recognition classifier based on BP neural network system containing four inputs, 1 hidden layer with 10 neurons, an output. Then, BP network is trained and tested by using the above four flutter features. A large number of videos including smoke and smoke-like videos are used to verify the correctness of feature selection and get a stable network system. In the end, the experimental results of different scenarios showed that the proposed algorithm can effectively and efficiently recognize smoke videos.