Research On Flame Video Detection Of Forest Fire Based On Multi-feature

It will be hard to put out the fire if forest fire breaks out, which will not only cause a great economic loss, but also cause serious damage to the ecological balance. Using video images to monitor the forest has gradually become an important means of forest fire prevention and monitoring. Forest fire monitoring method based on video can well remedy defects of the traditional forest fire monitoring method. Video surveillance can not only monitor forest for 24 hours to obtain continuous and real-time data, but also greatly reduce the forest fire monitoring time and effectively improve the accuracy of forest fire monitoring.Because forest surveillance environment is a large wilderness space, the swing of leaves causes a lot of moving objects in video images, and strong sunlight, yellow leaves of autumn and red maple leaves cause false alert of fire detection. These features make current fire video detection methods that are used in indoor or static scene are unsuitable for forest fire detection. Flame is an important manifestation of forest fire. In this paper, flame feature vectors are extracted by static features and dynamic features considering large granularity based on video clip. Flame video detection recognition of forest fire based on AdaBoost-BP algorithm is proposed.Firstly, the objects in forest scene are mostly the green trees or grass, and color range of flame at the early stage is red to yellow, so flame color features can be used to exclude a large number of non-fire color videos. However, some flame-color objects such as yellow leaves of autumn, red maple leaves and lamplight in woodlands can’t be filtered only via color features. Considering flame is constantly changing while fire burning, therefore moving features is utilized to further remove some relatively static flame-color objects.Secondly, there is a spreading process when fire breaks out. Fire video clips contain a number of consecutive video frame images. So forest fire monitoring videos are divided into spatio-temporal video blocks according to sliding time window. Static features, dynamic features, temporal features and spatial features are analyzed considering large granularity based on video clip. This paper mainly analyzes two kinds of static features(circularity, texture features) and three kinds of dynamic features(flame area variation, shape similarity, flicker frequency) to extract flame feature vectors.Thirdly, considering the shortcomings of conventional flame classification algorithm, this paper proposes flame video recognition model of forest fire based on AdaBoost-BP algorithm. AdaBoost algorithm is employed to produce a strong classifier by combining a set of weak classifiers—BP neural network. This paper focuses on the impact of the number of iterations of AdaBoost algorithm and hidden layer nodes of BP neural network to classification errors. And determine the optimum parameters of iterations of AdaBoost algorithm and hidden layer nodes of BP neural network. Then AdaBoost-BP method is compared with BP method.Finally, design and develop the forest fire flame detection software system. Visual C and MATLAB are used to realize mixed programming. Extraction of suspected flame regions, extraction of forest flame feature vectors and flame recognition based on AdaBoost-BP algorithm are integrated in the video surveillance platform.