Video Flame Detection Based On Multiple Features

Fire is the most hazardous natural calamities. As it's unpredictable and quick combustion, it affects natural environment and social production life. Efficient fire detection algorithm can detect fire in the early stage and make an alarm immediately to avoid huge fire disaster, protect the natural environment, residents' life and property, which have great academic value and practical significance.Traditional fire detection methods use temperature sensors, light sensors, which is low-cost and simple. However, they only can detect fire in a limited distance and affect seriously by illumination. Video fire detection algorithm can not only detect in a large range, but also can locate the position and size information. So, video fire detection methods get more and more attention. To search a high accurate?low error rate video fire detection algorithm, this paper does the following research.The suspect fire area segmentation. Quick and correct segmentation of the candidate area can reduce the detect range and lay a good foundation for the following detection. Efficacious pretreatment can reduce noise and retain the useful information. This article compares different de-noising methods and chooses adaptive median filter. Motion detection includes frame difference and background difference, optical. The paper analyses them in detail. With respect to fire detection, the paper gives different color models and chooses RGB and HSI color space model.Fire feature extraction. Fire has flicker characteristic, especially in frequency domain. Wavelet can describe frequency information correctly. Traditional wavelet transform has the shortcomings of shift variance and small direction options. DT-CWT is introduced for it has feature of shift invariance and six direction options of ±15?±45?±75. In addition, fire has unique texture. Local binary pattern is used for describing fire texture as its illumination invariance and simple computation. The paper compares LBP?rotate-invariance LBP and DRLBP(dominant rotated LBP) and combines it with fire character, then uses modified LBP as the texture feature. Besides wavelet and texture, color moments are also used as fire feature for its simple and effective advantages.Multi-feature video fire detection system. The paper proposes a shape invariance algorithm based on block feature. Motion and color is used to get the candidate fire block. Extract the features of DT-CWT, DRLBP and color moments. SVM can be used to train fire and non-fire samples. The trained SVM is used to test fire. Moreover, temporal verification is applied to exclude objects similar with fire(such as car light, street lamp and objects with reflection light) and verify fire finally.At last, make experiments with the proposed algorithm and other algorithms. The conclusion can get that the proposed method has higher detection rates, lower error rates in fire videos of different scene, illumination and background and non-fire videos of different disruptors. It is a high-performance, real-time, widely used video fire detection algorithm.