Research On Video-Oriented Flame Detection And Tracking Algorithm

The frequent occurrence of fire directly brings unexpected disasters to humans.So,realizing high-efficiency monitoring and real-time warning of fire have always been an urgent problem in the current society.Recently,fire detection technology based on video-image has rapidly developed.Such detection technology has the advantages of high sensitivity,wide applicability,fast response speed and low cost,which is an effective method to fire prevention.However,due to the complexity of the fire scene,the interference of the similar light source and irregularity of flame motion,the current flame detection and tracking algorithms can not meet the practical requirement of accuracy,robustness and real-time simultaneously.Therefore,a series of algorithms for flame detection and tracking based on video sequence are proposed and improved in this paper,which are illustrated as follows:(1)Aiming at the problems of the existing false detection and miss detection in the flame segmentation algorithm and the slow convergence of the Chan-Vese model,an image segmentation algorithm combining flame color prior information and improved CV model is presented.First of all,the position of the initial contour curve of the CV model is set according to the candidate flame region obtained by the preliminary segmentation of the YCbCr color model.Then,the local weighted average and windowed filtering techniques are introduced to effectively suppress noise interference and reduce redundant contours.Ultimately,fast and effective flame segmentation is achieved.(2)In order to improve the accuracy of detection and tracking,the static and dynamic characteristics of the video flame are analyzed and extracted in detail.First,the morphological features of the flame are extracted based on the growth change of the flame region between successive frames.Then the number of brightness jumps of the flame pixels over a period of time is counted to analyze the flame flicker frequency,and a novel bag-of-words model is introduced to describe the characteristics of flame motion.Finally,it is proved that these characteristics are reasonable and effective as flame recognition criterion,which lays a foundation for flame tracking.(3)This paper propose a fast object tracking algorithm based on particle filter and regional information fusion feature to predict the spread of fire and achieve flame tracking location accurately.Firstly,a particle filter algorithm based on weighted colorhistogram is used to predict the flame target position.Then multi-feature extraction of dynamic flame is performed in the tracked region,and the classification method via multiple expert systems is introduced to fuse multiple features for determining flame target.Simultaneously,both the color feature and dynamic fusion feature of the flame are used as complementary observation information to update the weight of the sampled particle.Finally,a sliding window method is applied to establish data association and region matching between successive video frames in order to achieve fast and accurate flame tracking.In this paper,the testing experiments are carried out using the fire and interference video from different complex scenarios.Compared with similar algorithms,the simulation results show that the proposed algorithms have obvious improvement in terms of segmentation speed and detection accuracy as well as tracking performance.