Research On Method Of Early Flame Detection Based On Infrared Image Dynamic Features

Timely detection and alert of fire is crucial to reduce the loss of life and property caused by fire disaster.Compared with traditional fire sensors,video-based fire detection methods are more convenient,flexible and have wider range of monitoring.While providing sufficient fire information which helps a lot in fire control.To distinguish real fire and possible disruptors,infrared image sequences are used to analyze the features of shape deformation of flame and other disruptors.Based on the time domain distribution of the target contour pixels,an infrared flame recognition method was proposed.Experiments show that the precision ratio could reach 80.5%.The main research works involved are as follows:Combined with the relevant literature on video-based fire detection,features and detection method of flame with different scales and scenes are analyzed.A plan combined unmanned aerial vehicles,infrared image capture and processing model was determined.Possible flame regions are segmented with region growing and gray level thresholds.Individual targets from framed are saved and processed as sample sequences.Plenty infrared image sequences of real flame and disruptors were analyzed to find credible features of flame.Since only temperature distribution is reflected in infrared images,shape deformation should play a more important role in flame detection compared with visible images.Traditional geometrical descriptions of deformation are lack of accuracy while tested with various flame and disruptors.To describe the deformation process of flame more accurately,this paper presents a recognition method based on contour distributions of individual targets in sample sequences.Compared the distribution difference of contour points in different position during a period of time.It is proved that the deformation of flame is asymmetric and could be described as an unimodal distribution model.While most disruptors have significantly irregular behaviors.K-principal curve and normal pixels template are used to describe the unimodal distribution.Then a 9-dimensional feature descriptor is constructed to describe the distribution,also used as the input of SVM to recognize flame.Compared with geometrical descriptors,the false alarm rate is reduced by 16.9%.In addition,the local behavior of contours is analyzed further with connection domain behaviour.An additional feature is constructed to help recognizing flame with enough shape details.A Hidden Markov Model with dynamic parameters is applied to estimate the flicker frequency of the chaotic process of flame.The flicker procedure of flame is divided into three hidden states: stretching,shrink and a waiting state.To assess the deformation state,a direction-invariant and scale-invariant measure based on the limitations of contour distribution is built.A dynamic state transfer probability matrix is constructed according to the flicker procedure.Then the flicker frequency could be estimated with a decoding procedure of contour samples.