Study On Intelligent Video-based Fire Recognition And Dynamic Prediction

Nowadays,fire prevention and control have been the focused issue.The efficient fire prevention and control technology is a powerful guarantee to reduce the risk of fire accidents.The video-based auxiliary decision-making for fire prevention and emergency rescue developed significantly with the popularization of surveillance camera systems.However,video-based fire technology could only judge whether a fire occurs,and the complex model has a low recognition accuracy rate on the early fire and an insufficient extraction ability on the real-time fire image information,which severely restricts the engineering application of video fire technology.Therefore,it is key for the popularization and utilization of video-based fire technology that how to interpret and retrieve the real-time fire information based on the video data of fire scenes,and dynamically analyze and predict fire trends in advance based on the real-time disaster information.In this paper,to improve the current video-based fire technology,the distributed video fire detection model based on a multi-scale convolution neural network(CNN)is constructed,and the information inversion and dynamic analysis method for video-driven fire based on object-oriented segmentation is proposed.The real-time prediction model on the temperature of tunnel fire is developed employing the radial basis function(RBF)neural network,and the Kalman filter algorithm is used to assimilate data between the prediction model and real-time fire information.All the algorithms involved are programmed with Python.The contents and results are as follows:(1)A distributed video fire detection model based on multi-scale CNN is developed aiming at the small flame target and multi-scale characteristics of the initial fire in the video.The fire candidate areas are quickly located by employing the low-level visual features of flame,which promotes the detection efficiency of small flame target and broadens the detection range of model scale.A multi-scale convolutional neural network is constructed based on spatial pyramid pooling.The multi-scale expression of flame depth information is obtained,and the deep fusion of global and local flame features is achieved in the fire scene.A fast,accurate and robust network model is obtained by fully optimizing network parameters and training in a multi-scene and multi-scale data set.(2)A video fire information inversion and dynamic analysis model orienting object segmentation is proposed to acquire the fire information in real time.The flame candidate areas are segmented quickly and accurately by SEEDS superpixel and YCr Cb color space method.The real-time heat release rate can be retrieved by extracting the size characteristics of fire images following time combined with fire prior knowledge and camera calibration information.Furthermore,the model is successfully applied to three experiments with different fire types and scales: wood crib,diesel pool,and propane fire.The flame height,projection area,and volume are extracted over time,and the fire risk and development trends are evaluated.(3)A real-time prediction model on the temperature of tunnel fire is developed based RBF neural network,which improves the weak advance and poor real-time in the fire prediction.The big data of tunnel fire is constructed on the basis of numerical results of 100 cases with different fire sources power,position and wind speed taking an underwater tunnel as a prototype,the reliability of numerical data is verified by comparison with classical tunnel fire model.An RBF neural network prediction model on tunnel fire is developed based on the data set.Thus,the fire numerical model constraint of the known tunnel is blended in the prediction model and fire parameters at any time and location can be output at the millisecond level,which results in the rapid and relatively accurate prediction of tunnel fire temperature.The uncertainty of the AI model driven by CFD data is further analyzed.(4)In view of the complex and changeable fire actual situation,a data assimilation method of tunnel fire prediction model based on Kalman filter is proposed.Kalman filter algorithm is employed to fuse the RBFNN tunnel fire prediction model driven by CFD data with the measured real temperature.The input of prediction model is modified based on the actual heat release rate,wind speed and other fire parameters to promote the traditional neural network method and static mapping of numerical calculation.The fire state is estimated optimally by iteratively updating model parameters,integrating real-time fire information and correcting timely the prediction deviation.The model with high accuracy and advancement could provide reliable technical support for fire emergency rescue.