Flame Combustion Analysis And 3D Reconstruction Of Spontaneous Combustion State Based On Optical Flow Method

Flame combustion analysis and flame 3D reconstruction have important applications.Common flame analysis methods are based on the characteristics of flame combustion,which requires a numerical value that is difficult to measure in real fire.Smarter methods include segmentation based analysis,which analyses the flame by analyzing the number of pixels in the image,which is too static and not very predictable.However,flame 3D reconstruction algorithms such as tomography,dense point cloud matching,texture matching and other methods have complex process and high cost problems,which can not be widely used in practical industry.This paper mainly studies the application of optical flow estimation in flame trend analysis and three-dimensional reconstruction task.The main research work is as follows:1)The optical flow of flame is used in flame combustion trend analysis and three-dimensional reconstruction.In the optical flow estimation of flame,the change of flame brightness will affect the brightness of surrounding pixels,which makes the surrounding pixels produce optical flow and adversely affects the optical flow estimation of flame.In this paper,the semantic segmentation algorithm is used to segment the flame first to solve this problem.In view of the problem that the current semantic segmentation network is not accurate enough for the flame data set,an improved Unet network based on attention mechanism is proposed.The MIOU of the improved network on the flame data set increases by about 2.8%and MPA increases by about 3.1%.2)In flame combustion trend analysis and three-dimensional reconstruction,the accuracy of optical flow estimation directly determines its accuracy.In the optical flow estimation of flame,the effect of traditional algorithm is not ideal,and the deep learning algorithm does not specifically target at the network of non-rigid motion.In this paper,the Loss function is improved in PWC-Net network to make it more suitable for non-rigid data sets,and the pre-training model of PWC-Net is used to fine-tune the non-rigid data set MPI-Sintel.Experiments show that the EPE error of the improved network is reduced by 0.19 and 0.22 in the pure version and the noise version,respectively.3)In the flame trend analysis,this paper proposes a sequential segmentation algorithm based on key point selection to analyze the flame light flow curve,and then study the flame combustion trend.This method can automatically select the key points to divide the flame optical flow curve,and use the slope of each stage curve fitting line to judge the combustion state,so as to analyze the combustion state of the flame.This method can achieve the task of monitoring forest fire or boiler flame combustion.4)Aiming at the problems existing in the traditional flame 3D reconstruction algorithm,this paper first proposes to use the relationship between optical flow and depth to transform the optical flow of the flame into depth,so as to reconstruct the flame surface structure.This method can abandon some problems existing in the traditional flame reconstruction algorithm.