Three-Dimensional Density Estimation of Flame Captured From Multiple Cameras

Optical combustion measurement and analysis systems using multiple sensors have received considerable attention. In particular, the image-based flame 3D reconstruction approaches using computerized tomography have been widely applied for the flame 3D reconstruction from a set of views by constructing the optimized linear combinations of the 3D scene and projected images. Previous techniques were easily computed but were weak against noise and blurring due to the underlying least square-based loss function. This paper presents a 3D density flame reconstruction method, captured from the sparse multi-view images, as a constrained optimization problem between the flame and its projected images. For effective estimation of the flame with a complicated structure in an arbitrary viewpoint, we extract the 3D candidate region of the flame and, then, estimate the density field using the compressive sensing. The objective function is a linear combination of the photo consistency cost and sparsity regularization terms, which avoids blurring in the reconstruction. The proposed approach is a powerful matrix factorization method with each voxel represented as a linear combination of a small number of basis vectors. The approach also effectively simplifies the reconstruction process and provides the whole 3D density field in one step. The experimental results verify that the proposed 3D density estimation performs favorably from the few flame images.