Optical Flow-Based Motion Estimation Of Complex Flows

Flow visualization technology,such as particle image velocimetry(PIV)or laser induced fluorescence(LIF),can extract velocity fields of fluid flows from image data.As this kind of tech-niques is able to provide non-intrusive,global and quantitative measurements of flow motion,it is widely used in the research of experimental fluid mechanics,process industry and many other fields.For complex flow phenomena(e.g.,turbulent flow),flow visualization and measurement allow the researchers to get a deeper insight into the flow patterns and provide experimental evi-dences for theoretical fluid mechanics.However,the traditional image velocimetry algorithms(i.e.,cross-correlation and optical flow methods)cannot fully meet the requirements of the complex flow motion estimation.On the one hand,the cross-correlation method has disadvantages in estimation accuracy and spatial resolution.On the other hand,the traditional optical flow approach is sensitive to illumination noise and the computation is quite time-consuming.In addition,the basic assump-tions of correlation-based and optical flow methods do not take into account the physical property of fluid flows.In order to overcome these shortcomings,several image velocimetry methods are proposed in this thesis,including a variational optical flow method based on dynamic illumina-tion equation,a variational optical flow method based on stochastic turbulent transport equation,a fluid motion estimation algorithm based on deep neural networks,and a fluid motion estimation algorithm based on data assimilation framework.These novel techniques tend to improve the ro-bustness,accuracy,space-time resolution as well as the real-time performance of the fluid image velocimetry.In summary,the main contents and innovations of this thesis are listed as follows.1.A dynamic illumination optical flow(DIOF)algorithm is proposed for particle image ve-locimetry.Due to the uneven intensity of laser irradiation or external illumination changes,the brightness of the two successive images obtained by PIV experiment tends to be in-constant.As the traditional optical flow method is designed based on brightness constancy assumption,it is sensitive to these illumination changes.In this thesis,the dynamic illumi-nation equation is applied to replace the optical flow constraint of Horn&Schunck(HS)method.In addition,the multi-resolution algorithm is used to optimize the objective func-tion of the variational optical flow,which allows to improve the estimation accuracy of large displacement.Compared with the traditional HS method,the proposed algorithm is more suitable for real PIV images with illumination noise.2.An optical flow method with stochastic transport equation(OF4STE)is p.roposed.The tra-ditional variational optical flow method do not consider the physical model of fluid flow.Therefore,the estimation accuracy of the complex velocity fields is not high enough to some extent.To fix this problem,the optical flow is combined with turbulence model in(his the-sis.First,the stochastic transport equation is derived from the location uncertainty principle.Then,the data term and the regularization term of the objective function are proposed based on the stochastic model.The multi-resolution algorithm is also applied to improve the per-formance.Moreover,while the cross-correlation and optical flow methods are difficult to obtained good results for LIF scalar images,the proposed algorithm is applicable to both particle and scalar images.3.A fluid motion estimation algorithm based on deep neural networks is proposed.With the development of deep learning,it is possible to solve the problem of fluid image velocimetry by using convolutional neural network(CNN).In this thesis,the deep learning technology is innovatively applied to the PIV experiment.Specifically,two PIV neural networks are proposed based on FlowNetS and LiteFlowNet,respectively,which are used for optical flow estimation.The input of the networks is a particle image pair and the output is a global ve-locity field.In addition,a PIV data set is artificially generated for CNN training,which takes into account the physical properties and the image noise.The proposed CNN models are verified by a number of assessments and in real PIV experiments such as turbulent boundary layer.Without loss of precision,the computational efficiency is greatly improved compared with the variational optical flow method.This advantage provides possibility for real-time flow measurement and control.4.A fluid motion estimation algorithm based on data assimilation framework(EnVar+OF)is proposed.The traditional variational optical flow estimates an average velocity field between two successive images,with no guarantee to recover a consistent motion trajectory over the whole sequence.To solve this problem,an ensemble-based variational data assimilation(EnVar)framework is adopted for fluid motion estimation.In such framework,the physical model of the fluid flow is taken as the strong constraint of the objective function,while the images are regarded as measurement samples.The spatiotemporal evolution of the fluid motion can be calculated based on the initial conditions.Compared to the traditional methods which extract velocity from two images,the proposed approach can provide the continuous velocity trajectory from multiple images.