Learning Deep Optical Flow Estimation for Particle-Image Velocimetry

• Applicant: Professor Dr.-Ing. Wolfgang Schröder
• Subject Area: Fluid Mechanics
• Term: since 2022
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 470744243

Particle-Image Velocimetry (PIV) is a very powerful and efficient measuring technique that provides deep insights into complex flow phenomena. Based on cross- correlations current algorithms compute the most probable displacement between two particle images. However, these sophisticated, manually designed methods also have limitations and require a decent level of prior knowledge to obtain good results. For instance, if the particle displacement is complex due to strong fluctuations, non- constant velocity gradients, or out-of-plane displacement, the performance of state- of-the-art algorithms significantly decrease. While the estimated mean field matches ground-truth data fairly well, velocity fluctuations are usually underestimated. A similar bias error can be observed for inhomogeneous distributed tracer particles which is typical in near-wall flows. More importantly, only sparse velocity outputs can be obtained due to the spatial averaging inherent to the cross-correlation.To overcome the limitations of cross-correlation based approaches, the Institute of Aerodynamics develops an innovative holistic approach to the PIV analysis based on new ideas of deep learning in optical flow applications going far beyond the current state-of-the-art in terms of spatial resolution, manual effort, and accuracy. Unlike existing methods, these proposed approaches are general, near-automated, yield per-pixel flow estimates and side-steps the problem of manually designing an analytical pipeline by defining an end-to-end network.The main focus of this proposal lies in the application and modification of different optical flow architectures to the field of PIV processing in realistic, experimental fluid flow measurements, their systematic evaluation with respect to various particle image conditions and the analysis of different learning paradigms. The task of processing realistic PIV images requires networks capable of computing accurate flow estimates for real fluid flows under varying particle and light conditions and goes far beyond ideal benchmark scenarios such as Sintel / KITTI test cases or particle images under perfect experimental conditions which are not particularly representative for fluid flow experiments as demonstrated in the literature.Based on a systematic analysis of all key components of supervised, unsupervised, and self-supervised deep learning approaches for optical flow estimation a novel neural method for general 2D PIV data processing will be introduced overcoming current shortcomings of cross-correlation based PIV algorithms while significantly increasing the spatial resolution of the velocity field.The suggested neural PIV method will be a game-changer in current PIV experiments since it connects computer vision methods with physics and engineering applications.

• DFG Programme: Research Grants