Optical Flow Methods For Tracer-Based Imaging Techniques In Fluid Mechanics

Abstract

This paper is aimed at assessing the performance of Optical Flow methods to estimate flow velocities from images designed for 2D Particle Image Velocimetry (PIV). A benchmark is proposed, based on synthetic tracer images associated to rotation- and deformation-dominated flow configurations. The investigated Optical Flow methods are Lucas-Kanade, Horn-Schunck and Farnebäck, combined them with the Liu-Shen method. The true values of the flow field (ground truth) are compared with the results of PIV and Optical Flow methods. Relative and absolute errors are computed for different combinations of flow type, relative velocity, white noise level, particle spot size and particle concentration. The accuracy of Optical Flow methods varies with the flow type, particle sizes and pre-processing steps. Combination of methods improves accuracy. The most accurate method is (dense) Lucas-Kanade/Liu-Shen. The Farnebäck/Liu-Shen combination is highly accurate for 10 or 12 bit images. The range of applicability of Optical Flow methods seems higher than that of correlation-based PIV.