High Resolution and High-Speed Live Optical Flow Velocimetry

TL;DR

This work addresses the limitations of cross-correlation PIV by presenting a variational Optical Flow Velocimetry (OFV) method that yields dense, per-pixel velocity fields with high spatial resolution at real-time rates on a single GPU. It combines a Lucas–Kanade-based optical flow approach with a Gaussian-pyramid framework and intensity normalization to recover displacements at each pixel, achieving near-kilohertz post-processing and tens to hundreds of hertz in real-time for large image sizes. Synthetic benchmarks on Rankine vortices and HIT DNS validate accuracy and gradient resolution, showing OFV better captures sharp gradients and small-scale structures than CC-PIV, particularly with texture-rich seeding. The results demonstrate real-time capabilities (up to ~460 Hz for 4 MP and ~90 Hz for 21 MP) and substantial post-processing gains, enabling closed-loop flow control and rapid diagnostics in experiments and complex flows.

Abstract

Particle Image Velocimetry (PIV) is the most widely used optical technique for measuring two-dimensional velocity fields in fluids. However, with the standard cross-correlation (CC) algorithm, improving the spatial resolution of instantaneous velocity fields and obtaining dense velocity fields in real time remains challenging. Optical Flow Velocimetry (OFV) offers a way to overcome these limitations. In this study, we demonstrate that dense velocity fields (one vector per pixel) with high spatial resolution can be obtained in real time at frequencies up to thousands of Hertz using an optical flow approach. We show that high resolution is achievable with optimized seeding, and that computational speed can be increased by choosing appropriate parameters and running on a single GPU. Using this method, 21 Mp velocity fields can be computed in real-time at 90 Hz, while 4 Mp velocity fields can be computed up to 460 Hz. These measurements enable the computation of various flow quantities in real time, during the experiment. It makes this technique perfectly suitable for many new type of experiments, from closed-loop flow control experiments based on OFV measurements, to very low frequency measurements or monitoring of the flow to find rare events. They also greatly accelerate post-processing leading to potential large time and energy gain for post-processing.

Figures (24)

• Figure 1: Diagram showing the main steps used to compute the velocity vectors at the kernel scale.
• Figure 2: Theoretical displacement magnitude of the Rankine vortex for various radii and velocities. The core radius increases from left to right (from $r = 12$ to $r=150$ px) while the maximum displacement of the field increases from 8 to 32 px from top to bottom.
• Figure 3: Theoretical velocity profiles used to move the particles in the images. (a) Velocity magnitude profiles with a given magnitude ($|U|=8~px/frame$) and different core radii leading to variation of the spatial veolcity gradient. (b) Velocity profiles obtained for a given fixed radius ($r=25~px$) but for different amplitudes, also leading to variations of the velocity gradients. Both cases show zoomed-in profiles around the core's center.
• Figure 4: 3D representation of the selected plane from a velocity datacube snapshot from the DNS HIT dataset.
• Figure 5: Sample of DNS velocity fields. (a)$u(x,y)$. (b) $v(x,y)$.