Real-time and On-site Aerodynamics using Stereoscopic PIV and Deep Optical Flow Learning

Mohamed Elrefaie, Steffen Hüttig, Mariia Gladkova, Timo Gericke, Daniel Cremers, Christian Breitsamter

Abstract

We introduce Recurrent All-Pairs Field Transforms for Stereoscopic Particle Image Velocimetry (RAFT-StereoPIV). Our approach leverages deep optical flow learning to analyze time-resolved and double-frame particle images from on-site measurements, particularly from the 'Ring of Fire,' as well as from wind tunnel measurements for real-time aerodynamic analysis. A multi-fidelity dataset comprising both Reynolds-Averaged Navier-Stokes (RANS) and Direct Numerical Simulation (DNS) was used to train our model. RAFT-StereoPIV outperforms all PIV state-of-the-art deep learning models on benchmark datasets, with a 68% error reduction on the validation dataset, Problem Class 2, and a 47% error reduction on the unseen test dataset, Problem Class 1, demonstrating its robustness and generalizability. In comparison to the most recent works in the field of deep learning for PIV, where the main focus was the methodology development and the application was limited to either 2D flow cases or simple experimental data, we extend deep learning-based PIV for industrial applications and 3D flow field estimation. As we apply the trained network to three-dimensional highly turbulent PIV data, we are able to obtain flow estimates that maintain spatial resolution of the input image sequence. In contrast, the traditional methods produce the flow field of ~16x lower resolution. We believe that this study brings the field of experimental fluid dynamics one step closer to the long-term goal of having experimental measurement systems that can be used for real-time flow field estimation.

Introduction

In experimental fluid mechanics, Particle Image Velocimetry (PIV) is the gold standard for visualizing airflow. Researchers seed the air with tracer particles—such as helium-filled soap bubbles (HFSB)—illuminate them with a laser sheet, and record their movement with high-speed cameras. However, traditional PIV relies on cross-correlation, which has two major drawbacks: low spatial resolution due to interrogation window averaging and significant processing bottlenecks that can delay results by weeks.

Experimental setup for on-site 3D car aerodynamics using Helium-filled soap bubbles (HFSB) and high-speed cameras.

While deep learning has recently improved 2D PIV, industrial applications require Stereoscopic PIV to capture 3D velocity components. This paper introduces RAFT-StereoPIV, bridging the gap between methodology development and real-world industrial 3D flow estimation, specifically targeting complex scenarios like the "Ring of Fire" (RoF) car wake measurements.

Methodology

The RAFT-StereoPIV architecture is based on the Recurrent All-Pairs Field Transforms (RAFT) framework originally designed for computer vision. It employs a shared Convolutional Neural Network (CNN) for feature and context encoding. The model computes visual similarity between every pixel in a frame pair, creating a 4D correlation volume.

The RAFT-StereoPIV architecture featuring feature/context encoders, 4D correlation volumes, and iterative flow refinement.

Key technical features include:

  • Iterative Updates: A convolutional gated recurrent unit (Conv-GRU) performs 12 iterative updates to refine the flow estimate, allowing the model to capture both small and large displacements.
  • 3D Reconstruction: A geometric mapping function (calibration) fuses the 2D optical flow pairs from two cameras into a final 3D velocity field $(U, V, W)$.
  • Efficiency: The authors also introduced RAFT-StereoPIV-Small, a compact version with 59% fewer parameters (2.2M vs 5.3M) that maintains robust performance.

Datasets and Training

The model was trained and validated using a multi-fidelity synthetic approach. Problem Class 1 (PC1) consisted of high-quality CFD-generated flows (e.g., DNS turbulence) used as an unseen test set. Problem Class 2 (PC2) was designed to mimic real-world experimental conditions, featuring reduced particle density, lower signal-to-noise ratios (SNR), and higher displacements. Real-world deployment was tested on data from the "Ring of Fire" and wind tunnel tests of a 25%-scaled Ahmed body vehicle model.

Key Results and Insights

RAFT-StereoPIV outperformed all existing PIV deep learning models, showing a 68% error reduction on the PC2 validation set and a 47% reduction on the unseen PC1 test set.

The most significant advancement is in spatial resolution. Traditional methods produce one vector for every ~16-48 pixels. In contrast, RAFT-StereoPIV produces a vector for every single pixel, maintaining the full spatial resolution of the input camera sensor—a ~16x improvement over commercial cross-correlation software. Furthermore, the model is computationally efficient; once trained, it processes a $1024 \times 1024$ image pair in approximately 40 seconds on a single NVIDIA A100 GPU, moving toward real-time on-site aerodynamic analysis.

Conclusion and Future Work

RAFT-StereoPIV successfully extends deep learning-based PIV to industrial 3D applications, providing high-resolution flow estimates that were previously unattainable in real-time. While raw predictions still contain some spurious vectors requiring post-processing, future directions include integrating noise and reflection removal directly into the neural network and exploring physics-informed loss functions to ensure predicted flows obey fluid dynamics laws like mass conservation.

Summary updated

Abstract

We introduce Recurrent All-Pairs Field Transforms for Stereoscopic Particle Image Velocimetry (RAFT-StereoPIV). Our approach leverages deep optical flow learning to analyze time-resolved and double-frame particle images from on-site measurements, particularly from the 'Ring of Fire,' as well as from wind tunnel measurements for real-time aerodynamic analysis. A multi-fidelity dataset comprising both Reynolds-Averaged Navier-Stokes (RANS) and Direct Numerical Simulation (DNS) was used to train our model. RAFT-StereoPIV outperforms all PIV state-of-the-art deep learning models on benchmark datasets, with a 68$\%$ error reduction on the validation dataset, Problem Class 2, and a 47$\%$ error reduction on the unseen test dataset, Problem Class 1, demonstrating its robustness and generalizability. In comparison to the most recent works in the field of deep learning for PIV, where the main focus was the methodology development and the application was limited to either 2D flow cases or simple experimental data, we extend deep learning-based PIV for industrial applications and 3D flow field estimation. As we apply the trained network to three-dimensional highly turbulent PIV data, we are able to obtain flow estimates that maintain spatial resolution of the input image sequence. In contrast, the traditional methods produce the flow field of $\sim$16$\times$ lower resolution. We believe that this study brings the field of experimental fluid dynamics one step closer to the long-term goal of having experimental measurement systems that can be used for real-time flow field estimation.
BibTeX
@inproceedings{elrefaie2024real,
  author = {Mohamed Elrefaie and Steffen Hüttig and Mariia Gladkova and Timo Gericke and Daniel Cremers and Christian Breitsamter},
  title = {{Real-time and On-site Aerodynamics using Stereoscopic PIV and Deep Optical Flow Learning}},
  booktitle = {Experiments},
  year = {2024},
  doi = {10.1007/s00348-024-03914-w},
}
Figure