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 fast 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 with 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 three-component two-dimensional (3C2D) velocity estimation. 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 fast flow field estimation.

The Challenge of Industrial PIV

Particle Image Velocimetry (PIV) has long been the gold standard for non-intrusive fluid flow measurement. However, traditional PIV faces a significant "speed limit" due to its reliance on cross-correlation algorithms. These methods are computationally expensive and typically produce spatially sparse vector fields, as motion is averaged over discrete interrogation windows. The challenge intensifies with stereoscopic PIV, which is required to capture three-component (3C) velocity fields from two-dimensional planes.

In industrial aerodynamics, such as analyzing the wake of a vehicle in motion, researchers often employ large-scale setups like the "Ring of Fire." These experiments generate massive volumes of image data. Processing this data with classical software often requires days of manual effort and significant expertise. While deep learning has emerged as a potential solution for 2D PIV, existing models frequently struggle with the noise, varying particle densities, and out-of-plane motions inherent in real-world industrial environments.

RAFT-StereoPIV: Architecture and Methodology

To address these challenges, the authors introduce RAFT-StereoPIV, an architecture based on Recurrent All-pairs Field Transforms (RAFT). This model is specifically designed to handle the complexities of stereoscopic particle images and provide high-resolution, dense flow fields.

Overview of the RAFT-StereoPIV architecture and processing pipeline, including feature encoding, correlation volumes, and iterative updates.

The RAFT-StereoPIV pipeline operates in four distinct stages:

  1. Feature and Context Encoding: Twin Convolutional Neural Networks (CNNs) extract per-pixel feature maps from the input image pairs.
  2. All-Pairs Correlation: The model computes the visual similarity between every pixel in the first frame and every pixel in the second, constructing a 4D correlation volume that captures potential displacements.
  3. Iterative Updates: A Convolutional Gated Recurrent Unit (Conv-GRU) iteratively refines the estimated flow field by looking up values in the correlation volume.
  4. 3C2D Reconstruction: Finally, a geometric mapping function integrates the 2D optical flow from two camera perspectives to reconstruct the full three-component (3C) velocity field.

For applications requiring higher throughput, the team also developed RAFT-StereoPIV-Small, which reduces the parameter count by 59% (from 5.3M to 2.2M) while maintaining competitive performance.

Experimental Setup and Datasets

The model's robustness was tested using a combination of synthetic benchmarks and real-world industrial data. A key application was the "Ring of Fire" (RoF) setup, an on-site measurement facility where a full-scale vehicle (e.g., a Volkswagen up!) drives through a tunnel seeded with helium-filled soap bubbles (HFSB).

The 'Ring of Fire' experimental setup for on-site 3D car aerodynamics measurement.

In this setup, high-speed cameras capture the vehicle's wake as it passes through a laser light sheet. RAFT-StereoPIV was able to process these complex, noisy images to reveal fine turbulent structures that are often lost or blurred in traditional sparse vector fields. Additional validation was performed using wind tunnel data from a 25%-scaled Ahmed body, ensuring the model's accuracy against controlled experimental benchmarks.

Key Results and Performance

RAFT-StereoPIV demonstrates a significant leap in performance over previous state-of-the-art (SOTA) deep learning models for PIV. On the "Problem Class 2" validation set—designed to mimic imperfect experimental conditions—the model achieved a 68% error reduction. Even on completely unseen test data (Problem Class 1), it maintained a 47% error reduction, proving its exceptional generalizability.

Beyond accuracy, the primary advantage is resolution. By providing per-pixel flow estimates instead of window-averaged vectors, RAFT-StereoPIV offers a spatial resolution approximately 16 times higher than traditional cross-correlation methods. This allows for the visualization of micro-scale flow features previously inaccessible in large-scale industrial tests. Furthermore, while training is resource-intensive, inference is fast; processing a 1024 × 1024 px² image pair takes approximately 40 seconds on a modern GPU.

Limitations and Future Outlook

The authors acknowledge that high-resolution comes with a trade-off: some of the smallest-scale structures observed might be artifacts of the optical flow processing rather than true physical turbulence, especially in regions where the laser sheet is thick.

Future research will focus on integrating 3D particle tracking and data assimilation to further enhance spatial resolution. Additionally, the team plans to investigate the impact of physical parameters like camera orientation and light-sheet spacing on AI accuracy, and validate the model against a dedicated synthetic Stereo PIV dataset, as current benchmarks are largely 2D.

Summary updated

Abstract

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 fast 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 with 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 three-component two-dimensional (3C2D) velocity estimation. 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 fast flow field estimation.
BibTeX
@inproceedings{elrefaie2024site,
  author = {Mohamed Elrefaie and Steffen Hüttig and Mariia Gladkova and Timo Gericke and Daniel Cremers and Christian Breitsamter},
  title = {{On-site aerodynamics using stereoscopic PIV and deep optical flow learning}},
  booktitle = {Experiments},
  year = {2024},
  doi = {10.1007/s00348-024-03914-w},
}
Figure