On Scene Flow Computation of Gas Structures with Optical Gas Imaging Cameras

Johannes Rangel, Robert Schmoll, Andreas KrollWACV 2020

TL;DR

This paper introduces a novel framework for computing the 3D velocity of gas structures using stereo optical gas imaging (OGI) cameras. By combining optical flow and stereo disparity estimation, the method provides a more robust way to track gas emissions, crucial for industrial safety and environmental monitoring.

What Problem Does This Paper Address?

Detecting and quantifying gas leaks, particularly greenhouse gases like methane, is critical for industrial safety, economic reasons, and mitigating climate change. While optical gas imaging (OGI) cameras are effective at visualizing these leaks, accurately measuring the velocity of the escaping gas has been challenging. Existing methods often rely on single cameras and make simplifying assumptions about gas flow direction, leading to inaccuracies in real-world scenarios with unpredictable wind and temperature fluctuations. This paper aims to overcome these limitations by developing a more robust 3D gas velocity estimation technique.

Methods

The core of the proposed method is a variational framework that estimates "scene flow" – the 3D motion of points in a scene – from stereo OGI camera pairs. The process involves several key steps:

  • Pre-Processing: This includes correcting for geometric and radiometric differences between the two cameras, as well as mitigating temporal shifts (jitter) caused by the camera's cooling system. Differential images are created to enhance gas contrast.
  • Optical Flow Computation: A classical variational method (based on Brox et al.) is used to estimate the 2D motion of gas structures within each camera's image. This step focuses on brightness constancy and spatial gradient similarities.
  • Disparity and Disparity Change Computation: A joint optimization approach is employed to compute the stereo disparity (depth information) and its change over time. This is crucial for inferring 3D motion.
  • Image Segmentation: A variational segmentation approach is used to identify regions likely containing gas information, separating them from the background. This helps to focus subsequent analysis on relevant areas.
  • 3D Gas Velocity Computation: The computed optical flow, disparity, and disparity change are combined using stereokinematic constraints to derive the 3D velocity field of the gas structures. The resulting velocity field represents a path-averaged and concentration-weighted motion.

Datasets

  • Middlebury Dataset: Used for benchmarking and tuning the optical flow parameters.
  • Synthetic Data: A simulation environment was developed to generate synthetic stereo image sequences of incompressible buoyant fluid flows. This allowed for evaluation under controlled conditions without real-world noise or complexities. Three different synthetic sequences were generated, varying in flow magnitude and direction.
  • Real Field Data: Measurements were taken in a biogas facility, where methane was released. This data was used to evaluate the system's performance under actual operating conditions.

Key Insights and Ideas

  • Stereo Vision for Gas Flow: Utilizing stereo OGI cameras significantly improves the robustness of gas velocity estimation compared to single-camera approaches, as it allows for direct 3D reconstruction.
  • Variational Methods for Gas: Variational methods, commonly used for visible objects, are adapted and shown to be effective for estimating scene flow in gas structures, even with their inherent semi-transparent and low-contrast nature.
  • Joint Disparity and Disparity Change: Computing disparity and its change jointly leads to smoother and more reliable 3D velocity fields, reducing the variance compared to decoupled methods.
  • Segmentation is Crucial but Challenging: While image segmentation helps to isolate gas regions, its performance is highly dependent on gas contrast and can be affected by moving objects in the background, leading to potential errors.

Key Results

  • Synthetic Data: The proposed method showed comparable or better performance than alternative decoupled scene flow methods (SGM+Brox, SGM+Fbck) in terms of relative mean velocity error (RMVE) and 3D average endpoint error (3D AEE) on synthetic data. The computed 3D velocity fields generally agreed well with the dominant gas velocity and direction of the reference.
  • Field Measurements: In real-world biogas facility experiments, the system was able to capture gas velocity magnitudes and directions that approached those measured by a 3D ultrasonic anemometer. The joint computation of disparity and disparity changes resulted in smoother results (higher CUSF values) compared to a decoupled method. However, the magnitude of the computed gas velocity was generally underestimated compared to wind speed.

Limitations and Future Work

  • Photometric and Geometric Assumptions: The accuracy of the method is still affected by the inherent assumptions of photometric consistency and geometric calibration, especially in low-contrast gas regions or when dealing with untextured areas.
  • Segmentation Sensitivity: The image segmentation approach, while useful, is sensitive to image contrast and can misclassify moving objects as gas or background.
  • Underestimation of Velocity: The computed gas velocity magnitude was observed to be underestimated in field experiments.
  • Future Work: The authors suggest exploring more sophisticated gas descriptors beyond motion parameters for image segmentation to improve accuracy. Further investigation into the causes of velocity underestimation and potential improvements to the photometric consistency assumptions for semi-transparent gas structures would be beneficial. The portability and applicability of the system in real industrial environments were demonstrated, suggesting potential for practical deployment.

Summary updated

BibTeX
@inproceedings{rangel2020scene,
  author = {Johannes Rangel and Robert Schmoll and Andreas Kroll},
  title = {{On Scene Flow Computation of Gas Structures with Optical Gas Imaging Cameras}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2020},
}
Figure