TL;DR
This paper presents a novel geometric calibration method specifically designed for telescopes with large focal lengths and complex off-axis optics, addressing a gap in existing calibration techniques. The method was successfully applied to ESA's CaSSIS instrument on the Trace Gas Orbiter, significantly improving the quality of scientific products like color images.
What Problem Does This Paper Address?
Calibrating cameras is crucial for accurate image analysis, but standard methods often fall short when dealing with specialized instruments like telescopes. Telescopes, particularly those with large focal lengths and off-axis designs (meaning the optical axis of the lens doesn't align with the center of the detector), exhibit complex distortions that are not well-handled by off-the-shelf calibration tools. These tools typically rely on simple distortion models and require calibration targets that would need to be impractically large (kilometers in size) and placed very far away (tens of kilometers) for a telescope. This lack of specialized, effective calibration methods for such instruments hinders the production of high-quality scientific data from space missions.
Methods
The authors developed a geometric calibration method that combines an intrinsic camera model, a rational optical distortion model, and an extrinsic model.
- Intrinsic Model: This describes the basic transformation from 3D camera coordinates to 2D image coordinates. For CaSSIS, this model is simplified to only one degree of freedom (DOF) as the principal point is assumed to be at the image center.
- Rational Optical Distortion Model: This is the core of their novel approach. Unlike simpler radial or Brown-Conrady models, the rational model can accurately represent the complex, non-radial distortions found in off-axis optical systems. The paper highlights that this model can even represent its own inverse, allowing for simultaneous estimation of distortion and correction.
- Extrinsic Model: This accounts for the transformation from a reference frame (e.g., spacecraft) to the camera frame, involving rotation and translation.
The calibration process involves two main stages:
- On-ground Calibration: Initially, they attempted calibration using a dotted target and a collimator. However, due to limitations with the collimator, this method was not fully successful in capturing the precise distortion.
- In-flight Calibration: This proved to be the more effective approach. It utilizes images of star fields, leveraging the precisely known positions of stars from astronomical catalogs. The process involves:
- Image Assembly & Preprocessing: Combining raw data packets into full images, followed by de-noising and flattening to remove artifacts.
- Star Field Recognition: Using the Astrometry.net library to detect stars and match them to known catalog positions.
- False Detection Removal: Identifying and discarding stars that are not consistently detected across multiple images.
- Camera Rotation Initialization: Estimating initial camera rotations for each image using SPICE kernels.
- Iterative Bundle Adjustment (BA): A crucial step where the focal length and camera rotations are refined simultaneously for all images, iteratively rejecting outliers to improve accuracy.
- Rational Optical Distortion Estimation: With the intrinsic and extrinsic models "frozen," the rational distortion model is then estimated to minimize remaining projection errors.
Datasets
The paper utilizes the following datasets:
- Simulated CaSSIS optical distortion data: Used for initial model selection and comparison (described in Tab. 3).
- On-ground calibration data: Images of a dotted target captured through a collimator.
- In-flight calibration data: Star field images acquired by CaSSIS during the TGO mission. Specifically, three datasets were used:
- "mcc motor" (July 2016)
- "pointing cassis" (July 2016)
- "commissioning2" (April 2016)
The "mcc motor" and "pointing cassis" datasets were combined to form a "training set" for model estimation, while "commissioning2" served as a "validation set."
Key Insights and Ideas
- The inadequacy of standard calibration for complex optics: The paper strongly emphasizes that off-the-shelf calibration tools and models are insufficient for telescopes with large focal lengths and off-axis designs due to their complex distortion characteristics.
- The power of the rational distortion model: The rational model proves to be highly effective in accurately capturing the intricate optical distortions of the CaSSIS telescope, outperforming simpler radial and Brown-Conrady models.
- In-flight star field calibration as a robust solution: Utilizing star fields for calibration is a well-established technique, but the authors demonstrate its successful adaptation for a complex telescope by integrating it with advanced models and iterative refinement processes.
- The importance of iterative outlier rejection in Bundle Adjustment: The iterative outlier rejection scheme within the Bundle Adjustment process is critical for removing erroneous star detections, which would otherwise degrade the accuracy of the subsequent optical distortion estimation.
- The generalizability of the method: While developed for CaSSIS, the authors assert that their method is general and can be applied to calibrate other telescopes with similar optical complexities.
Key Results
- The developed geometric calibration method significantly improved the quality of CaSSIS scientific products.
- Distortion Model Performance:
- Simple radial and Brown-Conrady models resulted in errors greater than 1 pixel.
- Bi-cubic and rational models achieved errors less than 0.1 pixels, with the rational model being chosen for its suitability.
- In-flight Calibration Accuracy:
- The average Euclidean error before Bundle Adjustment was 3.65 pixels.
- After the first BA iteration, the error reduced to 2.78 pixels.
- After the fourth (final) BA iteration, the error was 2.56 pixels.
- The refined focal length was found to be 875.93 mm, slightly shorter than the nominal 880 mm.
- After fitting the rational distortion model, the average Euclidean distance between observed and predicted star positions was reduced to 0.66 pixels.
- On a separate validation set, the average projection error with the refined model was 0.47 pixels, compared to 3.56 pixels with the nominal model.
- Qualitative Improvement: Map-projected color images of Mars showed a dramatic reduction in color fringes and stitching artifacts when using the refined geometric parameters, resulting in a "near perfect" image.
Limitations and Future Work
- Collimator Limitations: The on-ground calibration using a collimator was hampered by an unknown residual distortion from the collimator itself, preventing precise estimation of the CaSSIS distortion using this method.
- Color Band Misalignment: During the color image experiment, map-projected images of individual color bands showed misalignments along the track (1-10 pixels), possibly due to slight deviations in the camera's off-nadir pointing relative to its rotation axis. This issue requires further investigation.
- Photometric Calibration Artifacts: Some prominent banding artifacts in the final color images are attributed to incorrect photometric calibration, not geometric calibration, and are being investigated independently.
- Future Work: The authors encourage the re-use of their calibration code and data, suggesting that their method can be applied to other telescopes. The paper also implicitly points to the need for further investigation into the observed color band misalignments and photometric calibration issues.
Summary updated