A Novel Systematic Error Compensation Algorithm Based on Least Squares Support Vector Regression for Star Sensor Image Centroid Estimation

Jun Yang, Bin Liang, Tao Zhang, Jingyan SongSensors 2011 · DOI

TL;DR

This paper tackles the critical problem of systematic errors in star centroid estimation, a key factor for accurate satellite attitude determination. They propose a novel compensation algorithm using Least Squares Support Vector Regression (LSSVR) that significantly improves centroid estimation accuracy, reducing errors from 0.06 pixels to an impressive 6 x 10⁻⁵ pixels.

What Problem Does This Paper Address?

Star trackers are essential for determining a satellite's orientation in space, and their accuracy heavily relies on precisely locating stars in images. The process of estimating a star's center (centroid estimation) is prone to systematic errors, which can lead to significant inaccuracies in attitude determination (arc-seconds level). These errors arise from the inherent limitations of digital image processing, specifically the discretization of continuous signals and the finite size of the sampling window used to capture star light. Existing methods struggle to fully address these systematic errors, especially when dealing with varying star intensity profiles. This paper aims to thoroughly analyze these systematic errors and introduce a robust method to compensate for them.

Methods

The authors first conduct a theoretical analysis of systematic errors using frequency domain methods and numerical simulations. They break down the systematic error into two components:

  • Approximation Error: This arises from discretizing the continuous star signal and the sampling frequency limitations of the detector.
  • Truncation Error: This occurs when the sampling window used for centroid calculation is smaller than the actual spread of the star's light (Gaussian width).

Based on this analysis, they propose a novel Least Squares Support Vector Regression (LSSVR) algorithm with a Radial Basis Function (RBF) kernel for systematic error compensation. LSSVR is chosen for its efficiency and ability to find global solutions. The LSSVR model learns the relationship between the actual measured star centroid position and the systematic error. Once trained, it predicts the systematic error for a given star, allowing for its subtraction to obtain a more accurate ideal star centroid position.

Datasets

The paper does not introduce new datasets. Instead, it relies on:

  • Simulated Star Images: Numerical simulations are used extensively to generate training data and test the proposed algorithm. These simulations vary parameters like Gaussian width and star position to cover different scenarios.
  • Actual Star Images: A set of real night sky images captured by a CANON 20D camera are used for experimental validation. These images are processed to extract star centroid positions before and after compensation.

Key Insights and Ideas

  • Systematic Error Decomposition: Clearly separating systematic error into approximation and truncation components provides a structured way to understand and address the problem.
  • Window Size Criterion: The paper offers a practical criterion for selecting the sampling window size to minimize truncation error, suggesting it should be slightly larger than the star's Gaussian width.
  • LSSVR's Efficacy: LSSVR, particularly with an RBF kernel, proves to be a powerful tool for modeling and compensating complex, non-linear systematic errors in centroid estimation.
  • Sinusoidal vs. Linear Error Behavior: The systematic error exhibits a sinusoidal relationship with the ideal star position when the Gaussian width is small, transitioning to a linear relationship for larger widths. This observation is crucial for understanding the error dynamics.

Key Results

  • Accuracy Improvement: The LSSVR compensation algorithm significantly improves star centroid estimation accuracy. The residual systematic error is reduced from an initial value of up to 0.06 pixels to less than 6 x 10⁻⁵ pixels.
  • Performance Comparison: The proposed LSSVR method demonstrates accuracy that is one to two orders of magnitude higher than previously proposed methods like neural networks or analytical compensation algorithms.
  • Robustness: The algorithm shows good performance across different Gaussian widths, indicating its adaptability to varying star intensity profiles.
  • Real-world Validation: Experiments with actual star images confirm the effectiveness of the LSSVR compensation algorithm in reducing systematic errors.

Limitations and Future Work

  • Random Noise: The paper acknowledges that the influence of random noise on the LSSVR compensation algorithm was not considered. The authors suggest that further study is needed to understand and improve performance under large random noise conditions.
  • Computational Cost: While LSSVR is noted for its lower memory requirements compared to traditional SVM, the training phase for LSSVR can still be computationally intensive, especially with a large number of training samples.
  • Generalization to Other Errors: The current work focuses on systematic errors inherent to the COM algorithm. Future work could explore compensation for other types of errors or their combined effects.

Summary updated

Abstract

The star centroid estimation is the most important operation, which directly affects the precision of attitude determination for star sensors. This paper presents a theoretical study of the systematic error introduced by the star centroid estimation algorithm. The systematic error is analyzed through a frequency domain approach and numerical simulations. It is shown that the systematic error consists of the approximation error and truncation error which resulted from the discretization approximation and sampling window limitations, respectively. A criterion for choosing the size of the sampling window to reduce the truncation error is given in this paper. The systematic error can be evaluated as a function of the actual star centroid positions under different Gaussian widths of star intensity distribution. In order to eliminate the systematic error, a novel compensation algorithm based on the least squares support vector regression (LSSVR) with Radial Basis Function (RBF) kernel is proposed. Simulation results show that when the compensation algorithm is applied to the 5-pixel star sampling window, the accuracy of star centroid estimation is improved from 0.06 to 6 × 10(-5) pixels.
BibTeX
@inproceedings{yang2011novel,
  author = {Jun Yang and Bin Liang and Tao Zhang and Jingyan Song},
  title = {{A Novel Systematic Error Compensation Algorithm Based on Least Squares Support Vector Regression for Star Sensor Image Centroid Estimation}},
  booktitle = {Sensors},
  year = {2011},
  doi = {10.3390/s110807341},
}
Figure