An Efficient and Robust Star Identification Algorithm Based on Neural Networks.

Area of Science:

• Astronomy
• Computer Science
• Artificial Intelligence

Background:

• Lost-in-space star identification is crucial for autonomous spacecraft navigation.
• Existing algorithms struggle with noise, dynamic conditions, and computational efficiency.

Purpose of the Study:

• To develop a robust and efficient lost-in-space star identification algorithm.
• To leverage deep learning for improved accuracy and speed in star pattern recognition.

Main Methods:

• A modified log-Polar mapping was employed to create rotation-invariant star patterns.
• A one-dimensional Convolutional Neural Network (1D CNN) was utilized for star pattern classification.
• Global average pooling replaced fully-connected layers to reduce model complexity and overfitting.

Main Results:

• The 1D CNN algorithm achieved high accuracy under various noise conditions (e.g., 98.1% with position noise, 97.4% with false stars).
• Demonstrated reliable performance in dynamic scenarios with 82.1% accuracy at 10 degrees/second angular velocity.
• Achieved a fast identification time of 32.7 ms and low memory footprint (1920 KB).

Conclusions:

• The proposed 1D CNN algorithm offers superior robustness and speed compared to traditional methods.
• The algorithm is well-suited for real-time, autonomous navigation in embedded systems.
• This deep learning approach significantly advances lost-in-space star identification capabilities.