Autonomous Optical Satellite Location Using Low-Power Neural Architectures for Orbital Surveillance

Overview

The purpose of this project is to develop an energy-efficient autonomous system for detecting satellites and space debris in high-resolution imagery. It was hypothesized that a deep learning approach using a Faster R-CNN architecture would enable high detection accuracy in deep space backgrounds, addressing the energy consumption problems with current methods. A ResNet-18 Feature Pyramid Network was trained over 150 epochs using a curriculum learning dataset. The model was initially exposed to 128px crops and progressively moved to 640px resolution to improve recognition of small objects. Testing included multiple inference trials on unseen image sequences using a Flask-based server to evaluate processing performance. The model achieved stable convergence, with Mean Average Precision of 0.45. Final testing demonstrated that the system accurately localized satellites across five-frame sequences with processing times under 1000ms. A distinct trend of reduced loss was observed as the model specialized in 640px imagery during the final 60 epochs, ending at 0.045. The results indicate that curriculum learning effectively bridges the gap between speed and sensitivity for tiny space objects. These findings suggest that starting with simpler tasks allows the network to build foundational weights that stabilize complex training. Sources of error included star background noise, which was mitigated by implementing Non-Maximum Suppression to merge overlapping frames. In conclusion, the project successfully created a functional detection program for satellite debris monitoring. The hypothesis was supported as the final model achieved high confidence scores on full-scale images. This system offers a large improvement in energy over other methods.