Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SatSplatYOLO: 3D Gaussian Splatting-based Virtual Object Detection Ensembles for Satellite Feature Recognition

Van Minh Nguyen, Emma Sandidge, Trupti Mahendrakar, Ryan T. White

TL;DR

SatSplatYOLO tackles autonomous detection of satellite components around unknown non-cooperative RSOs by learning a 3D representation of the target with 3D Gaussian Splatting, generating synthetic views from novel camera poses, and ensembling YOLOv5 in the rendered views. The approach achieves a hardware-efficient end-to-end pipeline suitable for on-board deployment, delivering reliable detections and enabling subsequent pose estimation. Key contributions include a systematic camera-generation method for 3DGS rendering, a rendering-based ensembling technique to improve satellite component detection, and hardware-in-the-loop experiments with realistic lighting and motion. The results show strong 3D reconstruction quality and improved unknown-satellite detection, indicating practical potential for autonomous guidance, navigation, and control in on-orbit servicing and ADR scenarios.

Abstract

On-orbit servicing (OOS), inspection of spacecraft, and active debris removal (ADR). Such missions require precise rendezvous and proximity operations in the vicinity of non-cooperative, possibly unknown, resident space objects. Safety concerns with manned missions and lag times with ground-based control necessitate complete autonomy. In this article, we present an approach for mapping geometries and high-confidence detection of components of unknown, non-cooperative satellites on orbit. We implement accelerated 3D Gaussian splatting to learn a 3D representation of the satellite, render virtual views of the target, and ensemble the YOLOv5 object detector over the virtual views, resulting in reliable, accurate, and precise satellite component detections. The full pipeline capable of running on-board and stand to enable downstream machine intelligence tasks necessary for autonomous guidance, navigation, and control tasks.

SatSplatYOLO: 3D Gaussian Splatting-based Virtual Object Detection Ensembles for Satellite Feature Recognition

TL;DR

SatSplatYOLO tackles autonomous detection of satellite components around unknown non-cooperative RSOs by learning a 3D representation of the target with 3D Gaussian Splatting, generating synthetic views from novel camera poses, and ensembling YOLOv5 in the rendered views. The approach achieves a hardware-efficient end-to-end pipeline suitable for on-board deployment, delivering reliable detections and enabling subsequent pose estimation. Key contributions include a systematic camera-generation method for 3DGS rendering, a rendering-based ensembling technique to improve satellite component detection, and hardware-in-the-loop experiments with realistic lighting and motion. The results show strong 3D reconstruction quality and improved unknown-satellite detection, indicating practical potential for autonomous guidance, navigation, and control in on-orbit servicing and ADR scenarios.

Abstract

On-orbit servicing (OOS), inspection of spacecraft, and active debris removal (ADR). Such missions require precise rendezvous and proximity operations in the vicinity of non-cooperative, possibly unknown, resident space objects. Safety concerns with manned missions and lag times with ground-based control necessitate complete autonomy. In this article, we present an approach for mapping geometries and high-confidence detection of components of unknown, non-cooperative satellites on orbit. We implement accelerated 3D Gaussian splatting to learn a 3D representation of the satellite, render virtual views of the target, and ensemble the YOLOv5 object detector over the virtual views, resulting in reliable, accurate, and precise satellite component detections. The full pipeline capable of running on-board and stand to enable downstream machine intelligence tasks necessary for autonomous guidance, navigation, and control tasks.
Paper Structure (13 sections, 8 equations, 7 figures, 1 table)

This paper contains 13 sections, 8 equations, 7 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Object Detection
  4. 3D Scene Rendering
  5. Optical Navigation for RPO
  6. Methods
  7. 3D Gaussian Splatting
  8. Camera Generation
  9. YOLOv5 Ensembling
  10. Datasets
  11. Performance Evaluation
  12. Results
  13. Conclusion

Figures (7)

  • Figure 1: Nearest point to non-intersecting camera lines
  • Figure 2: Visualization for generation of 64 new cameras with radius 1 from the original camera position. Note that the sphere is deformed due to scaling difference between x, y, z axes
  • Figure 3: Object Detection Dataset Mahendrakar2024
  • Figure 4: ORION Testbed wilde_orion_2016
  • Figure 5: High quality render of true camera and 2 of the generated camera array from the same true camera
  • ...and 2 more figures