YOLO-GRBI: An Enhanced Lightweight Detector for Non-Cooperative Spatial Target in Complex Orbital Environments.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Aerospace Engineering

Background:

• Autonomous on-orbit servicing and space situational awareness (SSA) require robust non-cooperative spatial target detection.
• Existing object detection methods face challenges with limited onboard computational resources and complex space imaging environments, leading to low accuracy for small, obscured targets.

Purpose of the Study:

• To develop an enhanced object detection network (YOLO-GRBI) that balances high accuracy with computational efficiency for spaceborne applications.
• To address the limitations of current frameworks in detecting small targets with low information entropy in noisy space imagery.

Main Methods:

• Implemented a reparameterized ELAN backbone for improved feature reuse and gradient flow.
• Introduced BiFormer and C2f-iAFF modules to enhance salient target attention and reduce detection errors.
• Integrated GSConv and VoV-GSCSP modules in the neck to optimize computational redundancy and preserve information entropy.
• Utilized focal loss for classification and confidence prediction to handle class imbalance.

Main Results:

• YOLO-GRBI demonstrated superior performance compared to the baseline YOLOv8n on a custom spacecraft dataset.
• Achieved a 4.9% increase in mAP@0.5 and a 6.0% increase in mAP@0.5:0.95.
• Successfully reduced model complexity and inference latency while maintaining high detection accuracy.

Conclusions:

• YOLO-GRBI offers a significant advancement in efficient and accurate non-cooperative spatial target detection for space applications.
• The proposed network effectively addresses the challenges of limited resources and complex imaging conditions in space.
• This work contributes to enhancing autonomous capabilities in space servicing and improving space situational awareness.