Title: Enhancing Space Object Detection in LEO Constellations Through Multi-Satellite Collaborative Vision

As the density of satellites within Low Earth Orbit (LEO) constellations continues to rise, the near-Earth space environment is facing severe congestion, rendering Space Object Detection (SOD) a critical imperative for maintaining space safety and sustainability. To prevent collisions and guarantee the uninterrupted nature of space operations, SOD systems are required to perform rapid and precise detections while operating under tight onboard computational limitations. This study explores the efficacy of integrating multi-viewpoint observations within a deep learning (DL) architecture to boost SOD capabilities.

We developed a practical multi-view processing pipeline alongside various input representations designed to feed multi-view data into YOLO-based detection models. Our experimental results demonstrate that multi-view input integration is generally feasible and consistently yields superior performance metrics, specifically regarding mAP50 and mAP50-95. For instance, when utilizing the YOLOv9-m model, transitioning from a single-view approach to a three-view fused RGB configuration increased mAP50 from 0.638 to 0.732, while mAP50-95 rose from 0.227 to 0.276. Furthermore, compared to single-view baselines, the optimal three-view grayscale setup achieved a 36.3% improvement in mAP50 and a 46.5% gain in mAP50-95. These outcomes confirm that multi-view fusion is a robust and effective methodology for SOD, offering significant benefits for enhancing space situational awareness in LEO constellation operations.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC