YOLO‑11 Enables 94% Detection of Near‑Earth Object and Satellite Streaks

The Spanish Royal Navy Academy’s observatory built the StreakMind system to automatically detect linear streaks left by satellites or asteroids in astronomical images. Using both real observations from the La Sagra telescope (Celestron C14 + Fastar f/2.1, SBIG ST‑10X CCD) and 280 synthetic streaks, the team assembled a dataset of over 2,000 images with 765 manually labeled streaks ranging from 8.5 to 1,161 pixels.

StreakMind’s core detection module employs the YOLO‑11‑OBB model, a single‑stage network that outputs oriented bounding boxes (OBB) with angles, suitable for the tilted, elongated streaks in sky images. The workflow converts FITS files to normalized PNGs, runs YOLO‑11‑OBB for initial detection, and cross‑references Gaia star catalogs to discard false positives near bright stars.

After initial detection, the system refines each streak geometrically: it analyzes the photometric profile along the OBB’s major axis, extends the box to the true start and end points, and uses corner clustering to locate stable endpoints and the center. Consecutive frames are then linked by comparing pixel velocity and direction, forming continuous trajectories.

Detected streaks are converted to the Minor Planet Center (MPC) standard format, cross‑matched with satellite ephemerides, and assigned confidence scores via a two‑component Gaussian model. All records, including trajectory IDs, observatory codes, and MPC identifiers, are stored in a SQLite database for downstream analysis.

Evaluation on an independent test set of 273 images (input resolution 640 px, confidence threshold 0.25, IoU 0.45) shows the model reaches 94% precision and 97% recall, correctly identifying 107 of 110 real streaks. Star‑related false positives drop by 77% after catalog filtering, and the geometric refinement corrects short bounding‑box issues. Compared with manual inspection, StreakMind offers markedly higher efficiency, repeatability, and sensitivity.

Overall, StreakMind demonstrates that large‑scale survey images can be processed automatically to extract reliable streak detections, providing a scalable solution for near‑Earth object monitoring and space‑environment surveillance.