Machine learning UAP detection refers to the use of automated ML pipelines to classify aerial objects from sensor data and identify anomalous trajectories that cannot be explained by known aircraft.
Galileo Project Pipeline
The galileo-project’s commissioning dataset (Jan–May 2024, laura-domine et al.) describes a two-stage ML pipeline:
- YOLO (You Only Look Once): Per-frame object detection; achieves ~36% mean frame-by-frame detection efficiency for ADS-B-equipped aircraft.
- SORT (Simple Online and Realtime Tracking): Trajectory reconstruction across frames; assembles detections into ~500,000 aerial-object trajectories.
Calibration: ADS-B transponder data serves as the ground-truth baseline for tuning detection parameters.
Outlier search: An algorithm flags ~16% of trajectories (~80,000) for manual review. After review, 144 remained truly ambiguous, primarily due to the absence of range (distance) data — the main limitation of single-station deployments.
Key Limitation
Single-station infrared observatories cannot measure the range to a detected object, making it impossible to derive velocity and size — the key parameters for discriminating UAP from known objects. The galileo-project plans to address this with stereoscopic multi-station Dalek deployments separated by a few miles.