Commissioning An All-Sky Infrared Camera Array for Detection Of Airborne Objects

Brief

The Galileo Project deployed an array of eight uncooled FLIR LWIR Boson 640 cameras at its Massachusetts development site and operated it from January through May 2024, reconstructing approximately 500,000 aerial trajectories using a YOLO detection model and SORT tracking algorithm. For ADS-B-equipped aircraft the system achieved a 41% acceptance rate and a 36% mean frame-by-frame detection efficiency, with performance heavily dependent on weather, range, and aircraft size. A sinuosity-based outlier search flagged ~16% of trajectories (~80,000); manual inspection reduced that set to 144 ambiguous cases that could not be resolved without additional sensor modalities. Applying a likelihood-based statistical test produced a 95%-confidence upper limit of 18,271 UAP-candidate outlier events over the five-month period.

Metadata

Category

Search

Venue

Sensors

Type

Peer-reviewed

Year

2024

Authors

Laura Domine, Ankit Biswas, Richard Cloete, Alex Delacroix, Andriy Fedorenko, Lucas Jacaruso, Ezra Kelderman, Eric Keto, Sarah Little, Abraham Loeb, Eric Masson, Mike Prior, Forrest Schultz, Matthew Szenher, Wesley Watters, Abigail White

DOI

10.3390/s25030783

arXiv

2411.07956

Access

Open access

Length

20.1 M

Programs

Galileo Project

Instruments

FLIR LWIR Boson 640 (8-camera all-sky array), Dalek IR camera array, Nvidia Jetson Orin NX, ZWO ASI462MC near-infrared camera

Data sources

ADS-B (Automatic Dependent Surveillance–Broadcast), OpenSky Network historical database, YOLO object detection model, SORT tracking algorithm, Segment Anything Model (SAM)

Tags

UAP-detection, infrared-sensing, aerial-census, instrument-commissioning, machine-learning, technosignature

Key points

• Eight uncooled FLIR LWIR Boson 640 cameras (spectral band 7.5–13.5 μm) arranged in a 360° azimuth / +80° elevation hemispherical array, nicknamed 'Dalek', deployed at a Massachusetts site within 5 miles of a regional airport.p.4
• Approximately 500,000 aerial trajectories reconstructed from five months of commissioning data (January–May 2024).p.1
• 41% acceptance rate for ADS-B-equipped aircraft; 36% mean frame-by-frame detection efficiency; both metrics degrade significantly with adverse weather, longer range, and smaller aircraft size.p.1
• ~16% of trajectories (~80,000) flagged as outliers via sinuosity-based filter; manual IR-image review reduced the ambiguous set to 144 trajectories that cannot be further elucidated without range, kinematic, or multi-modal sensor data.p.1
• Likelihood-based statistical test places a 95%-confidence upper limit of 18,271 UAP-candidate outlier events over the five-month interval; the test is designed to generalize to all future outlier searches.p.1
• Novel extrinsic calibration method uses ADS-B airplane positions as reference targets, a workaround necessitated by LWIR cameras' inability to image stars, which is the standard reference grid for optical astronomical cameras.p.9
• Detection pipeline combines YOLO (You Only Look Once) for frame-by-frame object detection with SORT (Simple Online and Realtime Tracking) for trajectory reconstruction.p.1
• Observatory is designed as multi-modal and multi-spectral (IR, visible, UV, acoustic, radio spectrum, magnetic field, charged particle, and weather sensors); the LWIR array is the first instrument to complete commissioning.p.2

Verbatim

• “we find an acceptance rate (fraction of in-range airplanes passing through the effective field of view of at least one camera that are recorded) of 41% for ADS-B-equipped aircraft, and a mean frame-by-frame aircraft detection efficiency (fraction of recorded airplanes in individual frames which are successfully detected) of 36%.”

  p.1

• “About 16% of the trajectories are flagged as outliers and manually examined in the IR images. From these ∼ 80,000 outliers and 144 trajectories remain ambiguous, which are likely mundane objects but cannot be further elucidated at this stage of development without information about distance and kinematics or other sensor modalities.”

  p.1

• “Our observed count of ambiguous outliers combined with systematic uncertainties yields an upper limit of 18,271 outliers for the five-month interval at a 95% confidence level.”

  p.1

Most interesting

• The eight-camera LWIR array is nicknamed 'Dalek'; the team evaluated configurations from four to fourteen cameras before settling on eight as the optimal balance of angular resolution and cost.
• LWIR cameras cannot image stars, so the team developed a novel extrinsic calibration method using commercial ADS-B aircraft transponder data as a substitute for the standard celestial reference grid used in optical astronomy.
• Intrinsic calibration required a metal chessboard cooled to −20°C placed against a plate heated to 50°C to generate sufficient thermal contrast for the LWIR sensors, no optical chessboard can be used.
• The fiberglass enclosure dome is rated via finite element analysis to withstand 200 mph wind gusts, roughly double the local wind record, with less than 1 mm of radial deformation.
• The first-generation germanium windows used a 95% anti-reflection coating that degraded significantly within one year of outdoor exposure; replacements with a hydrophobic HP/AR coating (~96% transmissivity) showed substantially less deterioration.
• The 144 ambiguous trajectories are described as 'likely mundane objects', the system cannot classify them further without range measurements or data from additional sensor modalities, illustrating the hard limit of single-modality IR-only analysis.