Commissioning an All-Sky Infrared Camera Array for Detection of Airborne Objects

To date, there is little publicly available scientific data on unidentified aerial phenomena (UAP) whose properties and kinematics purportedly reside outside the performance envelope of known phenomena. To address this deficiency, the Galileo Project is designing, building, and commissioning a multi-modal, multi-spectral ground-based observatory to continuously monitor the sky and collect data for UAP studies via a rigorous long-term aerial census of all aerial phenomena, including natural and human-made. One of the key instruments is an all-sky infrared camera array using eight uncooled long-wave-infrared FLIR Boson 640 cameras. In addition to performing intrinsic and thermal calibrations, we implement a novel extrinsic calibration method using airplane positions from Automatic Dependent Surveillance–Broadcast (ADS-B) data that we collect synchronously on site. Using a You Only Look Once (YOLO) machine learning model for object detection and the Simple Online and Realtime Tracking (SORT) algorithm for trajectory reconstruction, we establish a first baseline for the performance of the system over five months of field operation. Using an automatically generated real-world dataset derived from ADS-B data, a dataset of synthetic 3D trajectories, and a hand-labeled real-world dataset, 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%. The detection efficiency is heavily dependent on weather conditions, range, and aircraft size. Approximately 500,000 trajectories of various aerial objects are reconstructed from this five-month commissioning period. These trajectories are analyzed with a toy outlier search focused on the large sinuosity of apparent 2D reconstructed object trajectories. 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. We demonstrate the application of a likelihood-based statistical test to evaluate the significance of this toy outlier analysis. 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. This test is applicable to all of our future outlier searches.