As newer, larger, and faster cameras are attached to bigger telescopes, finding sources that move across images in a trajectory consistent with space rocks has become more complex. The computing power to sift through, detect, and produce a sensible number of alerts has grown accordingly. In order to reach any survey’s potential one must maximize the efficiency of the search while limiting the number of false positive detections.
In particular we focus on the deep stack problem of combining a long series of consecutive images fixed on the sky. The usual approach is to subtract a best static sky from each image and shift-and-stack them to account for the rate of motion of a putative object in hopes of increasing its signal-to-noise. This method has proven successful albeit computationally expensive. Furthermore, this method produces a fairly large number of false positives as the number of shifts grows.
We introduce a novel AI-based search model for finding moving targets in astronomical data. Using deep learning methods like R-CNN over a set of combined images we develop a method that extends the usual statistics to find moving sources. We show the power in this approach by sifting through the Transit Exoplanet Survey Satellite (TESS) and extending its detection limit to T~20. We show its potential as an Interstellar Object Survey. However, where this method excels is in its negligible number of false positives which make it a suitable for other large synoptic surveys like the LSST on the Vera C. Rubin Observatory.