Detecting near-Earth objects using real-time synthetic tracking surveys

Introduction

The detection of near-Earth objects (NEOs) is a critical task for planetary defense and of significant importance for astronomy and astronautics. While existing surveys have successfully identified practically all large (>1km) NEOs and have successfully progressed for those >100 m, smaller objects are still challenging to detect, being overwhelmingly detected only during close flybys.

Traditional detection methods, such as the "blink" technique, rely on large aperture telescopes to detect smaller, fainter, objects. However, doing so is increasingly expensive, with the costs becoming prohibitive. In comparison, Synthetic Tracking (ST) is a detection technique [1][2] that allows the use of (many) smaller telescopes with longer integration times, where the detections can be significantly under the noise floor of individual images, by combining the images to increase the signal to noise ratio across all possible trajectories of a faint potential moving object. Despite the advantages of Synthetic Tracking, due to the perceived need for vastly more computational power, the backbone of current surveys is still the blink method. This factor is particularly prominent for NEOs, whose apparent velocities are much higher than for other types of minor planets, leading to a large number of potential object trajectories.

Methods

We present Synthetic Tracking on Umbrella (STU), a moving object detection software which leverages the power of modern GPUs to detect NEOs in real time. STU is developed on top of the Umbrella2 library[3], which we have previously developed to implement a blink detection pipeline, including both the detection algorithm and auxiliary functionality necessary for a complete detection processing pipeline. Thanks to innovative search strategies, efficient use of hardware and a multi-step candidate rejection process, STU can perform real time detection of fast-moving Near Earth Asteroids, even on large, multi-CCD instruments. STU can do so using modest hardware resources, such as a single off-the-shelf GPU, while being robust to outlier noise.

 

Results

Previously, we successfully demonstrated the functionality of STU through an observation archive spanning more than 100000 images from various telescopes (Stanescu et al., 2025, accepted), using images obtained in various conditions (variable seeing, detectors with different noise levels, pixel scales, field of view, objects with different apparent magnitudes, and moving rates). Furthermore, we have further tested STU in realistic survey conditions using the 2.54m Isaac Newton Telescope (INT), the 1.6m Korean Astronomy and Space Science Institute (KASI) telescopes[4].

For the INT, real-time capability was demonstrated with the Wide Field Camera (WFC) on the 2.54m Isaac Newton Telescope (INT), which has 4 x 9 MPix CCDs, at a resolution of 0.33"/pix. We continuously acquired data in 12-minute fields of 12 exposures, in several multi-night observing runs. For each field, image processing (sensor correction and plate solving) took 8 minutes and the STU detection pipeline took <2 minutes. Moreover, the trajectory scan took only 8 seconds. This runtime allowed real-time processing on a PC with a GPU under 1000 euro. For the much larger KASI telescope we have used the Korean Microlensing Telescope Network Camera (KMTCam), of 4 x 85 Mpix, at a resolution of 0.4"/pix. The STU runtime for the KASI telescope has been 7 minutes at a maximum apparent motion of 10"/min. This runtime is the result of STU having a pixel processing rate (PPR), which is the number of image pixels co-added per unit time, of 1700 GPix/sec. In the newer versions of STU (v0.7), we have added support for repurposing "small integer dot product" instructions, originally introduced in newer GPU hardware for accelerating machine learning workloads, for co-adding pixels faster. This improved the PPR to 2400 GPix/sec.

                                                                               

As an example of the results in these observing runs, we present two examples of asteroids detected by our software. First, in left figure, the detection of NEAs 2023 DZ₂, formerly catalogued as Virtual Impactor, discovered and recovered by our group between February 27 and March 1 2023 (MPEC 2023-F12, https://minorplanetcenter.net/mpec/K23/K23F12.html), Second, in right figure, the detection of 2024 CW₂, at an apparent motion of 9.54"/min. This object was discovered by our team in the night of 11/12 February 2024, and paired soon after being published to 2007 EG.

Acknowledgements

This work had been supported by a grant of the Romanian National Authority for Scientific Research -- UEFISCDI, project number PN-III-P2-2.1-PED-2021-3625.

 

References
[1] B. Gladman, et al. (1997) Astronomy and Astrophysics 317:L35.

[2] C. Zhai, et al.(2018) Technical note: Asteroid detection demonstration from skysat-3 b612 data using synthetic tracking.

[3] M. Stănescu, et al. (2021) Astronomy and Computing 35:100453.

[4] O. Vaduvescu, et al. (2025) New Astronomy 119:102410 ISSN 1384-1076.