One of the most challenging tasks in surveying space objects with optical telescopes is the need of real time image processing to quickly perform follow-up observations and acquire additional images in the visibility window. This is instrumental to performing sufficiently accurate initial orbit determination that allows scheduling further measurements with other sensors. Nowadays, the state of the art for tracklet extraction is mainly associated with segmentation techniques characterized by a processing time not suitable for quick sensor tasking. In this paper, a novel approach based on U-Net, a deep neural network, is proposed to process images of optical telescopes in real time, with the aim of extracting the tracklet information. As in all the other machine learning applications, a series of steps is required to obtain a working system: dataset creation, pre-processing, training, testing and post-processing. The dataset creation and pre-processing steps are needed to properly prepare the training step, while the post-processing one is applied to refine the output of the trained network. In order to succeed in the learning process, a sufficiently vast dataset composed by a set of images coupled with a label, in this case black&white masks (with a black background and a white tracklet), is needed. In order to be general, 360 realistic synthetic images and labels have been generated. Subsequently, images are rescaled (to speed up training) and normalized. The trained network is tested against a set of scaled synthetic and real images. Images are altered to reduce vignetting and to level out the background brightness, downconverted to 8 bit and normalized. The post-processing step performs a centroid identification and estimates the object right ascension and declination. The average processing time per real image is less than 4 s and bright tracklets are easily recognized with a mean centroid angular error of 0.25° in 75 % of test cases using a 2° field of view telescope

Real Time Space Object Tracklet Extraction from Telescope Survey Images with Machine Learning

De Vittori, A.;Cipollone, R.;Di Lizia, P.;Massari, M.
2020-01-01

Abstract

One of the most challenging tasks in surveying space objects with optical telescopes is the need of real time image processing to quickly perform follow-up observations and acquire additional images in the visibility window. This is instrumental to performing sufficiently accurate initial orbit determination that allows scheduling further measurements with other sensors. Nowadays, the state of the art for tracklet extraction is mainly associated with segmentation techniques characterized by a processing time not suitable for quick sensor tasking. In this paper, a novel approach based on U-Net, a deep neural network, is proposed to process images of optical telescopes in real time, with the aim of extracting the tracklet information. As in all the other machine learning applications, a series of steps is required to obtain a working system: dataset creation, pre-processing, training, testing and post-processing. The dataset creation and pre-processing steps are needed to properly prepare the training step, while the post-processing one is applied to refine the output of the trained network. In order to succeed in the learning process, a sufficiently vast dataset composed by a set of images coupled with a label, in this case black&white masks (with a black background and a white tracklet), is needed. In order to be general, 360 realistic synthetic images and labels have been generated. Subsequently, images are rescaled (to speed up training) and normalized. The trained network is tested against a set of scaled synthetic and real images. Images are altered to reduce vignetting and to level out the background brightness, downconverted to 8 bit and normalized. The post-processing step performs a centroid identification and estimates the object right ascension and declination. The average processing time per real image is less than 4 s and bright tracklets are easily recognized with a mean centroid angular error of 0.25° in 75 % of test cases using a 2° field of view telescope
2020
71st International Astronautical Congress (IAC 2020)
File in questo prodotto:
File Dimensione Formato  
DEVIA01-20.pdf

Accesso riservato

Descrizione: Paper
: Publisher’s version
Dimensione 1.22 MB
Formato Adobe PDF
1.22 MB Adobe PDF   Visualizza/Apri

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11311/1162341
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus 0
  • ???jsp.display-item.citation.isi??? ND
social impact