Salm Jérémy, Leurent Gauthier, Debons Robin, Bachelor project, Spring 2023
This work is carried out within the Space Situation Awareness EPFL team (SSA), based at the Federal Institute of Technology in Lausanne (EPFL). It takes place as a improvement of the work done last year in the same association. First of all, it will be a question of making comparisons between some orbital parameters for several satellites obtained during this work with reference values. In particular, the study of their relative errors will be carry out. In addition, the angular variation for different temporal modifications will be examined for some of these parameters. Finally, several cases will be considered by modifying the angles that locate the satellites in order to understand the associated sensitivity.
Luca Hartman, Luc Luigi Guyot & Julien Thierry Burri, Bachelor project, Spring 2022
This report summarises the work done to provide a basis for the utilisation of the telescope and the processing of images taken by the Space Situational Awareness (SSA) École Polytechnique Fédérale de Lausanne (EPFL) association. Among these efforts, the Point Spread Function (PSF) was measured after taking into account the different optical aberrations. Satellite tracks have been modelled. Then, the influence of the environment has been quantified. Several perturbations have been studied such as light pollution between two locations and the clarity of the sky. Finally, the trajectory of satellites has been computed numerically. The convergence study of two different algorithms, Runge-Kutta 4 and Störmer-Verlet, has been made. The influence of several forces on the satellite has been studied.
Aymeric Labarbe, Anna Schwab & Philippe Gigon, Bachelor project, Spring 2022
In this paper, different ways of preliminary orbit determination from optical sightings with a telescope, are implemented and tested. Starting from long exposure images taken with a Celestron RASA 36 telescope, the aim is to detect satellites and satellite debris on the image, extract their position and interpolate a first estimate of the object’s orbit. The orbit determination is done with two different preliminary orbit determination algorithms, namely a Laplace and a Gaussian algorithm. The two algorithms are tested in various cases and with data provided by Stellarium. In the tested cases the two algorithms lead to similar results. In most of the tested cases the obtained heights are up to few percent of relative error close to the real orbits. Nevertheless, in some situations the preliminary orbit determination fails and large errors on the satellite’s height and velocity occur. It was not possible to uniquely identify the reasons for those failures, different studied parameters such as the elevation of the observed satellite and the time between two observations have been identified as critical parameters. Tests with real data have been done and give more imprecise results. It could be due to the difficulty of extracting the angles from the images. In fact, further algorithmic work is needed in order to improve those results.
Theo Patron, Master project, Spring 2022
The increasing number of satellites orbiting the earth strengthen the need of a catalog containing relevant information about each object. In order to create such a catalogue, satellite streaks have to be detected in astronomical images.
The goal of this project is to develop a reliable method that allows to detect the streaks on the images taken by the SSA’s telescope. This has to be done under the constraint of a low data regime.
Two approaches will be tested. The first one is based on the previous work from  and  and uses a U-Net based network to detect the streaks. Synthetic data and an iterative training will be used in order to train the network in an unsupervised way. The second approach is based on the work of  and  and uses a novel architecture that is called HT-LCNN. Transfer learning will be used here since we will start from the model pretrained on the ShanghaiTech dataset .
Even if some good results (0.7 mean IoU) are achieved with the first method, a pixel-wise data annotation is needed to assess the model performance. Furthermore the robustness of the network to unseen data cannot be assured since only a dozen real streaks were available at this time. The second method yields very good results even with the model trained on the ShanghaiTech dataset (0.844 sAP). Fine-tuning on our dataset allowed to improve the performance (0.988 sAP). This method is prefered since the annotation is easier to do and the training is easily extendable to new data.
Falkenback Tanguy, Minor project, Spring 2022
This report is a study of different systems to detect space objects using radio frequencies. Three systems are studied to determine the best option: bistatic radar, Frequency Modulated Continuous Waves (FMCW) and passive detection. The purpose of the detection is to gain information on the location, speed and range of the object through signal and data processing.
Alexandre Di Piazza, Master project, Spring 2021
This report studies the detection of slow satellite tracks on single epoch images produced by OMEGACAM on the VLT Survey Tele- scope. We first generate synthetic data on which we train a Neural Network, and then fine-tune our network by freezing some layers and training again on real images of slow satellite tracks for a better gen- eralization to true images. It is in continuation with another semester project on the same subject.
Noah Kaltenrieder, Bachelor project, Spring 2021
We present an approach to extract the intensity along the satellite streaks that can be found in images that were obtained by OMEGACAM on the VLT Survey Telescope. In this project, we only analyzed the long tracks that were created by high velocity objects, to try to get an estimation about their rotation. We used Fourier transform to get the periodicity of the intensity along the tracks. Due to the lack of real data about rotation, we first had to create fake streaks with sinusoid intensity and different angles. We manually got the tracks to run the code on them, so ideally this project should be integrated with DetectSat  so that all the process can be done automatically. We found very great result with them and run the pipeline on real tracks. By making assumptions about the satellite, like their altitude, we get possible results on real streaks, but since we don’t have access to the real values, we can’t say if it is close to reality or not.
Manon Béchaz, Baptiste Claudon, Jules Eschbach, Salomon Guinchard, Bachelor project, Spring 2021
Avec l’augmentation exponentielle des objets en orbite, la mise en place de méthodes de recensement et de catalogues est aujourd’hui essentielle. Ce travail s’attelle à démontrer la possibilité d’utiliser du matériel optique accessible au grand public pour la détection, depuis le sol, de ces objets en orbite. Trois montages avec différents matériels on été testés, et se sont avérés pour deux d’entre eux à mêmes de détecter efficacement des traces de satellites. En parallèle, des algorithmes ont été développés pour traiter les images obtenues et en détecter automatiquement les traces, permettant le calcul automatique d’orbites et donc la mise en place d’un catalogue. Le traitement total des images s’est avéré à même de détecter correctement 22% des centres des traces observables, permettant d’obtenir 11 orbites sur les 117 images considérées pour l’étude statistique.
Yann Bouquet, Master project, Fall 2020
We present two approaches to detect and extract information about satellite streaks in single epoch images produced by OMEGACAM on the VLT Survey Telescope. We distinguish long tracks from high velocity objects, usually crossing an important part of the image, and short tracks from low velocity objects. The first method is based on image processing technique. By applying filters and morphological transforms on the image we manage to detect the long satellite streaks with Hough transform technique. Facing the limitations of this approach for short tracks, we propose a neural network model inspired from the UNet model, for image segmentation, and train it on synthetic data to identify pixels belonging to satellites streaks with a f1 score of 95%. We show the feasibility of our methods by using example images the gravitationally lensed quasar fields SDSSJ0924+0219.
Dunant Joachim, Master project, Spring 2020
This project is divided in two parts. The first part will focus on building a database which efficiently supports insertion of newly detected objects and the load of information about each space objects in LEO, by doing trajectory simulations of these objects through time. Multiple database structures will be explored, by looking at different ways to compute and optimise a Nearest Neighbour (NN) search. Then the goal of the second part is to explore (but not implement) multiple methods for the computation of collision probabilities of each satellite and debris, to see their pros and cons in the case of space orbits. In this manner, we can achieve the objective of having a regularly updated database containing all the objects detected in LEO and that can also provide an effective basis for calculating the risk of collision between every space object.