Annonce

Les commentaires sont clos.

Proposition de thèse : Deep Learning for Radar Detection of Targets in Clutter

20 December 2021


Catégorie : Doctorant


Deep learning for radar detection of targets in clutter

Start of the thesis: 10/01/2022

Deadline to apply: 06/01/2022

Keywords: Target detection, Deep neural networks, Cognitive Radar

Required knowledge:

-Signal processing, statistical signal processing

-Learning methods, Neural network and deep learning

-Matlab, Python

-Knowledge of radar signal processing would be a plus

Host laboratory at ONERA, Department: Electromagnetism and Radar Department, Unit : MATS (Advanced Methods in Signal Processing)

Location (ONERA center): Palaiseau 91120

Contact: Christèle Morisseau, Phone. : +33(0) 1.80.38.62.99 Email: christele.morisseau@onera.fr

Thesis Director

Name: Chengfang Ren, Laboratory: SONDRA (CentraleSupelec, NUS)

E-mail : chengfang.ren@centralesupelec.fr

Pour plus d’informations : https://www.onera.fr/rejoindre-onera/la-formation-par-la-recherche

 

Context

Making a decision about the presence or absence of targets in the measured data is a critical step in the search process for a surveillance radar. This decision-making is made difficult by the presence of clutter echoes generated by reflections on the elements of the surrounding scene (ionosphere, waves, or buildings, trees, etc.). These clutter echoes can be particularly troublesome for target detection, especially when characterized by a speed band that is potentially that of targets of interest.

Current detection methods are based on adaptive processing and conventional processing methods such as pulse compression (or adaptive filtering) to make targets stand out from ambient noise or clutter. These adaptive treatments generally seek firstly to reduce the clutter statistics to a known or exploitable statistic, that is to say, making it possible to set up an effective detection test secondly. However, these methods depend on the clutter model, which is often unknown and complex, and therefore does not allow adaptive processing to be associated with it (the case of ionospheric clutter on a surface wave radar for example). The adaptive clutter processing also requires secondary data, the selection of which is complicated because it is generally distributed over an unknown zone (after distance, or Doppler, or angular processing) and in a limited time. All of these limitations reduce the performance of radar target detection algorithms.

Finally, some radars installed on the territory record measurements which are not used for the following measurements. For example, the ROS radar (surface wave HF radar) developed at ONERA, stores information on the distribution of ionospheric noise which is not used to recognize this noise on the signals subsequently received. One of the interests of learning is to use this information as secondary data.

 

Goals

 

The objective of this thesis is thus to study the potential contribution of deep learning methods in the context of the detection of multiple radar targets in the presence of clutter and noise.

It is thus a question, on the one hand of constituting very large learning bases by simulation of radar data, then by selection of real data (in particular data from the ROS radar), and on the other hand of setting up a learning network capable of ingesting this simulated data. The learning network can be set up after "classic" radar processing (that is to say pulse compression) or instead of this processing, or even integrated into the radar processing chain, beforehand. data processing. This thesis must therefore conclude on the best participation of the learning network in the radar detection chain, according to the criteria for evaluating the performance of the detector. Also, special attention will be placed on the network's ability to use datasets belonging to the complex vectors space, instead of real ones, as they are used in conventional" radar processing. Also it will be analyzed what type of measurement space is more relevant to use for the classification (Doppler / Distance or Azimuth) and if the learning principle should rather use measurements only related to clutter signals and / or containing targets. Finally, within the framework of the study with application to ROS data, this thesis may propose a network capable of discriminating between ships and planes by their signature.

Particular interest is given to the target detection and clutter rejection capabilities provided by the considered method, but also to :

• the types of decompositions provided by the first layers of the network, in order to possibly try to propose, subsequently, other opportunities of treatment having a more classic approach but inspired by these networks,

• joint detection and characterization capabilities of the targets of these networks.

The problem of this thesis can be generalized to many fields, such as detection for image processing, medical diagnosis or detection on complex signals of one dimension.

 

Planning

1st year :

- Understanding of the context, bibliographic study on classical radar processing and deep learning detection.

- Implementation of a simple simulation chain of radar data containing targets, clutter and noise. Selection and handling of measurement signals, brief bibliographic study on the acquisition chain and the propagation model of measurement signals.

- First tests of deep learning type methods for target detection. Study of the influence of the type of data: complex or real.

2nd year :

- Implementation of neural networks adapted to the extraction of targets from simulated and measured data. Proposal of several processing chains for radar detection. Determination of the best representation space for the measurements.

- Analysis of the performances and comparison of the proposed algorithms.

3rd year :

- Study of joint capacities for detection and characterization of targets by networks and / or,

- Study of the characterizations by the first layers of the network with a view to neuronal / adaptive hybrid processing implementation of target detection.

- Thesis manuscript.

 

The doctoral student will be hosted in the MATS (Advanced Methods in Signal Processing) unit of the Electromagnetism and Radar Department on the ONERA site in Palaiseau. Throughout his thesis, he will be required to train in the problematic of signal processing (detection, estimation), in processing by deep learning, as well as training in other subjects by following training courses offered in particular by the doctoral school.

 

References:

T. Bucciarelli, G. Fedele and R. Parisi, "Neural networks based signal detection," Proceedings of the IEEE 1993 National Aerospace and Electronics Conference-NAECON 1993, Dayton, OH, USA, 1993, pp. 814-818 vol.2, doi: 10.1109/NAECON.1993.290838.

F. Amoozegar and M. K. Sundareshan, "A robust neural network scheme for constant false alarm rate processing for target detection in clutter environment," Proceedings of 1994 American Control Conference - ACC '94, Baltimore, MD, USA, 1994, pp. 1727-1728 vol.2, doi: 10.1109/ACC.1994.752367.

Gálvez, Nélida & Pasciaroni, José luis & Agamennoni, Osvaldo & Cousseau, Juan. (2004). RADAR SIGNAL DETECTOR IMPLEMENTED WITH ARTIFICIAL NEURAL NETWORKS.

Mezzoug, Karim & Bachir, Djebbar. (2009). Étude Comparative d'un Détecteur CFAR Neuronal de Plusieurs Cibles Radar dans un Fouillis de type K-Distribution.

Gálvez, Nélida & Cousseau, Juan & Pasciaroni, José luis & Agamennoni, Osvaldo. (2012). Improved Neural Network Based CFAR Detection for non Homogeneous Background and Multiple Target Situations. Latin American applied research Pesquisa aplicada latino americana = Investigación aplicada latinoamericana.