Skip to Main content Skip to Navigation

Drone recognition with Deep Learning

Abstract : We work on the recognition of radar micro-Doppler signals of drones thanks to Deep Learning tools. The micro-Doppler phenomenon consists in a frequency modulation set created by the intern movements of the observed target. First, we analyze the different existing data : simulations, collected data available. We examine their limitations and carry out a measurement campaign to tackle them. Once our data is collected, we study the impact of the different space representations to propose a standard format adapted for Deep Learning. We continue our research on a major radar problem : the lack of data. Thus, we explore the data augmentation with GANs. We propose a measure of the quality of these algorithms based on utility criteria, and not on the realism of generated data. We observe a statistically significant improvement of classifications thanks to the signals generated by our GANs. Encouraged by this result, we implement more advanced GANs conjugating ground truth and real data. As we identified possible resolution axes we currently develop them.
Complete list of metadata
Contributor : ABES STAR :  Contact
Submitted on : Wednesday, April 13, 2022 - 2:53:10 PM
Last modification on : Friday, August 5, 2022 - 9:27:29 AM
Long-term archiving on: : Thursday, July 14, 2022 - 6:41:18 PM


Version validated by the jury (STAR)


  • HAL Id : tel-03640378, version 1


Julien Gérard. Drone recognition with Deep Learning. Signal and Image processing. Université Paris-Saclay, 2022. English. ⟨NNT : 2022UPASG002⟩. ⟨tel-03640378⟩



Record views


Files downloads