2018 19th International Radar Symposium (IRS) 2018
DOI: 10.23919/irs.2018.8447963
|View full text |Cite
|
Sign up to set email alerts
|

Temporal Deep Learning for Drone Micro-Doppler Classification

Abstract: Our work builds temporal deep learning architectures for the classification of time-frequency signal representations on a novel model of simulated radar datasets. We show and compare the success of these models and validate the interest of temporal structures to gain on classification confidence over time.

Help me understand this report

Search citation statements

Order By: Relevance

Paper Sections

Select...
2
1
1
1

Citation Types

0
26
0

Year Published

2019
2019
2024
2024

Publication Types

Select...
4
2
1

Relationship

2
5

Authors

Journals

citations
Cited by 38 publications
(26 citation statements)
references
References 17 publications
0
26
0
Order By: Relevance
“…To this end, we propose a partially complex network for which the output complex representation is log-scaled after the passage to absolute value, and heuristically study the impact on performance of the complex-to-real (C2R(x) = 10log(|x| 2 |)) function's position in the layer hierarchy. The conclusion is conceptually satisfying as it places the C2R right after the final temporal representation layer, ie right before the convolutionalized fully-connected layers [6], as represented in figure 1; in practice the ante-penultimate convolutional layer of the network proposed in [3]. This result leads to a rather natural interpretation: while the complex spectral representation of the signal in a real-valued network stops at the Fourier transform, the latter in a CRNet explores a hierarchy of further filter banks in addition to the Fourier filtering.…”
Section: Number Of Parametersmentioning
confidence: 85%
See 3 more Smart Citations
“…To this end, we propose a partially complex network for which the output complex representation is log-scaled after the passage to absolute value, and heuristically study the impact on performance of the complex-to-real (C2R(x) = 10log(|x| 2 |)) function's position in the layer hierarchy. The conclusion is conceptually satisfying as it places the C2R right after the final temporal representation layer, ie right before the convolutionalized fully-connected layers [6], as represented in figure 1; in practice the ante-penultimate convolutional layer of the network proposed in [3]. This result leads to a rather natural interpretation: while the complex spectral representation of the signal in a real-valued network stops at the Fourier transform, the latter in a CRNet explores a hierarchy of further filter banks in addition to the Fourier filtering.…”
Section: Number Of Parametersmentioning
confidence: 85%
“…In this section we show experimental results on synthetic radar data, issued by the simulator introduced in [3]. Approximately 20 minutes of signal are generated for each of 3 different classes of drones; signals are passed through the models 35ms at a time.…”
Section: Resultsmentioning
confidence: 99%
See 2 more Smart Citations
“…The Fourier transform, whether it be purely frequencial or varying in time, exhibits a plethora of properties allowing it to reveal even to the naked eye many interesting characteristics of the signal at hand [1]: as such, perhaps the most popular representation for radar data analysis is the windowed Fourier transform, or spectrogram, also called micro-Doppler signature [2], which takes the form of an image where spatial locality is replaced by temporal and frequencial locality. Recent radar classification schemes thus apply modern Computer Vision algorithms on micro-Doppler signatures, such as convolutional neural networks [3] (CNNs) or recurrent networks [4] (RNNs).…”
Section: Introduction and Related Workmentioning
confidence: 99%