Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
2933972
Figure 1.
High-level block diagram of a cognitive radar (adapted from Guerci [2]).
during a measurement campaign at the NLR RPAS Test UAV body is removed to center the Doppler spectrum
Centre in Marknesse, The Netherlands. The mini-UAVs around 0 Hz. Examples of the measured micro-Doppler
included two types of fixed wing aircraft (Robbe Air spectrograms for all 5 mini-UAVs can be seen in Figure 2.
Trainer 140 and Sky Walker X8), an Align T-REX 550 The coherent integration time of the spectrograms is
helicopter, a Mikrokopter QuadroXL quadcopter, and a 10.7 ms. The modulations of the radar signal due to the
Mikrokopter OktoXL octocopter, see Figure 2. moving parts can clearly be observed.
Measurements of the mini-UAVs were conducted with
an experimental continuous wave radar that operates at a
frequency of 9.5 GHz and samples the radar signal at a
rate of 96 kHz. The complex radar signals are converted
MICRO-DOPPLER SIGNAL SIMULATIONS
to micro-Doppler spectrograms using a short time fourier In addition to the measurements, micro-Doppler signals
transform (STFT). The Doppler frequency of the mini- have been simulated using computer models of mini-
Figure 2.
Photos (1st row), measured spectrograms (2nd row), computer models (3rd row), and simulated spectrograms (4th row) of five mini-UAVs.
Figure 3.
Workflow for the classification of mini-UAVs with deep neural networks using micro-Doppler spectrograms.
UAVs to test the idea that a deep neural network can be CLASSIFICATION OF MINI-UAVS WITH DEEP NEURAL
trained with simulated micro-Doppler spectra. The trained
neural network can then be used to classify micro-Doppler NETWORKS
spectra measured with a real radar system during opera-
Figure 3 shows the workflow for the classification of mini-
tions. This concept would reduce the need for expensive UAVs with deep neural networks. The training phase con-
and time-consuming radar measurements to populate the sists of generating a training set with (a) simulated radar
knowledge base of a cognitive radar. It also could enable signals using target models or (b) real target signals mea-
the classification of mini-UAVs of which physical charac- sured with a radar. Micro-Doppler spectrograms are cre-
teristics are available, but no measurements can be made ated by applying an STFT successively to overlapping
due to security reasons. sequences of the time-domain signals. The preprocessing
The simulation of micro-Doppler spectra of targets, includes removal of the speed of the mini-UAV body, and
such as mini-UAVs is difficult because of the complex
normalization of the power spectrogram. In the test phase,
mechanical structure of these targets and the time varying the trained neural network is used to classify the measured
characteristic of the radar cross section (RCS) due to the radar signals.
moving parts. To limit the computation time for micro- Table 1 shows the number of micro-Doppler spectro-
Doppler signal simulations, an approach was chosen in grams that have been extracted from the measured
which the mini-UAVs are represented with so-called and simulated radar signals for the training and test sets
principal scatterers for which the analytic expressions of in the mini-UAV classification experiments described
their RCS exist [12]. Examples of these principal scatter- below.
ers are point scatterers, cylinders, ellipsoids, flat triangu-
lar plates, and thin wires. For each sample of the radar
signal, the position and attitude of each principal scat- Table 1.
terer is determined and the associated RCS for this scat-
terer is computed. Finally, the reflected radar signal for Micro-Doppler Spectrograms for the Five Mini-UAVs
the entire target is calculated by adding the contributions
of all principal scatterers. This simulation approach is rel- Training Test
atively fast, but neglects the occlusion of some scatterers Target
by other scatterers on the body, multiple reflections, dif- Simulated Measured Simulated Measured
fraction, and materials other than metal. Figure 2 shows
the computer models of the five mini-UAVs considered Air trainer 506 570 525 499
in this paper and examples of the associated simulated Skywalker 475 886 545 775
micro-Doppler spectrograms. Figure 2 clearly shows a
T-REX 550 965 1284 1080 1092
resemblance between the measured and the simulated
micro-Doppler spectrograms, but also some measured QuadroXL 477 841 546 869
characteristics that are not present in the simulated OktoXL 480 1370 555 1222
spectrograms.
Table 2.
Deep Neural Network Test Cases for Classification of Mini-UAVs Using Micro-Doppler Spectrograms
Figure 4.
Adversarial training of an auto-encoder with the GANomaly method to detect spectrograms of unknown targets.
the decimation. This simple and effective preprocessing application of a threshold to the output of the Soft-max
step is apparently difficult to learn by the CNN from the layer of a CNN [15]. If the maximum output of the Soft-
limited number of training samples in test case C. max layer does not exceed the threshold, an unknown tar-
Another conclusion is that the reduction in coherent get class is declared. This Soft-max approach is also
integration time in the Doppler filtering from 10.7 ms to referred to as a reject option for a classifier [16]. By vary-
2.7 ms also gives a significant improvement in classifica- ing the value of the threshold, a receiver operation charac-
tion accuracy. This improvement can be explained by the teristic (ROC) curve can be obtained, which shows the
fact that the modulations by the rotating propeller or rotor probability of detecting an unknown target class (true pos-
blades are less smeared in the Doppler spectrum due to itive rate) as a function of the probability of falsely declar-
the shorter integration time, and the changes in the Dopp- ing an unknown target class (false positive rate).
ler spectrum become clearer. An alternative to the Soft-max approach called GANo-
Although the use of an LSTM-RNN proved to be less maly is based on an auto-encoder that is trained in an
accurate than the CNN in the overall classification, an adversarial manner [17]. This technique has been used to
important advantage of an LSTM-RNN is that it can pro- investigate the potential of deep learning to screen avia-
vide good classification results already after a few coher- tion luggage for anomalous items using X-ray screening.
ent processing intervals, whereas the CNN only provides a Figure 4 shows the configuration of the GANomaly net-
classification result after an entire spectrogram has been work that consists of three subnetworks. The auto-encoder
processed [14]. Another advantage of the LSTM-RNN is acts as a generator network that learns a compact latent
that it is capable to deal with transitions in target behavior, space representation z of the input spectrograms x by try-
e.g., a fixed wing mini-UAV that takes off vertically, and ing to reconstruct the input spectrogram. The encoder
then, proceeds to horizontal flight. compresses the reconstructed spectrogram to an estimate
of the latent representation using the same architecture as
the encoder in the auto-encoder. The discriminator deter-
mines, with an encoder architecture, if the input is a real
DETECTION OF UNKNOWN TARGET CLASSES spectrogram or a fake spectrogram.
During military operations, cognitive radars will often be The detection of a spectrogram originating from an
employed in conditions where novel target classes are unknown target assumes that the auto-encoder is not able
observed of which there are no examples of micro-Dopp- to reconstruct such a spectrogram accurately because the
ler spectrograms in the training set of the deep neural net- network is trained only with spectrograms from known
work. In this case, the cognitive radar should detect the target classes. A reconstructed spectrogram for an
presence of this unknown target class and may decide to unknown spectrogram will also lead to discrepancies
schedule specific radar measurements to add spectrograms between the latent vector and its estimate.
of the unknown target class to the training set. This discrepancy is used by the GANomaly method
The detection of unknown target classes can be during the test phase to detect a spectrogram of an
achieved in several ways. The simplest approach is the unknown target class.
Figure 5.
ROC curve showing the true positive rate versus the false positive
rate for detecting an unknown Align T-REX 550 mini-UAV with Figure 6.
the Soft-Max and GANomaly methods. AUC results for the detection of unknown target classes with Soft-
max and GANomaly. The error bars represent variations due to the
use of random seeds for the initialization of the neural networks.
The objective function for training the GANomaly net-
works consists of three loss functions. The adversarial loss
function Ladv computes the L2 distance between features network and hyperparameters are as described in
computed by the function f from an intermediate layer of “CLASSIFICATION OF MINI-UAVS.” The models for
the discriminator network for the real and generated spec- GANomaly and Soft-max are trained for, respectively, 15
trograms. This feature matching approach reduces the and 50 epochs.
instability of GANs during training [18]. The context loss Figure 6 shows that both the Soft-max and the GANo-
function Lcon penalizes the auto-encoder for reconstruction maly achieve AUC ranging from 0.5 to 0.8. Note that
errors by measuring the L distance between the real and comparing the performance of deep learning methods is
generated spectrograms. Finally, the encoder loss Lenc often tricky because the result can highly depend on the
measures the L1 distance between the latent representations effort that has been put in optimizing the networks, train-
of the input and reconstructed spectrograms. ing parameters, and data preprocessing steps. Especially
Experiments with the Soft-max and GANomaly methods adversarial training methods, such as GANomaly, are
for the detection of unknown target classes have been con- notorious for instability issues. In the experiments, it was
ducted using the dataset with mini-UAV spectrograms, as observed that the final performance of the GANomaly
described in “CLASSIFICATION OF MINI-UAVS.” Each method was not yet fully satisfactory, and perhaps better
of the mini-UAVs has been used alternatively as an unknown performances can be achieved with other GAN configura-
target class when training the neural networks on the micro- tions for the detection of unknown target classes.
Doppler spectrograms from the remaining ‘“known” mini-
UAVs. Figure 5 shows an example of an ROC curve for a
test where the T-REX 550 helicopter was used as an
unknown target class. The GANomaly method performs bet-
DENOISING OF SPECTROGRAMS
ter at low and high false positive rates, whereas the Soft-max The micro-Doppler spectrograms collected during the
method works better at intermediate false positive rates. measurement campaign described in “CLASSIFICATION
Figure 6 shows the area-under-curve (AUC) for the OF MINI-UAVS” have a relatively high signal-to-noise
detection of unknown target classes with the Soft-max and ratio (SNR) due to the short range between the radar and
GANomaly methods. Each experiment is performed three the mini-UAVs. However, the SNR will be lower at longer
times to capture the performance variations due to the random ranges, and therefore, the performance of a target classifier
initialization of the neural networks and due to the random will degrade. To investigate if deep learning can be used
selection of the data samples in the test set and training set. to improve the SNR of the spectrograms in the preprocess-
The above results are obtained using the measured ing before classification, and hence, increase the range at
spectrograms of test case E. All spectrograms are truncated which a cognitive radar can recognize a mini-UAV, an
after the first 64 time samples resulting in square input experiment was conducted with a denoising adversarial
data of 6464 pixels. This input size enabled the use of auto-encoder. Adversarial trained auto-encoders can be
network configuration and training parameters described used to restore input data, such as images that are cor-
in the paper by Akcay et al. [17] that are proven to be rupted by noise [19]. The architecture of the denoising
successful for the GANomaly method. For Soft-max, the adversarial auto-encoder is like the architecture of the
Figure 7.
Denoising adversarial auto-encoder applied to spectrograms corrupted with synthetic noise.
GANomaly network, as shown in Figure 4, except that the imbalance in the training set used in this paper
separate encoder is omitted and the objective function for (see Table 1) is approximately a factor of two. This imbal-
training of the auto-encoder only consists of the mean ance can have a negative impact on the convergence of
square error between noise corrupted versions of the spec- the classifier during training and the overall accuracy [20].
trograms and the uncorrupted spectrograms. There are several ways to mitigate the effect of an imbal-
Figure 7 shows two examples of measured micro- anced training set. The first category concerns methods
Doppler spectrograms of a T-REX 550 helicopter with a that leave the training set intact and modify the training
high SNR in the first column. The second column shows procedure or the classifier to deal with the imbalance. The
the same spectrograms to which synthetic Gaussian noise second category involves changes to the training set itself
has been added to lower the SNR. The third column shows by generating data for classes that are underrepresented.
the reconstructed spectrograms using a denoising adver- In this section, a generative network called InfoGAN has
sarial auto-encoder. Although some details of the micro- been investigated for the generation of realistic training
Doppler modulations are distorted, it is still possible to data for underrepresented target classes [21]. When com-
recognize the micro-Doppler contribution due to the tail pared with a standard GAN, an InfoGAN has the advan-
rotor and, up to a certain extent, the blade tip contribution. tage that it learns in an unsupervised way, interpretable
In a second experiment, the denoising adversarial and disentangled representations of challenging datasets.
auto-encoder has been tested with measured spectrograms Figure 9 shows the architecture of an InfoGAN network
that are characterized by a lower SNR. As apparent from
the reconstructed spectrograms shown in Figure 8, a signal
to background ratio enhancement in the order of 20 dB is
achieved by the denoising auto-encoder. Furthermore,
information associated to tail rotor signature and blade tip
rotation is partially restored. Although the SNR of the
spectrograms seems to be significantly enhanced by the
denoising adversarial auto-encoder, the impact of this pre-
processing on the accuracy of a target classifier, such as a
CNN still has to be investigated.
Figure 10.
InfoGAN applied to micro-Doppler spectrograms with examples of synthesized spectrograms per category of the latent code.
integration time in the Doppler filtering had a higher Dutch Radar Center of Expertise (D-RACE), a strategic
impact on the classification performance than the type of alliance between Thales Nederland B.V. and TNO.
the deep neural network (CNN versus LSTM-RNN) in the
classifier. The use of simulated spectrograms to classify
measured spectrograms requires a higher fidelity modeling REFERENCES
of the micro-Doppler signals than the modeling approach
[1] S. Haykin, Y. Xue, and P. Setoodeh, “Cognitive radar: Step
based on principal scatterers used in this paper.
toward bridging the gap between neuroscience and engi-
When a cognitive radar measures a micro-Doppler
neering,” Proc. IEEE, vol. 100, no. 11, pp. 3102–3130,
spectrogram from a target class that is not represented in
Nov. 2012.
the training set, an anomaly detection algorithm could pro-
[2] J. R. Guerci, Cognitive Radar: The Knowledge-Aided
vide a trigger to the cognitive radar scheduler to collect
Fully Adaptive Approach, Norwood, MA, USA: Artech
micro-Doppler spectrograms from this unknown target.
House, 2010.
The performance of two deep learning techniques, Soft-
[3] M. S. Greco, F. Gini, P. Stinco, and K. Bell, “Cognitive
max and GANomaly, for the detection of spectrograms
radars: On the road to reality: Progress thus far and possibil-
from unknown target classes has been compared. Both
ities for the future,” IEEE Signal Process. Mag., vol. 35,
techniques had a similar performance with respect to the
no. 4, pp. 112–125, Jul. 2018.
probability of detecting unknown classes with the GANo-
[4] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,”
maly technique showing a larger performance variation
Nature, vol. 521, no. 7553, pp. 436–444, 2015.
due to the difficulty in training adversarial networks. Dif-
[5] D. A. E. Morgan, “Deep convolutional neural networks for
ferent GAN architectures for anomaly detection that are
ATR from SAR imagery,” Proc. SPIE, vol. 9475, 2015,
easier to train, such as a Wasserstein GAN may provide a
Art. no. 94750F.
better performance [22].
[6] Y. Kim and T. Moon, “Human detection and activity clas-
A denoising adversarial auto-encoder has been trained
sification based on micro-Doppler signatures using deep
to remove noise from spectrograms. An improvement of
convolutional neural networks,” IEEE Geosci. Remote
20 dB in the SNR of a measured spectrogram has been
Sens. Lett., vol. 13, no. 1, pp. 8–12, Jan. 2016.
observed. If this denoising technique does not alter the €
[7] M. S. Seyfioglu, S. Z. G€urb€uz, A. M. Ozbayo glu, and
micro-Doppler characteristics significantly, a considerable
M. Y€uksel, “Deep learning of micro-Doppler features for
increase in the range at which a cognitive radar can recog-
aided and unaided gait recognition,” in Proc. IEEE Radar
nize mini-UAVs could be achieved.
Conf., 2017, pp. 1125–1130.
To mitigate the effects of class imbalances in the num-
[8] R. P. Trommel, R. I. A. Harmanny, L. Cifola, and
ber of training examples on the performance of a classi-
J. N. Driessen, “Multi-target human gait classification
fier, an InfoGAN can be used to generate additional
using deep convolutional neural networks on micro-
spectrograms of underrepresented classes. A quantitative
Doppler spectrograms,” in Proc. Eur. Radar Conf., 2016,
study concerning the benefits of this approach will be con-
pp. 81–84.
ducted in the future research.
[9] B. Dekker, S. Jacobs, A. S. Kossen, M. C. Kruithof,
The overall conclusion of this paper is that deep learn-
A. G. Huizing, and M. Geurts, “Gesture recognition with a
ing techniques have a high potential for improving the
low power FMCW radar and a deep convolutional neural
classification of mini-UAVs using micro-Doppler spectro-
network,” in Proc. Eur. Radar Conf., 2017, pp. 163–166.
grams in cognitive radar. However, the selection and tun-
[10] P. Molchanov, K. Egiazarian, J. Astola, R. I. A. Harmanny,
ing of the appropriate neural network architectures for
and J. J. M. de Wit, “Classification of small UAVs and birds
different processing tasks, such as classification and
by micro-Doppler signatures,” in Proc. Eur. Radar Conf.,
denoising is tedious and time consuming. A better under-
2013, pp. 172–175.
standing of the theory behind deep neural networks, poten-
[11] R. Harmanny, J. J. M. de Wit, and G. Premel-Cabic,
tially enabled by the link with compressive sensing and
“Radar micro-Doppler mini-UAV classification using
sparse signal representations, would not only accelerate
spectrograms and cepstrograms,” Int. J. Microw. Wireless
the development process, but also improve the acceptance
Technol., vol. 7, no. Special Issue 3/4, pp. 469–477, 2015.
of these advanced techniques in military radar
[12] V. C. Chen, F. Li, S.-S. Ho, and H. Wechsler, “Micro-
applications.
Doppler effect in radar: Phenomenon, model, and simula-
tion study,” IEEE Trans. Aerosp. Electron. Syst., vol. 42,
no. 1, pp. 2–21, Jan. 2006.
ACKNOWLEDGMENT
[13] P. J. Speirs, A. Schr€oder, M. Renker, P. Wellig, and
TNO has received funds for this study from the Netherlands A. Murk, “Comparisons between simulated and measured
Ministry of Defense, under Grant V1512 and Grant V1908. X-band signatures of quad-, hexa- and octocopters,” in
The study has been conducted within the scope of the Proc. 15th Eur. Radar Conf., 2018, pp. 325–328.
[14] G. Klarenbeek, R. I. A. Harmanny, and L. Cifola, “Multi- [19] A. Creswell and A. A. Bharath, “Denoising adversarial
target human gait classification using LSTM recurrent neu- autoencoders,” IEEE Trans. Neural Netw. Learn. Syst.,
ral networks applied to micro-Doppler,” in Proc. 14th Eur. vol. 30, no. 4, pp. 968–984, Apr. 2019.
Radar Conf., 2017, pp. 167–170. [20] M. Buda, A. Makib, and M. A. Mazurowski, “A system-
[15] A. Bendale and T.E. Boult, “Towards open set deep atic study of the class imbalance problem in convolutional
networks,” in Proc. IEEE Conf. Comput. Vis. Pattern Rec- neural networks,” Neural Netw., vol. 106, pp. 249–259,
ognit., Jun. 2016, pp. 1563–1572. 2018.
[16] D. M. J. Tax and R. P. W. Duin, “Growing a multi-class [21] X. Chen, Y. Duan, R. Houthooft, J. Schulman, I. Sutskever,
classifier with a reject option,” Pattern Recognit. Lett., and P. Abbeel, “InfoGAN: Interpretable representation
vol. 29, no. 10, pp. 1565–1570, 2008. learning by information maximizing generative adversarial
[17] S. Akcay, A. Atapourabarghouei, and T. P. Breckon, nets,” in Proc. 30th Int. Conf. Neural Inf. Process. Syst.,
“GANomaly: Semi-supervised anomaly detection via 2016, pp. 2172–2180.
adversarial training,” in Proc. ACCV, Perth, Australia, [22] T. Schlegl, P. Seeböck, S.M Waldstein, G. Langs, and
Dec. 2018. U. Schmidt-Erfurth, “f-anogan: Fast unsupervised anomaly
[18] T. Salimans, I. Goodfellow, W. Zaremba, W. Cheung, detection with generative adversarial networks,” Medical
A. Radford, and X. Chen, “Improved techniques for train- image analysis, vol. 54, pp. 30–44, May 2019.
ing GANs,” in Proc. 30th Int. Conf. Neural Inf. Process.
Syst., 2016, pp. 2234:2242.