LSTM-based System for Multiple Obstacle Detection using Ultra-wide

Band Radar

Amira Mimouna

1,2 a

, Anouar Ben Khalifa

2 b

, Ihsen Alouani

1 c

, Abdelmalik Taleb-Ahmed

1 d

Atika Rivenq

and Najoua Essoukri Ben Amara

IEMN-DOAE, UMR CNRS 8520, Polytechnic University Hauts-de-France, 59300 Valenciennes, France

Universit

e de Sousse, Ecole Nationale d’Ing

enieurs de Sousse, LATIS - Laboratory of Advanced Technology

and Intelligent Systems, 4023, Sousse, Tunisie

Keywords:

Obstacle Detection, UWB Radar, Deep Learning, LSTM, Intelligent Transportation Systems.

Abstract:

Autonomous vehicles present a promising opportunity in the future of transportation systems by providing

road safety. As signiﬁcant progress has been made in the automatic environment perception, the detection of

road obstacles remains a major challenge. Thus, to achieve reliable obstacle detection, several sensors have

been employed. For short ranges, the Ultra-Wide Band (UWB) radar is utilized in order to detect objects in

the near ﬁeld. However, the main challenge appears in distinguishing the real target’s signature from noise

in the received UWB signals. In this paper, we propose a novel framework that exploits Recurrent Neural

Networks (RNNs) with UWB signals for multiple road obstacle detection. Features are extracted from the

time-frequency domain using the discrete wavelet transform and are forwarded to the Long short-term memory

(LSTM) network. We evaluate our approach on the OLIMP dataset which includes various driving situations

with complex environment and targets from several classes. The obtained results show that the LSTM-based

system outperforms the other implemented related techniques in terms of obstacle detection.

1 INTRODUCTION

Seemingly during few years, autonomous vehicles

have turned a major interesting topic for researchers.

This is due to the exploitation of artiﬁcial intelli-

gence in the automotive ﬁeld, which opened the way

to the deployment of self-driving cars. In fact, self-

driving cars have the potential to radically transform

the road’s environment by saving road users’ lives,

reducing pollution and fuel savings (Skrickij et al.,

2020) (Khalifa et al., 2020).

Real-time obstacle detection is one of the main

challenges presented in developing perception sys-

tems for Intelligent Transportation Systems (ITS)

(Tarchoun et al., 2020b) (Tarchoun et al., 2020a). To

achieve this, an ITS deploys numerous sensors. The

widely employed sensors are cameras, lidar and radar

sensors (Koci

c et al., 2018). These sensors can be

https://orcid.org/0000-0003-3658-643X

https://orcid.org/0000-0002-9946-0829

https://orcid.org/0000-0001-5102-8087

https://orcid.org/0000-0001-7218-3799

used in different circumstances, as each sensor has

typical advantages and disadvantages in this ﬁeld of

research (Jegham et al., 2018). In this work, we focus

our attention on radar sensors.

The radar is a radio frequency-based sensor that

transmits electromagnetic waves and determines the

vehicle surrounding characteristics based on the re-

ﬂected echo. Recently, the radar sensor has played

a key role in military applications (Lee et al., 2020),

people counting (Choi et al., 2017) and positioning

(Nguyen and Pyun, 2015), and it is speciﬁcally em-

ployed in the automotive ﬁeld (Hakobyan and Yang,

2019). In fact, the radar is able to detect obstacles in

the vehicle’s surrounding by estimating their range,

velocity and angle. It is robust to bad weather con-

ditions like the fog and rain, and it is not affected by

poor light circumstances. Furthermore, it is employed

in short and long distances depending on the devel-

oped advanced driver assistance systems. In fact,

narrow-band radars are specially utilized to detect ob-

stacles in long ranges. For short range applications,

the Ultra-Wide Band (UWB) technology is the most

exploited in this area of research as it provides rich

418

Mimouna, A., Ben Khalifa, A., Alouani, I., Taleb-Ahmed, A., Rivenq, A. and Ben Amara, N.

LSTM-based System for Multiple Obstacle Detection using Ultra-wide Band Radar.

DOI: 10.5220/0010386904180425

In Proceedings of the 13th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2021) - Volume 2, pages 418-425

ISBN: 978-989-758-484-8

information. The UWB radar emits narrow pulses

with low energy in the order of picoseconds to few

nanoseconds (Saito et al., 2003). This type of radar

has been in the spotlight owing to its high resolution,

good penetration rate, and anti-interference ability ac-

quired from other devices (Zetik et al., 2007).

The preeminent property of such a radar consists in

the deformation of the emitted pulse. This distortion

depends on the obstacle’s characteristics, thereby it is

labeled as the object signature. This signature is im-

pacted by the object shape, material and size. For ex-

ample, the signature of a metallic object has higher

amplitude than a pedestrian one. Accordingly, the

employment of such a technology remains promising

to detect objects in near ranges. The data acquired

from the UWB radar can be represented under two

forms: a 1 dimensional (1D) signal, which is the re-

ﬂected echo, and 2 dimensional (2D) data that can be

a 2D feature map or a converted image.

Despite the fact that the UWB radar provides rich

information, one of the fundamental challenges that

could occur is distinguishing the real target’s signa-

ture from noise. For this purpose, various studies

have been conducted by proposing UWB-based de-

tectors. These detectors can be characterized into

two categories: hand crafted-based systems and deep

learning-based systems.

The most popular hand crafted-based detection

algorithms are: the correlation technique (Sakkila

et al., 2006), the Constant False Alarm Rate (CFAR)

(Rohling, 1983) based techniques and the Higher Or-

der Statistics (HOS) (Mendel, 1991). The aforemen-

tioned algorithms have been exploited ever since the

development of the conventional radar. CFAR-based

approaches analyze the characteristics of the signal’s

noise in order to deﬁne a threshold. The target is de-

tected if its magnitude is above the deﬁned thresh-

old. A HOS-based technique is a correlation-based

method that uses higher-order moment spectra.

Currently, with the signiﬁcant advancements con-

ducted by deep learning techniques, there exist vari-

ous studies that have investigated it with UWB signal

measurements, particularly in indoor environments

(Briso et al., 2019) (Tong et al., 2020) (Kumar et al.,

2019). Even though the interest of processing UWB

data via deep learning techniques is growing, there is

no work that has deeply treated UWB signals for out-

door environment perception to the best of our knowl-

edge. In this paper, we propose the ﬁrst framework

that exploits UWB signals with a Long Short-Term

Memory (LSTM) network in an outdoor environment

involving complex urban driving situations for multi-

target detection. As the received UWB signal is time-

series data, the exploitation of a RNN is suitable to ex-

hibit the temporal dependencies. Therefore, the main

intention of the developed network is to discriminate

the real target from noise within a received UWB sig-

nal.

In this work, the main contributions are the fol-

lowing:

• An overview of the existing deep learning-based

methods using UWB signals is presented.

• A proposal of the ﬁrst framework that exploits

LSTM with UWB signals for multi obstacle de-

tection in an outdoor complex environment.

• A comparison between our suggested approach

and the state-of-the-art techniques based on ex-

panded experiments using an extensive environ-

ment perception dataset (Mimouna et al., 2020)

is detailed. The obtained results show that our

method outperforms the conventional techniques.

The remainder of this paper is arranged as follows.

In section 2, a review of related work using UWB sig-

nals is presented. The proposed approach is detailed

in section 3. In section 4, we expose the experimental

setup and the obtained results. Finally, we conclude

the paper in section 5.

2 RELATED WORK

In subsection 2.1, we present the reviewed work that

relies on using the UWB radar with hand-crafted de-

tectors. For subsection 2.2, we review the deep UWB-

based systems.

2.1 Hand-crafted UWB-based Systems

Most of the studies on UWB-based obstacle detec-

tion systems have taken advantage of the well-known

correlation technique, CFAR and HOS techniques.

These techniques are known as hand-crafted based

detectors. In (Sakkila et al., 2008), the correlation

was used to detect either a car, a metal plate, a mo-

torway barrier or a pedestrian. The HOS technique

was commonly employed to suppress the Gaussian

noise (Mendel, 1991). Several algorithms have been

extended from the CFAR including the Cell Averag-

ing CFAR (CA-CFAR) (Conte et al., 1988), the order

static CFAR (Blake, 1988), the Smallest Of CFAR

(Cai et al., 2010), etc. A new thresholding method

based on the CFAR technique for UWB-based de-

tection application was put forward in (Quan et al.,

2020). The approach took into consideration false

alarm and miss-detection criteria. A new parameter,

entitled constant miss-detection rate (CMDR), is de-

ﬁned. The ﬁnal threshold was calculated by adding

LSTM-based System for Multiple Obstacle Detection using Ultra-wide Band Radar

419

the CFAR and CMDR rates. The experiments showed

good performances, although they were just carried

out in indoor environment using self-recorded dataset.

In the work presented in (Kim, 2020), metal lanes

would be detected if the energy value of the received

UWB signal surpassed a deﬁned threshold. In (Sadli

et al., 2018), a UWB radar detector for ITS applica-

tions was proposed by combining the CA-CFAR and

HOS techniques. Nevertheless, the experiments were

performed under controlled scenarios.

In fact, most of the aforementioned related work

has been a threshold-based methods. Otherwise, these

techniques depend essentially on the amplitude of the

object’s signature. Moreover, it should be pointed that

the considered environment and scenarios are con-

trolled using restrained datasets.

2.2 Deep Learning UWB-based Systems

To the best of our knowledge, UWB-based systems

that rely on deep learning techniques have been em-

ployed only for indoor applications such as: activity

recognition (Sadreazami et al., 2018), indoor people

localization (Poulose and Han, 2020) and recognition

of movements during sleep (Piriyajitakonkij et al.,

2020).

Deep learning methods have been used with 2D

UWB data, so, this dilemma is considered as an image

processing-based challenge. In (Chen et al., 2019), an

SFCW-UWB radar was used for fall detection, which

generated a time-frequency spectrum as UWB data.

Hence, the employed deep learning architecture was

the ﬁne-tuned Alexnet model. The authors in (Ko and

Lee, 2016) converted the time-series UWB data to

a time-frequency representation by Stockwell trans-

form. Afterwards, the reshaped images serve as in-

puts of the Convolutional neural network(CNN). The

exploited CNN model was LeNet. In (Ahmed and

Cho, 2020), UWB data were stored as a 2D matrix

including the slow-time and the fast-time properties.

Subsequently, it was converted to a grayscale image,

and later to an RGB one. GoogLeNet was adopted as

a deep learning architecture for hand gesture recogni-

tion.

In fact, 2D radar imaging based systems are no

longer considered a signal processing dilemma, but

rather an image processing challenge.

Otherwise, 1D UWB signals are likewise em-

ployed with deep learning for indoor applications.

To enhance transportation safety, a UWB radar is

installed in the rear view mirror to estimate the num-

ber and location of the in-vehicle people. A multi-

layer perception is employed where the time-sampled

radar signal data are the input of the network. To de-

ﬁne the suitable parameters, the number of hidden

layers is adjusted. Compared to machine learning

techniques, the suggested network achieves better re-

sults (Lim et al., 2020). For activity recognition, a

CNN-LSTM network using three UWB radars is im-

plemented in (Maitre et al., 2021). The features are

extracted using a CNN architecture that includes: two

1D convolution layers with 64(1x3) ﬁlters, a Relu ac-

tivation function, and a 1D max pooling layer. Sub-

sequently, the output is ﬂattened in to a 1D vector to

feed the LSTM network that contains 2 LSTM lay-

ers. Promising results are obtained. Jiang et al. (Jiang

et al., 2020) trained a CNN-LSTM to classify Line-

Of-Sight and None-Line-Of-Sight signals in the con-

text of indoor positioning applications. The UWB

channel impulse response was used as input of the

CNN that deployed two conventional layers. After-

wards, the CNN outputs were linked to the LSTM

network. A bi-direction LSTM and a stacked LSTM

were used. The achieved accuracy was equal to 81%,

but the training dataset was limited.

The aforementioned applications have considered

only indoor environments, using either 2D radar

imaging or 1D UWB data. Nevertheless, the ITS en-

vironment includes complex driving situations with

various types of targets. Moreover, in our case, the

choice of adopting a 1D radar signal rather than using

a 2D radar data representation is justiﬁed by the fact

that we deal with an ITS application where response

time is a crucial criterion.

3 PROPOSED METHOD

The received UWB radar signal is time-series data,

so, the adoption of the LSTM is recommended. Thus,

in this section, we ﬁrstly present the background of

the LSTM network, followed by a description of the

proposed approach.

3.1 LSTM Background

The LSTM network is a special architecture of the

artiﬁcial RNN developed in 1997 by Hochreiter and

Schmidhuber (Hochreiter and Schmidhuber, 1997). It

was introduced to avoid the issues that would occur

when modeling long-term dependencies with RNN

such as the vanishing or exploding gradient problem.

Therefore, the LSTM network is able to learn short-

time as well as long-term dependencies. It is espe-

cially known by its effectiveness to treat time-series

data (Ameur et al., 2020). In other words, the LSTM

network is able to model the temporal changes in a

series of data owing to its memory units and recur-

ICAART 2021 - 13th International Conference on Agents and Artiﬁcial Intelligence

420

𝑌

𝑡

tanh

Forget Gate

Input Gate

Output Gate

𝐶

𝑡−1

𝐶

𝑡

𝑂

𝑡

𝐼

𝑡

𝐹

𝑡

𝐹

𝑡

𝐻

𝑡

𝑋

𝑡

Figure 1: Architecture of LSTM unit.

rent architecture. The LSTM units are connected se-

quentially. Each LSTM cell includes its own memory

with three gates: the input, the output and the forget

gates. These gates are responsible for protecting and

controlling the ﬂow of information through the cell.

Otherwise, they decide which information has to be

forgotten or reminded. The LSTM cell is expressed

by Equation 1 and detailed in Figure 1:







= σ(W

t−1

] + b

)

= σ(W

t−1

] + b

)

= σ(W

t−1

] + b

)

= F

∗C

t−1

+ I

∗tanh(W

t−1

] + b

)

= O

∗tanh(C

)

= so f tmax(W

+ b

)







(1)

where:

• X ={X(1),X(2),..,X(N)} is a an input sequence,

where N is the length of the time series sequence.

• I

, F

, O

and C

are respectively the input, the

forget gate, the output gates and the memory cell

state.

• H

is the cell output and Y

is the ﬁnal output.

• tanh and σ are respectively the hyperbolic tangent

and the logistic sigmoid activation functions.

• W and b are respectively the input weights, asso-

ciating the LSTM cell to the inputs, and the bias

vectors.

3.2 Proposed UWB-based System for

Obstacle Detection

Despite the fact that UWB reﬂected signal incorpo-

rates rich information, the discrimination between the

object’s signature from noise is a fundamental chal-

lenge. The received UWB signal r(t) can be modeled

following Equation 2:

r(t) = s(t)+ n(t) (2)

where s(t) is the received echo and n(t) is the noise

of the transmission channel of the UWB radar.

In fact, the process of emitting and receiving an

impulse by the UWB radar is labeled as a radar scan,

and the received echo of the j

radar scan s

(t) is

presented by Equation 3:

(t) =

∑

k=1

x(t − t

) (3)

where: x(t) is the emitted pulse which is acquired as

reﬂected signals, a

is the amplitude. And, t and

are respectively the reﬂection duration of the pulse

after the transmission process and the delay of the k

received echo in the j

radar scan.

Thereby, based on the UWB property that indicates

that each obstacle has its own signature, noise also

should be different.

On the other hand, the target’s wave reﬂection am-

plitude is generally changing over time due to phys-

ical characteristics (material, shape, size, etc.), and

it gets attenuated when moving away from the radar

sensor. Otherwise, the received echo is a function of

time, and is also a function of the obstacle distance

from the sensor. This time-distance relation is ex-

pressed by Equation 4:

R =

τc

(4)

where τ and c are respectively the delay between

the transmission and reception of the UWB pulse, and

the speed of light.

Hence, based on the temporal changes presented

in the UWB signal and the fact that the noise parti-

tions should be different from a real object’s signa-

ture, we adopt the idea of employing the LSTM net-

work. This choice is made on account of the fact

that this type of RNNs is able to recognize and syn-

thesize the dynamics variations within the received

UWB echo. Thereby, in this work we put forward

the ﬁrst framework that exploits the LSTM network

with UWB signals for distinguishing obstacles from

noise in a vehicle environment perception context.

The proposed framework is explained in Algo-

rithm 14 and represented in Figure 2. Primarily, the

received signal is split into time series sequences [X

These sequences present the useful regions which

contain the real targets (illustrated by the orange win-

dows in Figure 2) and noise partitions (depicted by

green rectangles in Figure 2). For data variety, the

noise parts are randomly selected from the UWB sig-

nal. In fact, the window’s size (W sig) is empirically

set according to the objects’ signature length. After-

wards, features from the time-frequency domain are

extracted from the deﬁned regions. Thus, the Discrete

Davelet Transform (DWT) is utilized. We extract four

features from the approximation coefﬁcients (Ca) and

LSTM-based System for Multiple Obstacle Detection using Ultra-wide Band Radar

421

…

LSTM

LSTM layer

Fully

Connected

Layer

Softmax

Layer

Feature Extraction

UWB received Signals Decomposition

Detection :

Real Object / Noise

Sequences

Feature vector

Discrete Wavelet

Transform

Real object

Noise

Figure 2: Proposed LSTM-based framework using UWB signals.

Algorithm 1: LSTM-based algorithm for obstacle

detection using UWB signals.

Data: Received signal r = [r

], Signature

window length: Wsig ; LST M

options

;

Position: pos

Output: Class

1 // Split time series into sequences

2 for t ← pos to length(W sig) do

3 X

= r(pos,t + pos);

4 end for

5 // Extract features using the DWT

6 Fvector = []

7 [Ca,Cd] = DWT(X

)

8 F1 = mean(Ca);F2 = std(Ca); F3 =

min(Cd); F4 = rms(Cd)

9 Fvector = append(F

ector,F1, F2, F3,F4)

10 // Feed the LSTM-based network

11 Out put

LST M

= LST M(Fvector,LST M

options

)

12 Out putFCL =

Fullyconnectedlayer(Out put

LST M

)

13 Class = So f tmax(Out putFCL)

14 return (Class)

the detail ones (Cd) for each sequence. Subsequently,

the fattened 1D descriptor vector feeds the LSTM net-

work. Finally, the output of the LSTM is linked to the

fully connected layer of size 2 followed by a Softmax

layer and a classiﬁcation layer.

To conclude, the suggested detector analyzes the

temporal changes within the UWB signal via learn-

ing the extracted time-frequency features that highly

present the signal characteristics in order to differen-

tiate a real target from noise.

4 EXPERIMENTS

To highlight the efﬁciency of the proposed architec-

ture, we compare it to the implemented state-of-the art

techniques: CA-CFAR, HOS and the work in (Sadli

et al., 2018). Furthermore, we conduct our experi-

ments on a variety of urban driving scenarios. In this

section, we provide the experimental setup and the

obtained results. The experimental results are eval-

uated using the Precision (P), Recall (R) and F1-score

metrics.

4.1 Experimental Setup

The heterOgeneous Multimodal Dataset for Ad-

vanced EnvIronMent Perception (OLIMP)

is ex-

ploited in this work (Mimouna et al., 2020) . It

contains several urban driving situations including

four main categories: pedestrian, cyclist, vehicle and

tram. The dataset includes over 47 thousands of UWB

signals collected under various conditions: different

weather conditions (as fog, snow), near and far tar-

gets, etc. The developed radar by the UMAIN Inc

company is the sensor that has been used to record

the OLIMP database (Anu, 2020) . It is entitled HST-

D3. The radar distance range is six meters and its fre-

quency range is [3GHZ, 4GHZ] with a bandwidth of

0.45-1Ghz. The HST-D3 radar is composed of a mod-

ule radar called HST-S1 Pi and a Raspberry Pi 3. In

addition, The directional antennas are utilized as they

guarantee a better target echo-to-clutter and noise ra-

tio. The UWB radar signal comprises 660 samples

per frame. The radar is illustrated in Figure 3.

https://sites.google.com/view/ihsen-alouani/datasets

ICAART 2021 - 13th International Conference on Agents and Artiﬁcial Intelligence

422

Table 1: Execution time results.

HOS CA-CFAR (Sadli et al., 2018) Proposed method

Execution time (s) 1.34 1.41 1.52 2.01

For our experiments, we use 2/3 of the data for

the training step and 1/3 for the test process. It shall

be mentioned that for training process, we exploit the

unidirectional network with 100 LSTM hidden units.

The number of epochs and mini batch size during ex-

periments are set to 100 and 64 instances respectively.

Moreover, we utilize the Adam optimizer (Bock et al.,

2018). The initial learning rate is set to 0.001. The

optimal set of parameters is selected based on pre-

liminary tests. For the DWT, the Daubechies wavelet

(db2) is employed to extract the features from the time

series sequences.

As regards to the comparative study, for the HOS

we take advantage of the 4

order cumulant that re-

lies on the Tuganit4 algorithm. Concerning the CFAR

technique, the CA-CFAR detector is considered with

an automatic threshold.

All the experiments are performed using Matlab

2020a on a PC with an Intel (R) core (TM) i7-8565U,

CPU @ 1.8 GHz, 16 GB of RAM.

The HST-D3 UWB Radar

Transmitted pulses

Received Echo

HST-S1 Pi radar module

Raspberry Pi 3

UWB Directional Antennas

Figure 3: Used HST-D3 UWB radar hardware speciﬁca-

tions.

4.2 Results and Discussion

The experimental results concerning the comparative

study is illustrated in Figure 4. In fact, it can be seen

from the ﬁgure that the obtained results show that our

deep learning-based method achieves the highest per-

formance. Our proposed method outperforms signif-

icantly the considered traditional detectors, particu-

larly in terms of recall and precision.

The performance of the CA-CFAR, the HOS and

their combination in the work of (Sadli et al., 2018)

depend essentially on the deﬁnition of the thresh-

old parameter. A higher threshold generates more

false negatives, but, the opposite case produces ad-

ditional false positives. Furthermore, the object de-

tection rate using the aforementioned techniques rely

Figure 4: Experimental results using HOS, CFAR, work in

(Sadli et al., 2018) and our method.

on the object’s amplitude. In low magnitude cases,

which means that the object is moving away from the

radar, the target cannot be detected and it is consid-

ered as noise.

Compared with these techniques, the proposed

system remains robust even with low amplitude cases

as it is not based basically just on the object’s ampli-

tude. In fact, the LSTM-based method can distinguish

noise from real targets thanks to the relevant features

that are extracted from the time-frequency domain,

and by learning the temporal relationship between the

data sequences. The DWT reveals the structure of the

signal, thus allowing the localization of its peculiar

characteristics in time and frequency domains. For

this reason, the time-frequency features lead to a high

performance as they can represent well the signal’s

characteristics.

It shall be mentioned that, some missclassiﬁcations

are still present due the challenges related to the in-

terclass similarity of the obstacle’s signature and the

noise signal when the object’s signature has a low

amplitude. Accordingly, further research can be ex-

panded to extract deep features to tackle this dilemma.

In terms of complexity, Table 1 depicts a compar-

ison of the execution time of our method with the

state-of-the-art techniques. As it can be seen from

the results, our architecture has the highest execution

time, though, advanced hardware resources are to be

deployed to acquire faster detection time.

5 CONCLUSION

In this paper, we put forward a novel framework

for obstacle detection using UWB signals with deep

learning techniques. The developed detector focuses

principally on distinguishing the real target from

LSTM-based System for Multiple Obstacle Detection using Ultra-wide Band Radar

423

noise within the received UWB signal. For this pur-

pose, features are extracted using the DWT. These

descriptors are used as inputs of the LSTM network

which is exploited as a classiﬁer. According to the

extensive experiments carried out on a large dataset.

Promising results are found, which outperform the

performances of the related techniques. For future

work, the employment of CNN to extract features

from the UWB signals can be investigated.

REFERENCES

(2020). Umain. http://www.umain.co.kr/en/uwb-radar-

sensor-module/.

Ahmed, S. and Cho, S. H. (2020). Hand gesture recognition

using an ir-uwb radar with an inception module-based

classiﬁer. Sensors, 20(2):564.

Ameur, S., Khalifa, A. B., and Bouhlel, M. S. (2020). A

novel hybrid bidirectional unidirectional lstm network

for dynamic hand gesture recognition with leap mo-

tion. Entertainment Computing, 35:100373.

Blake, S. (1988). Os-cfar theory for multiple targets and

nonuniform clutter. IEEE transactions on aerospace

and electronic systems, 24(6):785–790.

Bock, S., Goppold, J., and Weiß, M. (2018). An improve-

ment of the convergence proof of the adam-optimizer.

arXiv preprint arXiv:1804.10587.

Briso, C., Calvo, C., and Xu, Y. (2019). Uwb propagation

measurements and modelling in large indoor environ-

ments. IEEE Access, 7:41913–41920.

Cai, L., Ma, X., Yan, S., Hao, C., and Wang, R. (2010).

Some analysis of fuzzy cago/so cfar detector in non-

gaussian background. In 2010 2nd International

Workshop on Intelligent Systems and Applications,

pages 1–4. IEEE.

Chen, S., Fan, C., Huang, X., and Cao, C. (2019). Low

prf low frequency radar sensor for fall detection by

using deep learning. In 2019 IEEE 4th International

Conference on Signal and Image Processing (ICSIP),

pages 400–404. IEEE.

Choi, J. W., Yim, D. H., and Cho, S. H. (2017). People

counting based on an ir-uwb radar sensor. IEEE Sen-

sors Journal, 17(17):5717–5727.

Conte, E., Longo, M., and Lops, M. (1988). Performance

analysis of ca-cfar in the presence of compound gaus-

sian clutter. Electronics Letters, 24(13):782–783.

Hakobyan, G. and Yang, B. (2019). High-performance au-

tomotive radar: A review of signal processing algo-

rithms and modulation schemes. IEEE Signal Pro-

cessing Magazine, 36(5):32–44.

Hochreiter, S. and Schmidhuber, J. (1997). Long short-term

memory. Neural computation, 9(8):1735–1780.

Jegham, I., Khalifa, A. B., Alouani, I., and Mahjoub, M. A.

(2018). Safe driving: driver action recognition using

surf keypoints. In 2018 30th International Conference

on Microelectronics (ICM), pages 60–63. IEEE.

Jiang, C., Shen, J., Chen, S., Chen, Y., Liu, D., and

Bo, Y. (2020). Uwb nlos/los classiﬁcation using

deep learning method. IEEE Communications Letters,

24(10):2226–2230.

Khalifa, A. B., Alouani, I., Mahjoub, M. A., and Rivenq,

A. (2020). A novel multi-view pedestrian detec-

tion database for collaborative intelligent transporta-

tion systems. Future Generation Computer Systems,

113:506–527.

Kim, D.-H. (2020). Lane detection method with impulse

radio ultra-wideband radar and metal lane reﬂectors.

Sensors, 20(1):324.

Ko, S.-K. and Lee, B.-T. (2016). Object classiﬁcation of

uwb responses using s t-cnn. In 2016 International

Conference on Information and Communication Tech-

nology Convergence (ICTC), pages 794–796. IEEE.

Koci

c, J., Jovi

c, N., and Drndarevi

c, V. (2018). Sensors

and sensor fusion in autonomous vehicles. In 2018

26th Telecommunications Forum (TELFOR), pages

420–425.

Kumar, D., Sarkar, A., Kerketta, S. R., and Ghosh, D.

(2019). Human activity classiﬁcation based on breath-

ing patterns using ir-uwb radar. In 2019 IEEE 16th

India Council International Conference (INDICON),

pages 1–4.

Lee, D., Shaker, G., and Melek, W. (2020). Imaging of hu-

man walking behind the obstacle utilizing pulsed radar

technique in the c-band for military surveillance appli-

cations. JOURNAL OF ELECTRICAL ENGINEER-

ING & TECHNOLOGY, 15(3):1431–1439.

Lim, S., Jung, J., Kim, S.-C., and Lee, S. (2020). Deep neu-

ral network-based in-vehicle people localization using

ultra-wideband radar. IEEE Access.

Maitre, J., Bouchard, K., Bertuglia, C., and Gaboury, S.

(2021). Recognizing activities of daily living from

uwb radars and deep learning. Expert Systems with

Applications, 164:113994.

Mendel, J. M. (1991). Tutorial on higher-order statistics

(spectra) in signal processing and system theory: the-

oretical results and some applications. Proceedings of

the IEEE, 79(3):278–305.

Mimouna, A., Alouani, I., Ben Khalifa, A., El Hillali,

Y., Taleb-Ahmed, A., Menhaj, A., Ouahabi, A., and

Ben Amara, N. E. (2020). Olimp: A heterogeneous

multimodal dataset for advanced environment percep-

tion. Electronics, 9(4):560.

Nguyen, V.-H. and Pyun, J.-Y. (2015). Location detection

and tracking of moving targets by a 2d ir-uwb radar

system. Sensors, 15(3):6740–6762.

Piriyajitakonkij, M., Warin, P., Lakhan, P., Leelaarporn, P.,

Kumchaiseemak, N., Suwajanakorn, S., Pianpanit, T.,

Niparnan, N., Mukhopadhyay, S. C., and Wilaiprasit-

porn, T. (2020). Sleepposenet: Multi-view learning

for sleep postural transition recognition using uwb.

IEEE Journal of Biomedical and Health Informatics.

Poulose, A. and Han, D. S. (2020). Uwb indoor localization

using deep learning lstm networks. Applied Sciences,

10(18):6290.

Quan, X., Choi, J. W., and Cho, S. H. (2020). A new thresh-

ICAART 2021 - 13th International Conference on Agents and Artiﬁcial Intelligence

424

olding method for ir-uwb radar-based detection appli-

cations. Sensors, 20(8):2314.

Rohling, H. (1983). Radar cfar thresholding in clutter

and multiple target situations. IEEE transactions on

aerospace and electronic systems, (4):608–621.

Sadli, R., Tatkeu, C., Hamidoun, K., El Hillali, Y., and

Rivenq, A. (2018). Uwb radar recognition system

based on hos and svms. IET Radar, Sonar & Navi-

gation, 12(10):1137–1145.

Sadreazami, H., Bolic, M., and Rajan, S. (2018). On the

use of ultra wideband radar and stacked lstm-rnn for

at home fall detection. In 2018 IEEE Life Sciences

Conference (LSC), pages 255–258. IEEE.

Saito, A., Harada, H., and Nishikata, A. (2003). Devel-

opment of band pass ﬁlter for ultra wideband (uwb)

communication systems. In IEEE Conference on Ul-

tra Wideband Systems and Technologies, 2003, pages

76–80. IEEE.

Sakkila, L., Deloof, P., Elhillali, Y., Rivenq, A., and Niar, S.

(2006). A real time signal processing for an anticolli-

sion road radar system. In IEEE Vehicular Technology

Conference, pages 1–5.

Sakkila, L., Tatkeu, C., Boukour, F., El Hillali, Y., Rivenq,

A., and Rouvean, J. (2008). Uwb radar system for

road anti-collision application. In 2008 3rd Inter-

national Conference on Information and Communi-

cation Technologies: From Theory to Applications,

pages 1–6.

Skrickij, V.,

Sabanovi

c, E., and

Zuraulis, V. (2020). Au-

tonomous road vehicles: recent issues and expecta-

tions. IET Intelligent Transport Systems, 14(6):471–

479.

Tarchoun, B., Ben Khalifa, A., Dhifallah, S., Jegham, I.,

and Mahjoub, M. A. (2020a). Hand-crafted features

vs deep learning for pedestrian detection in moving

camera. Traitement du Signal, 37(2):209–216.

Tarchoun, B., Jegham, I., Khalifa, A. B., Alouani, I., and

Mahjoub, M. A. (2020b). Deep cnn-based pedestrian

detection for intelligent infrastructure. In 2020 5th

International Conference on Advanced Technologies

for Signal and Image Processing (ATSIP), pages 1–6.

IEEE.

Tong, H., Xin, N., Su, X., Chen, T., and Wu, J. (2020).

A robust pdr/uwb integrated indoor localization ap-

proach for pedestrians in harsh environments. Sen-

sors, 20(1):193.

Zetik, R., Sachs, J., and Thoma, R. S. (2007). Uwb short-

range radar sensing - the architecture of a baseband,

pseudo-noise uwb radar sensor. IEEE Instrumentation

Measurement Magazine, 10(2):39–45.

LSTM-based System for Multiple Obstacle Detection using Ultra-wide Band Radar

425