IEEE TRANSACTIONS ON RADAR SYSTEMS, VOL.

1, 2023 75

Improved Target Detection Through DNN-Based

Multi-Channel Interference Mitigation in
Automotive Radar
Shengyi Chen , Marvin Klemp, Jalal Taghia, Uwe Kühnau, Nils Pohl , Senior Member, IEEE,
and Rainer Martin , Fellow, IEEE

Abstract— Deep learning methods have triggered significant spectrum and the probability of mutual interference increases
progress in automotive radar-based object detection and classifi- accordingly. Since unexpected interference can either reduce
cation. However, with an increasing number of radar sensors the signal-to-noise ratio (SNR) or create ghost targets [3], the
on the road, mutual interference is unavoidable since these
sensors share the same frequency spectrum. Mutual interference accuracy of target detection can be impeded by interference if
affects the robustness of radar processing schemes and thereby no countermeasures are taken.
the object detection accuracy. Unlike many recent works which In order to achieve an efficient sharing of the frequency
focus on interference mitigation for a single receive channel, this bands dedicated to automotive radar, cooperative interfer-
paper proposes a multi-channel mitigation approach and seeks ence mitigation strategies [4], [5] have been proposed. How-
to analyze the effect of mutual interference in a multiple-input
multiple-output (MIMO) radar. To this end, we first formulate ever, additional coordination units are required to imple-
a general signal model for multi-channel interference scenarios. ment cooperation algorithms which will add extra costs and
Then, a novel signal separation neural network is proposed for operating time to the existing radar system. Furthermore,
multi-channel interference mitigation which eliminates the efforts several approaches based on waveform design have been
of interference detection. We assess the impact of interference in introduced for tackling the interference issue [6], [7]. In [6],
terms of the reconstruction error, angle estimation error, and
the target detection accuracy in both real-world and simulated a waveform with random phase has been designed, where the
interference scenarios. It is demonstrated that the proposed consecutively transmitted waveforms are orthogonal to each
neural network can provide superior signal recovery, massively other. Thus, interference signals can be reduced by matched
reduces the false-positive rate, and significantly improves the filtering. A phase-coded frequency-modulated continuous-
accuracy of object detection even in the presence of severe wave (FMCW) and a phase-coded linear-frequency-modulated
interference.
(LFM) continuous waveform are proposed in [7]. The main
Index Terms— Automotive radar, interference mitigation, sig- challenge for these cooperative interference mitigation strate-
nal separation, target detection, deep learning.
gies and waveform design approaches resides in its coexistence
with the chirp sequence modulation radar, a widely used
I. I NTRODUCTION variant of the FMCW radar in automotive radar systems,
as well as the older generations of automotive radars such
R ADAR based environment analysis is receiving increas-
ing attention in advanced driver assistance systems
(ADAS) [1], [2]. In contrast to camera and lidar sensors,
as the continuous-wave (CW) radar.
Besides the cooperative interference mitigation strategies
radar sensors are robust in adverse weather conditions such as and the waveform design approaches exemplified above, var-
snow, heavy rain, as the wavelength of millimeter-wave radar ious signal reconstruction algorithms have been developed
is longer than the particle size of hydrometeors. In order to for solving the interference problem. These signal recovery
achieve a 360-degree view, modern vehicles use multiple radar algorithms are mainly designed to recover four different signal
sensors for ADAS functions. However, due to frequency spec- types: the discrete-time beat signal [3], [8], [9], [10], [11],
trum regulation, most radar sensors share the same frequency [12], [13], [14], [15], [16], the range profile (obtained after
applying the fast Fourier transform (FFT) to fast-time samples
Manuscript received 12 December 2022; revised 5 March 2023 and 17 May in the discrete beat signal) [17], [18], [19], the frequency spec-
2023; accepted 19 May 2023. Date of publication 22 May 2023; date of
current version 7 June 2023. (Corresponding author: Shengyi Chen.) trum obtained after applying the short-time Fourier transform
Shengyi Chen and Rainer Martin are with the Institute of Communica- (STFT) to the discrete beat signal [20], and the range-Doppler
tion Acoustics, Ruhr-Universität Bochum, 44780 Bochum, Germany (e-mail: (RD) spectrum (obtained after applying the second FFT to the
[email protected]; [email protected]).
Marvin Klemp is with the Institute of Measurement and Control Systems, range profile in the chirp direction) [17], [21], [22], [23], [24].
Karlsruhe Institute of Technology, 76131 Karlsruhe, Germany (e-mail: mar- For eliminating the interference-contaminated discrete beat
[email protected]).
Jalal Taghia is with HELLA Aglaia Mobile Vision GmbH, 12109 Berlin,
signal segments, the zeroing method is employed in [3] where
Germany (e-mail: [email protected]). the signal segments disturbed by interference are detected and
Uwe Kühnau is with HELLA GmbH & Co. KGaA, 59555 Lippstadt, simply set to zero. However, when the number of disturbed
Germany (e-mail: [email protected]).
Nils Pohl is with the Institute of Integrated Systems, Ruhr-Universität
samples becomes large, the zeroing method results in the loss
Bochum, 44801 Bochum, Germany (e-mail: [email protected]). of target information in terms of both phase and amplitude.
Digital Object Identifier 10.1109/TRS.2023.3279013 Different to [3], the autoregressive (AR) model-based method
2832-7357 © 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: IMEC. Downloaded on June 23,2023 at 12:15:34 UTC from IEEE Xplore. Restrictions apply.
76 IEEE TRANSACTIONS ON RADAR SYSTEMS, VOL. 1, 2023

is used in [12] for the discrete beat signal interpolation in with the camera- or lidar-based solutions. In [1], a complete
the sample positions which are disturbed by interference. processing pipeline was developed for a semantic segmen-
An autoencoder was developed in [16] for the reconstruc- tation of the radar point cloud obtained from measurements
tion of disturbed discrete beat signals, where the disturbed of multiple radar sensors. The performance of the object
samples are detected by an edge detector as in [9] and the classification is improved in [28] by extracting the features
proposed autoencoder reconstructs interference-free samples. of point cloud clusters from 98 handcrafted feature candi-
In general, most of these time-domain signal reconstruction dates. Due to the application of the constant false alarm rate
algorithms require a reliable interference detection algorithm (CFAR) algorithm for the extraction of target points, the point
that can precisely determine the position of disturbed samples. cloud returned from the commercial radar sensors is usually
In [25], a method was presented in which the interference is sparse which might cause the loss of target information.
detected and suppressed by an iterative adaptive thresholding Most recently, the RadarResNet with one-stage anchor-based
procedure, however, it is difficult to locate interferences which detector and a fully convolutional network are proposed in [2]
have a similar amplitude as the interference-free signal [13]. and [30], respectively, for automotive radar object detection
The signal strength of disturbed samples may vary largely due using the low-level radar signals, namely the Range-Azimuth-
to the multi-path propagation and the distance between the Doppler (RAD) data cube. It should be noted that the impact
interferer and the victim radar. of mutual interference is not taken into account in the object
Similar to the approaches for time-domain signal recovery, detection accuracy achieved in these contributions.
the disturbed signal segments in the range profile or in the The problem of mutual interference involving a single
STFT frequency spectrum can be first detected and then receive antenna has been extensively studied in the literature;
interpolated. In [20], the interference-free signal segments yet, the propagation of interference across multiple receive
are reconstructed in the STFT frequency spectrum by lin- antennas and its impact on the target detection need to be
ear predictive coding (LPC). In [19] a generative adversar- further explored. In this paper, we investigate the mitigation
ial network (GAN) is proposed to recover the interference- of mutual interference across multiple receive antennas and
free signal segments in the complex range profile. However, analyze how the interference affects the accuracy of target
in some cases the signal segments disturbed by interference detection. The main contributions are summarized as follows:
may spread across the entire STFT frequency spectrum and the • We formulate a general signal model for multi-channel
range profile, therefore the available interference-free signal interference scenarios.
segments are not sufficient for these algorithms to provide • Using the above signal model, we create a data set by
1
adequate recovery. simulating different types of interference and combining
Various deep learning techniques have also been proposed it with an existing data set of interference-free discrete
to reconstruct the disturbed RD spectrum [17], [21], [22], beat signals acquired by a MIMO radar.
[23], [24]. In [17], a convolutional neural network (CNN) • We propose a signal separation neural network for sep-
is employed for noise suppression in the range profile and arating the interference-free discrete beat signal and the
the RD spectrum for a single receive channel. However, interference. The proposed signal separation neural net-
the target peak values may be distorted in the reconstructed work can directly reconstruct the interference-free dis-
RD spectrum [17]. The resource-efficient and complex-valued crete beat signal of multiple receive channels.
variants of CNN have been proposed in [22] and [24], respec- • The proposed signal separation neural network is evalu-
tively. A convolutional autoencoder is employed in [21] for ated in terms of signal reconstruction, angle estimation
the reconstruction of the RD spectrum, where the amplitude error, and improvement of the target detection accuracy
and phase components of the complex-valued RD spectrum in comparison to state-of-the-art algorithms.
are used as input. However, this approach cannot properly
The organization of this paper is as follows. In Section II,
reconstruct the phase information, since the phase component
the FMCW MIMO radar system model and the signal process-
of the target can resemble the phase of the noise (both vary
ing chain are introduced. Section III describes the proposed
in the same range) and is therefore more difficult for the
signal separation neural network for interference mitigation.
neural network to acquire. The deep learning approaches have
In Section IV, the performance of the proposed method is eval-
shown their strength in noise suppression in these studies when
uated and the influence of interference on the object detection
the amplitude of the target peaks is distinctly greater than
accuracy is investigated. Section V concludes this paper.
the background noise level in the disturbed RD spectrum.
However, when the weak target peaks are obscured by the
background noise, it may be difficult to recover them directly II. FMCW MIMO R ADAR S IGNAL M ODEL AND S IGNAL
from the disturbed RD spectrum [13]. P ROCESSING C HAIN
In order to increase the resolution of the angle estima- A. Signal Model and Interference
tion of commercial automotive radar with a low hardware
cost and small package size, the multiple-input multiple- Most commercial automotive radar systems currently use
output (MIMO) [26], [27] radar technology has been receiving chirp sequence modulation [31]. Considering the time-division
considerable attention recently. Environment perception using 1 For the sake of reproducibility of experiments, a data set containing
MIMO radar sensors has shown a comparable accuracy [1], the simulated interference of different types of radar sensors is provided in
[2], [28], [29] in the object detection and classification tasks gitlab.ruhr-uni-bochum.de/chenshfm/interference-data-set.

Authorized licensed use limited to: IMEC. Downloaded on June 23,2023 at 12:15:34 UTC from IEEE Xplore. Restrictions apply.
CHEN et al.: IMPROVED TARGET DETECTION THROUGH DNN-BASED MULTI-CHANNEL INTERFERENCE MITIGATION 77

The q-th chirp of one target’s echo signal at the n-th receiver
emitted from the m-th transmitter is delayed by τm,n :
rq,m,n (t) =
aq,m Aq,m,n s(t − (⌊q/M⌋TPRI + mT ) −τm,n ) + v(t), (4)
| {z }
tq,m

where
2R 2vt mdt sin(θ ) ndr sin(θ )
τm,n = + + + , (5)
c
|{z} c c c
τ0

dt and dr are the inter-element spacing of the transmit and

receive antennas, respectively, Aq,m,n is the received ampli-
tude, θ the direction of arrival (DoA) of waves,3 q ∈ {0,
MQ − 1}, m ∈ {0, M − 1}, n ∈ {0, N − 1}, and v denotes
complex-valued white noise. Here, R and v denote, respec-
tively, the distance and relative radial velocity between the
Fig. 1. An example of TDM MIMO with two transmit antennas (TX) and radar sensor and the object, and c represents the speed of
four receive antennas (RX).
light.
The beat signal in the baseband can be obtained after
stretch processing, namely mixing rq,m,n conjugately with the
multiplexing (TDM) MIMO scheme2 with M transmitters and transmitted signal:
N receivers [26], [27], each transmitter sends a sequence of Q
chirps, where the active transmit antenna, enumerated by m, ŷ q,m,n (t) = rq,m,n
∗
(t)s(t − tq,m ), (6)
is changed after each single chirp. For an antenna configuration
where tq,m has been defined in (4). Considering the phase term
with M transmit antennas and N receive antennas, M N virtual
of ŷ q,m,n (t), we have
receive channels can be synthesized (i.e., taking the receiver
φq,m,n (t) = φ(t − tq,m ) − φ t − tq,m − τm,n


antenna array and placing it at the position of each transmitter
antenna) [32, Chapter 3.1.2]. Fig. 1 shows an example of = 2π f c τm,n + κ(t−tq,m )τm,n − 0.5κ(τm,n )2 . (7)


the TDM MIMO technique. The waveform emitted by M
transmitters is given by After sampling with f s = 1/Ts , and collecting P samples
per chirp at time t = tq,m + pTs (p ∈ {0, P − 1}), the discrete
MX
Q−1 M−1
X beat signal in the case of a single target can be described
Tx (t) = aq,m s(t − ⌊q/M⌋TPRI − mT ), (1) as [27]
q=0 m=0 
f c +κtq,m

j2π κτ0 + 2v pTs
ŷ m,n ( p, q) ≈ aq,m Aq,m,n e c

sin(θ )
where TPRI is the pulse repetition interval (see Fig. 1), ⌊·⌋ · e j2π fc
2v
c tq,m e j2π(mdt +ndr ) λ e j2π fc τ0 + v, (8)
the floor function, T the chirp duration (M T ≤ TPRI ), q
enumerates the transmitted chirps across all transmit antennas, where λ = c/ f c is the wavelength. Note that the second
and order terms are neglected [33]. The term ( f c + κtq,m )2v/c
( corresponds to the range migration and Doppler-frequency
1, if q mod M = m shift effects which are typically neglected as well [27]. The
aq,m = (2) single-target MIMO signal model in (8) can now be further
0, otherwise.
extended to L targets:
The individual transmit chirp signal with a normalized ampli- L
(l)
e j2π κτl pTs e j2π fdl tq,m
X
tude is described as ŷ m,n ( p, q) ≈ aq,m Aq,m,n
l=1
sin(θl )
 
j2π(mdt +ndr )
s(t) = e jφ(t) rect
t
, (3) ·e λ e j2π fc τl + v, (9)
T
where τl = 2Rl /c denotes the round-trip propagation delay of
where φ(t) = 2π f c t +πκt is the phase of the local-oscillator,
2 the l-th target with Rl and vl denoting the distance and relative
f c the carrier frequency, κ = B/T the chirp rate with B radial velocity of this target, f dl = f c 2vc l , and θl represents its
denoting the sweep bandwidth, and rect(·) the square pulse DoA.
which is one in the interval [0, 1) and zero otherwise. With the assumption that the interfering signal has the car-
rier frequency f c I , the transmit waveform of an interferer radar
2 We only consider the TDM MIMO scheme in the signal model as well as 3 With co-located antennas [32] and the far-field approximation, i.e., for a
in our interference simulation as this technology is applied in the radar sensor given target, the azimuth angle of transmit signals is the same as the azimuth
being used for data acquisition in this work. angle of receive signals.

Authorized licensed use limited to: IMEC. Downloaded on June 23,2023 at 12:15:34 UTC from IEEE Xplore. Restrictions apply.
78 IEEE TRANSACTIONS ON RADAR SYSTEMS, VOL. 1, 2023

LPF-frequency limits are considered. Since the interference

signal arrives at the n-th receive antenna with a time delay
ndr sin(θ )/c (compared to its arrival time at the first receive
antenna), the additive in-band interference signal of the q-th
chirp at the n-th receive antenna is
X (q)
y̌ q,n (t) ≈ Ãi,n e j2π( fc I − fc )t e− j2π ndr sin(θ )/λ I
i∈ Z̃ q
 2 
(q) (q) (q)
  
j2π 0.5(κ̃−κ)t 2 −κ̃τ I t+0.5κ̃ τ I − fc I τ I t − qT
·e i i i
rect ,
T
(11)

where Z̃ q denotes the set of interfering chirps that disturb the

(q)
f c − f LPF − f c I +κ̃τ I
q-th chirp at the n-th receiving antenna, κ̃−κ
i
⩽t ⩽
(q)
f c + f LPF − f c I +κ̃τ I
κ̃−κ
(extending the derivations in [8] for the case
i

f c I ̸= f c ), and λ I = c/ f c I is the wavelength. Z̃ q is determined

by several factors, namely the modulation parameters of the
interferer and victim radars, and the collision time between the
interferer and victim radars. Z̃ q is an empty set if no interfering
Fig. 2. Examples of different interference scenarios. chirps are present. When TPRI of the transmitted waveform of
the victimradars is assumed to be equal to M T in (11), then
rect t−qT can be used to denote the time span of the q-th
is Tint (t) = z=0 sint (t) = Ãe j φ̃(t) rect T̃t
P Z −1


T
sint t − z T̃ , where
chirp. It should be noted that the received additive interference
with φ̃(t) = 2π f c I t + 0.5κ̃t 2 , Ã the amplitude, Z the


varies only at different receive antennas due to the phase shift
number of transmitted chirps, κ̃ = B̃/T̃ , B̃ and T̃ denote
and is not related to the transmitters. After sampling with a
the chirp slope, sweep bandwidth, and chirp duration of the
period of Ts at each receive channel, we now use y̌ m,n ( p, q̃)
interfering radar, respectively. It should be noted that this paper
to represent the additive in-band interference samples at the
focuses on incoherent interference, hence κ̃ is not considered
virtual receive channel (m N + n). The discrete beat signal
to be identical to κ. Assuming that the interferer has co-located
ym,n ( p, q̃) can be therefore summarized as
antennas, and with the far-field approximation, it makes no
difference to the receivers from which transmit antenna the (
ŷ m,n ( p, q̃) + v ( p, q̃) ∈ Hm,n
interference chirp originates. Therefore, for the description of ym,n ( p, q̃) =
the transmitted waveform of the interferer, chirps are simply ŷ m,n ( p, q̃) + y̌ m,n ( p, q̃) + v ( p, q̃) ∈
/ Hm,n ,
counted from 0 to Z − 1. (12)
Due to different chirp rates and bandwidths, the interference
scenario can vary case by case as shown in Fig. 2. By using where ( p, q̃) denotes the sample index ( p indicates the sample
(q) position inside the chirp and q̃ the chirp index within a chirp
the transmitted chirp sequence as reference, τ Ii is used to
denote the delay time between the start time of the i-th chirp sequence transmitted by the same TX) and Hm,n is the set
from interferer and the q-th transmitted chirp. Note that the of sample indices without interference at the virtual receive
(q)
delay time τ Ii is a negative number if the start time of the channel (m N + n).
i-th interfering chirp is earlier than that of the q-th transmitted
chirp. As shown in the scenario (c) in Fig. 2, during one chirp
B. Signal Processing Chain
time T multiple positions in the chirp can be disturbed. In this
case, τ I(0)
0
and τ I(0)
1
represent the delay time of the occurrence We now explain the signal processing chain in the auto-
of first and second interfering chirp w.r.t. the first transmitted motive radar system for target detection with the RAD data
chirp, respectively. The phase of the interference signal after cube in the presence of interference. To obtain the range,
stretch processing is: velocity, and DoA information of targets, few signal process-
ing steps need to be applied to the received discrete beat
(q) (q)
 
φ̌ Ii (t) = φ̃ t − τ Ii − φ(t) signal. Fig. 3 shows the signal processing chain of a FMCW

(q) TDM MIMO radar system with two transmit antennas and
= 2π ( f c I − f c )t + 0.5(κ̃ − κ)t 2 − κ̃τ Ii t four receive antennas. Since in this paper the interference

(q)
2
(q)
 mitigation method is applied to the disturbed multi-channel
+ 0.5κ̃ τ Ii − f c I τ Ii (10) discrete beat signal, the RAD data cube is obtained based
on the recovered discrete beat signal. Finally, the RAD data
(q)
  
where − f LPF ⩽ f c I + κ̃ t − τ Ii − ( f c + κt) ⩽ f LPF , f LPF cube is utilized for the object detection and classification. For
is the cut-off frequency of the low-pass filter (LPF), indicating a better comparison, the classical point cloud-based object
that only the interference signal segments falling within the detection is also shown in Fig. 3.

Authorized licensed use limited to: IMEC. Downloaded on June 23,2023 at 12:15:34 UTC from IEEE Xplore. Restrictions apply.
CHEN et al.: IMPROVED TARGET DETECTION THROUGH DNN-BASED MULTI-CHANNEL INTERFERENCE MITIGATION 79

Fig. 3. Signal processing pipeline in FMCW TDM MIMO radar system in the presence of interference.

Fig. 4. Signal separation neural network (SeparationNet) for interference mitigation. For clarity, only the real part of the input data is shown.

III. S IGNAL S EPARATION N EURAL N ETWORK FOR the target components. By including signals from multiple
I NTERFERENCE M ITIGATION receive channels in the training process, the neural network can
In this section we introduce the proposed signal separation acquire this phase relationship across multiple receive channels
neural network for separating target’s echo signals from the and hence utilize this additional information for improving
interference. signal reconstruction. Furthermore, the convolutional neural
network is inherently proficient in learning the signal features
directly from multiple channels (see Fig. 5). The kernels
A. Neural Network Architecture in various filters extract the signal features in each input
The architecture of the proposed signal separation neu- channel (yellow for the real signal part and green for the
ral network (SeparationNet) shown in Fig. 4 is based on imaginary signal part) and the filters summarize outputs of
convolutional neural network layers [34]. In contrast to kernels into feature maps that can later be employed for
state-of-the-art neural networks using the disturbed signal signal reconstruction. Note that the number of filters per layer
of a single receive channel for interference mitigation, Sep- corresponds to the number of its output channels. Thus, the
arationNet is designed to reconstruct multi-channel receive more filters used in a convolutional layer, the more signal
signals, as depicted in Fig. 4 and 5. The echo signals from features can be extracted (i.e., more channels in feature maps).
targets in different receive channels have a relationship in To this end, the complex disturbed discrete beat signal in
terms of DoA-dependent phase shifts, which help to determine (12) is rearranged into a tensor of the form Ys ∈ RP×Q×MN×2
Authorized licensed use limited to: IMEC. Downloaded on June 23,2023 at 12:15:34 UTC from IEEE Xplore. Restrictions apply.
80 IEEE TRANSACTIONS ON RADAR SYSTEMS, VOL. 1, 2023

Fig. 6. Radar sensor employed for the measurement of multi-channel

interference-free discrete beat signals.

where ϵ denotes the total number of samples in the input data

cube, eYs , b
Ys , e q I ∈ RP×Q×MN×2 are the tensor form of
Y I , and Y
the recovered discrete beat signal, the interference-free discrete
Fig. 5. An example of a Conv3D employed in SeparationNet. The number beat signal, the reconstructed interference signal, and the true
of filters per layer corresponds to the number of its output channels. The real
and imaginary signal parts are convolved with the distinct kernels in each
interference signal, respectively.
filter (the real part with the kernel displayed in yellow and the imaginary part Alternatively, we also consider the root mean-squared error
with the kernel displayed in green) and the filter summerizes the output of (RMSE) loss which is defined as
each kernel into feature maps. The kernels are swept over all signal samples s
in 3 directions with a given stride during convolution.  2 
1X e
2 
LR M S E = Ys − Ys + Y I − Y I
b e q . (14)
ϵ
consisting of the real and imaginary parts as the input for the
neural network. The first four convolutional layers (denoted as C. Data Set and Training Details
Conv3D) extract signal features from the mixed signals. The
Since the effort of collecting a large number of interference-
signal features of interference-free samples and interference
contaminated measurements with various interference sources
samples are supposed to be the output of the first half and
is prohibitive, a viable approach is to simulate the interfer-
the second half of the filters in the fourth convolutional layer,
ence [17]. Different to [17], where the discrete beat signals are
respectively. Then, each half of the extracted signal features
simulated only for one receive channel, we simulate several
are utilized to reconstruct the corresponding interference-free
types of multi-channel interference based on the real-world
signal segments and the interference segments. The recovered
measured discrete beat signals of a FMCW MIMO radar
interference-free signal can be used for further processing
(TI AWR1843BOOST5 ) [2] according to its signal param-
steps as shown in Figure 3. In the meantime, based on the
eters as shown in Table I. Fig. 6 shows the radar sensor
separated interference signal, the operating frequency band of
employed for measurement of the real-world discrete beat
the interferer can be estimated, the victim radar can therefore
signals. It integrates the phase-locked loop (PLL), transmitter,
change its operating frequency to other frequency bands which
receiver, and analog-to-digital converter (ADC) in a single
are free of interference [35]. Thus, the effort for interference
chip. The AWR1843BOOST device has three transmit and
mitigation for the next measurement cycle can be saved, if the
four receive channels and it is capable of operation in bands
interferer occupies the current frequency band for its further
ranging from 76 to 81 GHz. The two TXs at the same elevation
measurements and there are no new interferers. The kernel size
are activated for the radar measurements.
is set to 1x9x3 for increasing the receptive field by including
Since the measurements are recorded on the street, the
more samples in different chirps and receive channels.4 The
measured targets are distributed in the radar sensor’s field of
number of filters per layer and the kernel size can be adjusted
view and vary from scenario to scenario. The DoA of the
in case of different input dimensions.
simulated interferer corresponds to the field of view of the
victim radar.
B. Loss Function The signal separation neural network is trained using an
The loss function is defined as the mean absolute error loss NVIDIA RTXTM 3060 for 100 epochs with a batch size of
of the reconstructed discrete beat signal and the reconstructed four. It is found that 100 training epochs are sufficient for
interference signal. providing satisfactory signal reconstructions. The stride of the
convolution is one. The learning rate is set to 0.0001 and
1 X e qI ,

L AB S = Ys − bYs + e YI − Y (13) the ADAM optimizer [36] is used during training. The
ϵ ELU activation function was chosen because it avoids the
4 Increasing the kernel size in the first dimension (e.g., 2x9x3) can slightly “dying ReLU” problem [37]. According to our experiments,
improve the quality of signal reconstruction compared to the kernel size
1x9x3, but at the cost of longer inference time of SeparationNet. 5 https://www.ti.com/tool/AWR1843BOOST

Authorized licensed use limited to: IMEC. Downloaded on June 23,2023 at 12:15:34 UTC from IEEE Xplore. Restrictions apply.
CHEN et al.: IMPROVED TARGET DETECTION THROUGH DNN-BASED MULTI-CHANNEL INTERFERENCE MITIGATION 81

TABLE I
R ADAR PARAMETERS OF THE V ICTIM R ADAR AND
THE S IMULATED I NTERFERER

Fig. 7. Interference in the discrete beat signal domain: (a) Real-world

measured disturbed discrete beat signal, (b) simulated disturbed discrete beat
the reconstruction accuracy could be slightly improved by signal with a single interferer at a distance of about 2 meters, (c) simulated
increasing the number of filters in each layer and adding disturbed discrete beat signal with a single interferer at a distance of about
more layers to the current neural network architecture (see 20 meters, (d) simulated disturbed discrete beat signal with four different
interferers at different distances. For an intuitive visualization, only the real
Fig. 4). However, the training and inference time will also part of the complex discrete beat signal is shown.
increase if more filters and layers are included. The inference
time of the proposed SeparationNet is about five milliseconds
(corresponding to ca. 57.60 giga floating point operations)
real-world interference scenarios even though the interference
on NVIDIA RTXTM 3060. The data set contains 20374 real-
is purely simulated during training.
world measured interference-free complex discrete beat signals
having a size of [P, Q, M N ] which is split into training
data set and the evaluation data set containing a subset IV. E VALUATION M ETHODS AND R ESULTS
of 18342 and 2532 (including 500 highway and country
road scenarios) respectively. During the training process, the In this section, we first briefly analyze the convergence of
interference signals are randomly simulated in each training the two loss components (i.e., the loss of the reconstructed
epoch and thus the interference scenario is completely new signal segment and the loss of the reconstructed interference
for the neural network in each epoch. We therefore do not segment) in (13). Then, the performance of the proposed
specifically prepare a validation data set. During the training signal separation neural network is evaluated and compared to
and evaluation processes, the complex discrete beat signal is the state-of-the-art algorithms. The neural network trained on
then separated into real and imaginary parts, resulting in a L AB S is used for the evaluation, since L AB S and L R M S E per-
tensor of the form [P, Q, M N , 2]. In each training iteration, form similarly in our experiments. The signal-to-interference-
the additive interference of randomly one to four interferers plus-noise ratio (SINR) [13], the false-positive rate (FPR),
are simulated based on (11), where each interferer has different the true-positive rate (TPR), the peak phase mean-squared
signal parameters that vary randomly according to Table I. The error (PPMSE) [24], the mean absolute percentage error
amplitude of the additive interference varies between 50 and (MAPE) [13], and the F1-score are used as evaluation metrics.
10000. For normalization, each sample of the input data is The F1-score is defined as F1 = 2 TP +2 FP TP
+ FN
where TP,
divided by 250. This value is chosen based on the observation FP and FN denote the number of true-positive, false-positive,
that the maximum amplitude of the undisturbed discrete beat and false-negative estimates. The true positive rate and false
signals in the training data set is about 250 (i.e., the maximum positive rate are defined as TPR = TP TP + FN
and FPR =
FP
amplitude of the target signal segment). FP + TN
, respectively, where TN denotes the true-negative
Fig. 7 shows examples of real-world measured and simu- estimates. A cell averaging CFAR (CA-CFAR) target detection
lated disturbed discrete beat signals. In our simulation, we sim- algorithm is used to automatically detect the peak positions in
ulated not only the strong interference (Fig. 7 (b)) which has an both the interference-free and interference-contaminated RD
interferer at a distance of about 2 meters but also interference spectra. The peak positions detected in the interference-free
with medium amplitude (Fig. 7 (c)). In order to imitate RD spectra are then used as the ground truth for the calculation
more complex interference scenarios (e.g., intersection and of TP, FP, TN, and FN. If a peak detection is found only in
congestion scenarios), we simulated interference from multiple the reconstructed RD spectrum but not in the interference-
interferers (Fig. 7 (d)) at different distances. free RD spectrum, this detection is classified as FP. If a peak
Based on the signal model formulated in (11), the sim- detection is found in both the reconstructed RD spectrum and
ulated interference can appropriately imitate the real-world the interference-free RD spectrum, this detection is recognized
measured interference. The simulated interference is proven to as TP. The SINR, FPR, TPR, PPMSE, MAPE, and F1-score
be effective as the trained neural network can later handle the are calculated in each receive channel. Then, the mean values

Authorized licensed use limited to: IMEC. Downloaded on June 23,2023 at 12:15:34 UTC from IEEE Xplore. Restrictions apply.
82 IEEE TRANSACTIONS ON RADAR SYSTEMS, VOL. 1, 2023

TABLE II
A NALYSIS R ESULTS OF THE R ECONSTRUCTION IN T ERMS OF SINR, F1-
S CORE , FPR, AND TPR FOR D IFFERENT M ITIGATION T ECHNIQUES .
T HE R ESULTS OF THE P ROPOSED M ETHOD A RE
H IGHLIGHTED IN B LUE

completely obscured by the high level of background noise in

the disturbed RD spectrum. The target peak at a distance of
13.8 m with a velocity of −10.08 m/s can be easily detected in
the RD spectra recovered by the zeroing method [3], AR [12],
Fig. 8. Training loss curves of the reconstructed signal segment and the and the proposed SeparationNet, while it is absent in the
reconstructed interference segment. RD spectrum recovered by CNN [22]. The SeparationNet can
reconstruct the target peak more accurately than the zeroing
and AR methods, as shown in Fig. 10, which gives the range
of these evaluation metrics of all receive channels are used for profile of the recovered RD spectra at velocity −10.08 m/s.
the performance assessment. The mean value of SINR, F1-score, FPR, and TPR of
Moreover, based on the reconstructed discrete beat signals all receiving channels are summarized in Table II. Although
of these algorithms, the corresponding RAD data cubes can be SeparationNet is trained with purely simulated interference,
obtained (see Fig. 3) and the accuracy of target detection and it can still provide the best signal recovery in this real-
classification is further evaluated based on a state-of-the-art world measured interference scenario. The FPR of Separa-
automotive radar object detection neural network [2]. Finally, tionNet is two orders of magnitude less than the FPR for
the benefits of applying the proposed interference mitigation the disturbed RD spectra and one order of magnitude less
approach are demonstrated. than the state-of-the-art interference mitigation algorithms.
Note that all interference-contaminated samples are assumed
A. Convergence Analysis Based on Training Loss Curve to be perfectly detected for the zeroing method and the AR
model in our evaluation. Based on our experiments, when
The loss curves of the reconstructed signal segment and 1% of interference-contaminated samples are not detected, the
the reconstructed interference segment are presented in Fig. 8, SINR of the RD spectra reconstructed by the zeroing and AR
where we calculated the loss of signal and interference (see methods significantly reduces to ca. 16.35 dB and 17.21 dB
(13)) separately. Compared to the loss of the reconstructed in this real-world interference scenario, respectively.
interference segment, the loss of the reconstructed signal
segment has a smaller value and a faster convergence rate.
This is due to the fact that the amplitude of the interference C. Performance Evaluation on Signal Reconstruction With
segment fluctuates in a larger range than the signal segment, Simulated Interference
as described in section III-C. As a result, the error value
of the reconstructed interference segment may deviate to a Since it is quite difficult to obtain reference measurements
greater extent. Both losses converge to a steady state after for all real-world measured interference scenarios, we fur-
about 100 training iterations. ther evaluate the performance of these interference mitiga-
tion algorithms with simulated interference (11), where the
discrete beat signals are measured in the real world [2].
B. Performance Evaluation on Signal Reconstruction With In this section, the performance of proposed method along
Measured Interference with the simple zeroing method [3], the AR model [12], and
Recent research results have shown that the zeroing the CNN-based RD spectra reconstruction [22] are evaluated
method [3], the AR model [12], and deep learning algorithms in terms of SINR, PPMSE, MAPE, and F1-score. Here, all
like the CNN-based RD spectrum reconstruction [17], [22] interference-contaminated samples are also assumed to be
can be utilized for interference mitigation. We first evaluate perfectly detected for the zeroing method and the AR model
the SINR, F1-score, FPR, and TPR of the reconstructed RD in our evaluation. To demonstrate the advantages of training
spectra that obtained by applying the proposed SeparationNet the neural network with multi-channel discrete beat signals
and the state-of-the-art algorithms. In this real-world interfer- as input, the proposed SeparationNet is also trained with only
ence scenario, the interfering radar which has a chirp rate of one receive channel and then used to reconstruct multi-channel
1/3 of the victim radar, is placed at a distance of 2 meters in discrete beat signals. For distinction, the SeparationNet trained
front of the victim radar. with single-channel discrete beat signals is referred to as
Fig. 9 shows the undisturbed and disturbed RD spectra SeparationNet-S and the SeparationNet trained with multi-
as well as the recovered RD spectra. The target peaks are channel discrete beat signals is referred to as SeparationNet-M.

Authorized licensed use limited to: IMEC. Downloaded on June 23,2023 at 12:15:34 UTC from IEEE Xplore. Restrictions apply.
CHEN et al.: IMPROVED TARGET DETECTION THROUGH DNN-BASED MULTI-CHANNEL INTERFERENCE MITIGATION 83

Fig. 9. (a) Undisturbed RD spectrum, (b) disturbed RD spectrum, and RD spectra recovered by (c) CNN, (d) the zeroing method, (e) AR, (f) SeparationNet.

Fig. 11 shows the SINR of the recovered RD spectra

of three reference algorithms: AR [12], CNN [22], and the
zeroing method [3], as well as the proposed SeparationNet.
Note that we have extended the input channel of the original
single-channel CNN implementation in [22] to reconstruct the
multi-channel RD spectra. For the reconstruction of a multi-
channel discrete beat signal, the zeroing method, AR, and
the SeparationNet trained with single-channel interference are
applied in each signal channel and the reconstructed signal
in each individual channel is then combined for the further
analysis. The SINR of RD spectra recovered by the AR
and zeroing methods decreases as the interference intensity
increases, while the CNN and SeparationNet are less affected
by the interference intensity. The CNN has shown its strength
in RD spectra denoising and provides the best SINR of the
recovered RD spectra. However, when the interference mag-
Fig. 10. Range profile of RD spectra at velocity −10.08 m/s. nitude becomes very large (e.g., in the real-world measured
interference which causes a degradation of ca. −25 dB SINR
Due to the diversity of chirp rate, the collision time between of the RD spectrum), the SINR provided by CNN degrades to
interferer and victim radar, the number of interferers, the the suppression of both strong and weak target peaks in the
number of disturbed samples, and the signal strength of inter- reconstructed RD spectrum.
ference can vary in different interference scenarios. In order to As shown in Fig. 12, the proposed SeparationNet can deliver
evaluate the performance of these aforementioned algorithms, a superior F1-score under different interference intensities.
three classes of interference scenarios are simulated: weak, Due to the absence of some weak target peaks in the recon-
medium, and strong, corresponding to a degradation of ca. structed RD spectra (see Fig. 9 (c)), the CNN therefore does
−5 dB, −10 dB, and −15 dB SINR of the RD spectra not provide an optimal F1-score. The SeparationNet trained
(compared to the interference-free scenario), respectively. The with single-channel interference outperforms the state-of-the-
empirical mean and standard deviation of the SINR of the art interference mitigation algorithms, while the SeparationNet
interference-free RD spectra in the evaluation data set are trained with multi-channel interference further improves the
26.67 dB and 3.44 dB, respectively. Note that the proposed F1-score by 3 - 5%. Fig. 13 shows the evaluation results
neural network is trained with the interference scenarios that of the MAPE of reconstructed target peaks detected by the
cause -10 dB to -5 dB SINR degradation of the RD spectra. CA-CFAR. For all signal reconstruction methods, the MAPE

Authorized licensed use limited to: IMEC. Downloaded on June 23,2023 at 12:15:34 UTC from IEEE Xplore. Restrictions apply.
84 IEEE TRANSACTIONS ON RADAR SYSTEMS, VOL. 1, 2023

Fig. 14. PPMSE of target peaks in the reconstructed RD spectra of different

Fig. 11. SINR of the reconstructed RD spectra of different interference interference mitigation methods in the presence of weak (‘w’), medium (‘m’),
mitigation methods in the presence of weak, medium, and strong interference. and strong (‘s’) interference.

provides superior reconstruction under different interference

intensities. Different from the evaluation related to SINR, the
weak target peaks can cause a similar MAPE or PPMSE as
the strong target peaks. Therefore, the absence of weak target
peaks in the reconstructed RD spectra significantly degrades
the results of CNN. Even though the AR and zeroing meth-
ods provide lower MAPE and PPMSE than the CNN under
the assumption of a perfect detection of disturbed samples,
it should be noted that this assumption does not hold in most
real-world interference scenarios. Thus, the recovered results
of the AR and zeroing methods might deteriorate in real-
world scenarios due to the imperfect detection of the disturbed
samples.
Fig. 12. F1-score for the target peak detection in the reconstructed RD Compared to the strong target peaks such as cars or trucks,
spectra of different interference mitigation methods in the presence of weak, the relatively weak target peaks usually represent the vul-
medium, and strong interference.
nerable road users (e.g., pedestrians and bicyclists) whose
signals are more likely to be affected by mutual interference.
To evaluate the quality of the reconstructed target peaks at
different interference densities, an intersection scenario is
chosen where pedestrians, bicyclists, and cars are present,
which well represent the targets with low, medium, and high
SNR. The SNR of the weak target peaks is about 3 dB below
the medium target peaks, while the SNR of the medium target
peaks is about 4 dB smaller than the strong target peaks. The
angle estimation error of the weak, medium and strong target
peaks as well as the receiver operating characteristic (ROC)
of the detection of these target peaks are further evaluated in
this section.
Table III shows the evaluation results of the RMSE of
the angle estimation of the proposed method compared with
Fig. 13. MAPE of target peaks in the reconstructed RD spectra of different
other state-of-the-art interference mitigation algorithms. When
interference mitigation methods in the presence of weak (‘w’), medium (‘m’), the power of interference increases, the angle estimation of
and strong (‘s’) interference. weak targets based on the reconstructed signals of all methods
contains more errors, while the proposed SeparationNet pro-
vides the smallest RMSE. For the targets with medium and
increases with increasing interference intensity. The target strong SNR, the angle estimation errors of the SeparationNet,
peaks reconstructed by the proposed method have the smallest Zeroing, and AR methods are smaller than the angle estimation
MAPE. Similar results can also be found in the PPMSE of error for the weak targets. As the results presented in Fig. 14
the reconstructed target peaks in Fig. 14, where the proposed show, SeparationNet can provide the smallest phase errors
SeparationNet (both SeparationNet-S and SeparationNet-M) for the detected target peaks. Thus, the angle estimations of

Authorized licensed use limited to: IMEC. Downloaded on June 23,2023 at 12:15:34 UTC from IEEE Xplore. Restrictions apply.
CHEN et al.: IMPROVED TARGET DETECTION THROUGH DNN-BASED MULTI-CHANNEL INTERFERENCE MITIGATION 85

TABLE III
ROOT M EAN S QUARED E RROR ( IN R ADIANS ) OF A NGLE E STIMATION OF W EAK , M EDIUM , AND S TRONG TARGETS U NDER D IFFERENT I NTERFERENCE
S CENARIOS . T HE R ESULTS OF THE P ROPOSED M ETHOD A RE H IGHLIGHTED IN B OLD

downsampling blocks. The coordinate transformation layer

takes the feature maps extracted by RadarResNet and generates
Cartesian feature maps which can then be used for object
detection in Cartesian coordinates with the YOLO detection
head. The YOLO detection head is an anchor-based method
proposed in [40]. The output of the YOLO detection head
is the predicted bounding boxes for the detected objects.
More details regarding the RAD-based object detection neural
network can be found in [2]. To evaluate object detection
performance, the average precision (AP) [41] is used which
indicates the percentage of ground-truth objects among all
detected objects. A detection can be classified as true positive
or false positive by comparing the intersection over union
(IoU) of the ground-truth and predicted bounding boxes with
a predefined threshold. The IoU is used to measure how well
the ground-truth and predicted boxes match and is defined as
Area of intersection |C ∩ D|
Fig. 15. The ROC for the detection of target peaks by CA-CFAR in the IoU = = ,
reconstructed RD spectra of different interference mitigation methods in the Area of union |C ∪ D|
presence of strong interference.
where C and D are the prediction and ground truth bounding
boxes. Since multiple object classes are defined, the mean
SeparationNet for the medium and strong targets are iden- AP (mAP) is employed as a metric to measure the accuracy
tical to the angle estimations based on the undisturbed RD of object detection across all classes. A higher mAP score
spectra. indicates a more accurate detection.
Fig. 15 shows the ROC for the detection of target peaks The mAP of the object detection based on interference-free
by CA-CFAR in the reconstructed RD spectra of different RADs is compared with the mAP of the object detection based
interference mitigation methods in the presence of strong on the disturbed RADs (with three interference intensities)
interference. The area under the ROC (AUC) is an indicator of and the RADs reconstructed by the proposed SeparationNet.
the target detection performance. The maximum value of AUC Table IV shows the mAP of the object detection for three
is 1, which means that true-positive rate is high regardless object classes (car, person, and truck) at four different IoUs
of false-positive rate. The proposed SeparationNet provides (0.1, 0.3, 0.5, and 0.7). Even with weak interference, the mAP
the largest AUC, which means that most of the target peaks drops significantly for these four IoUs. However, with the
(including the weak target peaks) can be recovered after the proposed signal reconstruction method, the mAP can be signif-
interference mitigation has been applied. icantly improved even in the presence of strong interference.

D. Performance Evaluation of Reconstructed Discrete Beat E. Discussion

Signal on Target Detection With the proposed signal model and the corresponding data
In this subsection, we evaluate the impact of interference generation framework, a wide variety of interference scenarios
reduction on the object detection accuracy using a state-of- can be considered in the training process. However, since each
the-art RAD-based object detection neural network (RadarRes- manufacturer of automotive radar sensors has its own specific
Net) [2]. Note that the phase compensation [38] for the moving modulations of the radar signal, a comprehensive collection
target is not applied in the RAD-based object detection devel- of all interference scenarios for training is almost impossible.
oped in [2]. Since the aim of this work is to improve the object By simulating the radar signal parameters of the interferer
detection accuracy in the presence of interference, we use randomly in each training epoch, which can be considered as
the RadarResNet as is. The proposed signal separation neural a data augmentation technique [42], the trained neural network
network is integrated into the processing pipeline as shown in can properly handle a wide range of interference scenarios and
Fig. 16. The RadarResNet [2] is based on ResNet [39] which avoid overfitting. However, new modulation techniques may
consists of two types of blocks, namely the residual blocks and emerge. In this context it should be noted that the interference

Authorized licensed use limited to: IMEC. Downloaded on June 23,2023 at 12:15:34 UTC from IEEE Xplore. Restrictions apply.
86 IEEE TRANSACTIONS ON RADAR SYSTEMS, VOL. 1, 2023

Fig. 16. Dataflow of RAD-based object detection in the presence of interference extended based on [2] with the proposed signal separation neural network.

TABLE IV
M EAN AVERAGE P RECISION OF RAD-BASED O BJECT D ETECTION
AT D IFFERENT I O U S W ITH U NDISTURBED RAD S , I NTERFERENCE -
C ONTAMINATED RAD S , AND THE C ORRESPONDING RAD S R ECON -
STRUCTED BY THE P ROPOSED S EPARATION N ET

between the digital modulated radar and the chirp sequence

radar is not considered in this work.
In comparison to the proposed method, the zeroing and
AR methods are limited not only in the quality of the
reconstructed signal but also in the accuracy of the detected
interference segments, since the positions of the disturbed
samples must first be detected and the signal segments in
these disturbed positions can then be reconstructed. However,
it is difficult to accurately detect all disturbed samples, as also Fig. 17. Examples of discrete beat signals and RD spectra in a real-world
discussed in [22], where a detection accuracy of 90% is measured dynamic scenario: (a) without interference, (b) with interference
(which is difficult to be detected with the classical amplitude-based threshold-
assumed. As shown in Fig. 17, if the amplitude of the ing algorithm), and (c) recovered by the proposed method. The corresponding
disturbed signal samples is in the same range as the ampli- RD spectra of (a), (b), and (c) are presented in (d), (e), and (f), respectively.
tude of the interference-free signal samples, these disturbed The target peaks are circled in red ellipses. For an intuitive visualization, only
the real part of the complex discrete beat signal is shown in (a), (b), and (c).
samples are difficult to detect by the classical amplitude-
based thresholding algorithms. Although the amplitude of
these disturbed samples is relatively small (compared to the
interference scenarios shown in Fig. 7), the disturbance in the after FFT, so that the weak target peaks are obscured by the
RD spectrum is not negligible (see Fig. 17(e)). Nevertheless, high levels of background noise. However, reconstructing these
the proposed method can reconstruct the target signal samples few disturbed discrete beat signal samples in the time domain
well (see Fig. 17(c) and (f)). In contrast to the zeroing is easier and yields better results. Utilizing the signal features
and AR methods, the proposed method and the CNN for learned from multiple receive channels can further improve
RD spectra reconstruction [22] avoid the additional detection signal reconstruction results. For an imaging radar with mul-
process. tiple TXs at different elevations, the input of the proposed
The performance difference between the proposed method neural network can be formulated as a tensor in the form
and the CNN [22] is mainly caused by two factors, namely [P, Q, N , M, 2] instead of stacking all virtual channels on top
the type of input signal and the number of channels of the of each other (i.e., [P, Q, M N , 2]). Then, a 3D convolution
input signal used during training. Even if only a few discrete can be applied in the first three dimensions to extract the
beat signal samples are disturbed by strong interference, the phase information from the received signals of RXs (at the
energy of the interference spreads to the whole RD spectrum same elevation) coming from the same TX for interference

Authorized licensed use limited to: IMEC. Downloaded on June 23,2023 at 12:15:34 UTC from IEEE Xplore. Restrictions apply.
CHEN et al.: IMPROVED TARGET DETECTION THROUGH DNN-BASED MULTI-CHANNEL INTERFERENCE MITIGATION 87

mitigation. In this way, the phase offsets between the virtual [3] G. M. Brooker, “Mutual interference of millimeter-wave radar sys-
antenna elements of different transmitters would not affect the tems,” IEEE Trans. Electromagn. Compat., vol. 49, no. 1, pp. 170–181,
Feb. 2007.
reconstruction performance. [4] J. Khoury, R. Ramanathan, D. McCloskey, R. Smith, and T. Campbell,
Most commercial automotive radar sensors have dedicated “RadarMAC: Mitigating radar interference in self-driving cars,” in Proc.
antenna patterns for the intended applications (e.g., adaptive 13th Annu. IEEE Int. Conf. Sens., Commun., Netw. (SECON), Jun. 2016,
pp. 1–9.
cruise control). Thus, the antenna gain may vary on different
[5] C. Aydogdu, M. F. Keskin, N. Garcia, H. Wymeersch, and D. W. Bliss,
receive channels. The amplitude of the received signal can “RadChat: Spectrum sharing for automotive radar interference mitiga-
therefore vary and result in different signal distributions on tion,” IEEE Trans. Intell. Transp. Syst., vol. 22, no. 1, pp. 416–429,
each receive channel. By considering the received multi- Jan. 2021.
[6] Z. Xu and Q. Shi, “Interference mitigation for automotive radar using
channel signals as the input, the neural networks can learn the orthogonal noise waveforms,” IEEE Geosci. Remote Sens. Lett., vol. 15,
joint distribution and reconstruct the interference-free signal no. 1, pp. 137–141, Jan. 2018.
segments properly. [7] F. Uysal, “Phase-coded FMCW automotive radar: System design and
Since the TDM-MIMO scheme has the disadvantage of a interference mitigation,” IEEE Trans. Veh. Technol., vol. 69, no. 1,
pp. 270–281, Jan. 2020.
reduced unambiguous speed [43], [44], i.e., if the target’s true [8] F. Uysal and S. Sanka, “Mitigation of automotive radar interference,” in
radial velocity exceeds the maximum unambiguous velocity Proc. IEEE Radar Conf. (RadarConf18), Apr. 2018, pp. 405–410.
range of the TDM waveform, the measured velocity will be [9] S. Chen, J. Taghia, U. Kühnau, T. Fei, F. Grünhaupt, and
incorrect and aliased. It is of interest to investigate the perfor- R. Martin, “Automotive radar interference reduction based on sparse
Bayesian learning,” in Proc. IEEE Radar Conf. (RadarConf), Sep. 2020,
mance of the proposed algorithm in the cases of ambiguous pp. 1–6.
speed in future work. [10] S. Lee, J. Lee, and S. Kim, “Mutual interference suppression using
wavelet denoising in automotive FMCW radar systems,” IEEE Trans.
Intell. Transp. Syst., vol. 22, no. 2, pp. 887–897, Feb. 2021.
V. C ONCLUSION [11] J. Mun, S. Ha, and J. Lee, “Automotive radar signal interfer-
ence mitigation using RNN with self attention,” in Proc. IEEE
In this paper, we present a signal separation neural net- Int. Conf. Acoust., Speech Signal Process. (ICASSP), May 2020,
work for interference mitigation in FMCW MIMO radar. The pp. 3802–3806.
[12] M. Rameez, M. Dahl, and M. I. Pettersson, “Autoregressive
network estimates not only an interference-free multi-channel model-based signal reconstruction for automotive radar interfer-
beat signal but also the multi-channel interference. By synthe- ence mitigation,” IEEE Sensors J., vol. 21, no. 5, pp. 6575–6586,
sizing interference signals using a novel interference model Mar. 2021.
and adding them to real-world radar measurements at multiple [13] S. Chen, J. Taghia, U. Kühnau, N. Pohl, and R. Martin, “A two-
stage DNN model with mask-gated convolution for automotive radar
receive antennas, we create a new interference-contaminated interference detection and mitigation,” IEEE Sensors J., vol. 22, no. 12,
discrete beat signal data set. The evaluation demonstrates that pp. 12017–12027, Jun. 2022.
the proposed method yields superior recovery in terms of [14] J. Wang, M. Ding, and A. Yarovoy, “Matrix-pencil approach-
based interference mitigation for FMCW radar systems,” IEEE
SINR, FPR, F1-score, angle estimation, MAPE, and PPMSE Trans. Microw. Theory Techn., vol. 69, no. 11, pp. 5099–5115,
compared to the other reconstruction algorithms in this evalu- Nov. 2021.
ation. Thus, this data-driven pre-processing approach provides [15] J. Bechter, F. Roos, M. Rahman, and C. Waldschmidt, “Automotive radar
an effective countermeasure against interference. In addi- interference mitigation using a sparse sampling approach,” in Proc. Eur.
Radar Conf. (EURAD), Oct. 2017, pp. 90–93.
tion, the evaluation with real-world interference demonstrates [16] S. Chen, J. Taghia, T. Fei, U. Kühnau, N. Pohl, and R. Martin, “A DNN
the effectiveness of the proposed interference signal model. autoencoder for automotive radar interference mitigation,” in Proc.
Although the neural network has been trained only with IEEE Int. Conf. Acoust., Speech Signal Process. (ICASSP), May 2021,
pp. 4065–4069.
simulated interference, it also properly handles real-world
[17] J. Rock, M. Toth, E. Messner, P. Meissner, and F. Pernkopf, “Complex
measurements. With the help of the proposed method, the signal denoising and interference mitigation for automotive radar using
accuracy of target detection in the presence of interference convolutional neural networks,” in Proc. 22th Int. Conf. Inf. Fusion
(evaluated in terms of the mean average precision) is also (FUSION), Jul. 2019, pp. 1–8.
[18] F. Jin and S. Cao, “Automotive radar interference mitigation using
significantly improved. Finally, the proposed neural network adaptive noise canceller,” IEEE Trans. Veh. Technol., vol. 68, no. 4,
could also be combined with the classical frequency hopping pp. 3747–3754, Apr. 2019.
algorithm for interference mitigation, since the interference [19] S. Chen, W. Shangguan, J. Taghia, U. Kühnau, and R. Martin, “Auto-
signal segments are available as well. Since the data set used motive radar interference mitigation based on a generative adversarial
network,” in Proc. IEEE Asia–Pacific Microw. Conf. (APMC), Dec. 2020,
in this work was measured with TX antennas on the same pp. 728–730.
elevation, the proposed algorithm still needs to be verified on [20] S. Neemat, O. Krasnov, and A. Yarovoy, “An interference mitigation
data sets measured with TX antennas on different elevations technique for FMCW radar using beat-frequencies interpolation in the
STFT domain,” IEEE Trans. Microw. Theory Techn., vol. 67, no. 3,
in future work. pp. 1207–1220, Mar. 2019.
[21] J. Fuchs, A. Dubey, M. Lübke, R. Weigel, and F. Lurz, “Automotive
radar interference mitigation using a convolutional autoencoder,” in
R EFERENCES Proc. IEEE Int. Radar Conf. (RADAR), Apr. 2020, pp. 315–320.
[22] J. Rock, W. Roth, M. Toth, P. Meissner, and F. Pernkopf, “Resource-
[1] O. Schumann, J. Lombacher, M. Hahn, C. Wöhler, and J. Dickmann, efficient deep neural networks for automotive radar interference mitiga-
“Scene understanding with automotive radar,” IEEE Trans. Intell. Vehi- tion,” IEEE J. Sel. Topics Signal Process., vol. 15, no. 4, pp. 927–940,
cles, vol. 5, no. 2, pp. 188–203, Jun. 2020. Jun. 2021.
[2] A. Zhang, F. E. Nowruzi, and R. Laganiere, “RADDet: Range-Azimuth- [23] C. Jiang, T. Chen, and B. Yang, “Adversarial interference mitigation for
Doppler based radar object detection for dynamic road users,” in Proc. automotive radar,” in Proc. IEEE Radar Conf. (RadarConf), May 2021,
18th Conf. Robots Vis. (CRV), May 2021, pp. 95–102. pp. 1–6.

Authorized licensed use limited to: IMEC. Downloaded on June 23,2023 at 12:15:34 UTC from IEEE Xplore. Restrictions apply.
88 IEEE TRANSACTIONS ON RADAR SYSTEMS, VOL. 1, 2023

[24] A. Fuchs, J. Rock, M. Toth, P. Meissner, and F. Pernkopf, “Complex- Shengyi Chen received the B.Sc. degree in electrical
valued convolutional neural networks for enhanced radar signal and computer engineering from the Technical
denoising and interference mitigation,” in Proc. IEEE Radar Conf. University of Kaiserslautern, Germany, Fuzhou Uni-
(RadarConf), May 2021, pp. 1–6. versity, China, in 2016, and the M.Sc. degree in
[25] M. Umehira, T. Okuda, X. Wang, S. Takeda, and H. Kuroda, “An adap- electrical and computer engineering from the Tech-
tive interference detection and suppression scheme using iterative pro- nical University of Munich, Germany. He is cur-
cessing for automotive FMCW radars,” in Proc. IEEE Radar Conf. rently pursuing the Dr.-Ing. degree in electrical
(RadarConf), Sep. 2020, pp. 1–5. engineering and information technology with the
[26] S. Sun, A. P. Petropulu, and H. V. Poor, “MIMO radar for advanced Ruhr-Universität Bochum, Germany.
driver-assistance systems and autonomous driving: Advantages and His research interests are automotive radar sig-
challenges,” IEEE Signal Process. Mag., vol. 37, no. 4, pp. 98–117, nal processing, machine learning, and compressive
Jul. 2020. sensing.
[27] F. Engels, P. Heidenreich, M. Wintermantel, L. Stäcker, M. Al Kadi, and
A. M. Zoubir, “Automotive radar signal processing: Research directions
and practical challenges,” IEEE J. Sel. Topics Signal Process., vol. 15,
no. 4, pp. 865–878, Jun. 2021.
[28] N. Scheiner, N. Appenrodt, J. Dickmann, and B. Sick, “Radar-based road
user classification and novelty detection with recurrent neural network
ensembles,” in Proc. IEEE Intell. Vehicles Symp. (IV), Jun. 2019, Marvin Klemp received the B.Sc. and M.Sc. degree
pp. 722–729. in computer science from the Hochschule Bonn-
[29] N. Scheiner, F. Kraus, N. Appenrodt, J. Dickmann, and B. Sick, Rhein-Sieg, Sankt Augustin, Germany, in 2018 and
“Object detection for automotive radar point clouds—A comparison,” 2021, respectively. He is currently pursuing the
AI Perspect., vol. 3, no. 1, pp. 1–23, Nov. 2021. Ph.D. degree with the Institute of Measurement and
[30] A. Ouaknine, A. Newson, J. Rebut, F. Tupin, and P. Pérez, “CARRADA Control Systems, Karlsruhe Institute of Technology,
dataset: Camera and automotive radar with range-angle-Doppler anno- Karlsruhe, Germany.
tations,” in Proc. 25th Int. Conf. Pattern Recognit. (ICPR), Jan. 2021, His research interests include deep learning based
pp. 5068–5075. perception systems for intelligent transportation
[31] M. Kronauge and H. Rohling, “New chirp sequence radar waveform,” systems.
IEEE Trans. Aerosp. Electron. Syst., vol. 50, no. 4, pp. 2870–2877,
Oct. 2014.
[32] K. Rambach, Direction of Arrival Estimation Using a Multiple-Input-
Multiple-Output Radar With Applications to Automobiles. Stuttgart,
Germany: Universität Stuttgart, 2017.
[33] J. Gamba, Radar Signal Processing for Autonomous Driving, 1st ed.
Cham, Switzerland: Springer, 2019.
[34] Y. LeCun and Y. Bengio, Convolutional Networks for Images,
Speech, and Time Series. Cambridge, MA, USA: MIT Press, 1998,
pp. 255–258. Jalal Taghia received the B.Sc. degree in electrical
[35] J. Bechter, C. Sippel, and C. Waldschmidt, “Bats-inspired frequency engineering from Azad University, Ghazvin, Iran,
hopping for mitigation of interference between automotive radars,” in in 2006. He pursued further studies in electrical
IEEE MTT-S Int. Microw. Symp. Dig., May 2016, pp. 1–4. engineering and M.Sc. degree from Shahid Beheshti
[36] D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” University, Tehran, Iran, in 2009, and the Dr.-Ing.
in Proc. 3rd Int. Conf. Learn. Represent., (ICLR), San Diego, CA, USA, degree from the Institute of Communication Acous-
May 2015. tics (IKA), Ruhr-Universität Bochum, Bochum,
[37] L. Lu, Y. Shin, Y. Su, and G. E. Karniadakis, “Dying ReLU and initial- Germany, in 2016.
ization: Theory and numerical examples,” 2019, arXiv:1903.06733. From 2009 to 2012, he served as a Research Fel-
[38] J. Bechter, F. Roos, and C. Waldschmidt, “Compensation of low in the EU Marie Curie Initial Training Network
motion-induced phase errors in TDM MIMO radars,” IEEE AUDIS, focusing on “Digital Signal Processing in
Microw. Wireless Compon. Lett., vol. 27, no. 12, pp. 1164–1166, Audiology”. During his time at IKA from 2016 to 2018, he worked as a
Dec. 2017. Postdoctoral Researcher specializing in instrumental assessment of speech
[39] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for intelligibility, speech enhancement, and information theory for signal pro-
image recognition,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. cessing. Subsequently, from 2018 to 2022, he served as a Signal Processing
(CVPR), Jun. 2016, pp. 770–778. Engineer at HELLA GmbH & Co. KGaA, Lippstadt, Germany. His primary
[40] A. Bochkovskiy, C.-Y. Wang, and H.-Y. M. Liao, “YOLOv4: Optimal responsibility was the development of robust signal processing algorithms
speed and accuracy of object detection,” 2020, arXiv:2004.10934. for automotive radar systems. Since 2022, he has been working as a Senior
[41] R. Padilla, S. L. Netto, and E. A. B. da Silva, “A survey on performance SW developer and Radar Expert at HELLA Aglaia Mobile Vision GmbH in
metrics for object-detection algorithms,” in Proc. Int. Conf. Syst., Signals Berlin. In this role, he focuses on the development of embedded software for
Image Process. (IWSSIP), Jul. 2020, pp. 237–242. Automotive Radar Systems.
[42] A. Mikolajczyk and M. Grochowski, “Data augmentation for
improving deep learning in image classification problem,” in
Proc. Int. Interdiscipl. PhD Workshop (IIPhDW), May 2018,
pp. 117–122.
[43] F. Roos, J. Bechter, N. Appenrodt, J. Dickmann, and C. Waldschmidt,
“Enhancement of Doppler unambiguity for chirp-sequence modu-
lated TDM-MIMO radars,” in IEEE MTT-S Int. Microw. Symp. Dig., Uwe Kühnau received the Diploma degree in physics and the Dr. rer. nat.
Apr. 2018, pp. 1–4. degree in solid state physics from the University of Leipzig, Leipzig, Germany,
[44] M. Dikshtein, O. Longman, S. Villeval, and I. Bilik, “Automotive in 1993 and 1998, respectively.
radar maximum unambiguous velocity extension via high-order phase He started his industrial career in pre-development for automotive sensors
components,” IEEE Trans. Aerosp. Electron. Syst., vol. 58, no. 1, at Hella in 2000 and later headed the advanced engineering for radar systems.
pp. 743–751, Feb. 2022. He is currently Head of Radar Systems at Forvia.

Authorized licensed use limited to: IMEC. Downloaded on June 23,2023 at 12:15:34 UTC from IEEE Xplore. Restrictions apply.
CHEN et al.: IMPROVED TARGET DETECTION THROUGH DNN-BASED MULTI-CHANNEL INTERFERENCE MITIGATION 89

Nils Pohl (Senior Member, IEEE) received the Rainer Martin (Fellow, IEEE) received the
Dipl.-Ing. and Dr.Ing. degrees in electrical engi- M.S.E.E. degree from Georgia Institute of Technol-
neering from Ruhr University Bochum, Bochum, ogy, Atlanta, in 1989, and the Dipl.-Ing. and Dr.-Ing.
Germany, in 2005 and 2010, respectively. degrees from RWTH Aachen University, Aachen,
From 2006 to 2011, he was a Research Assis- Germany, in 1988 and 1996, respectively.
tant with Ruhr University Bochum, where he From 1996 to 2002, he was a Senior Research
was involved in integrated circuits for millimeter- Engineer with the Institute of Communication Sys-
wave (mm-wave) radar applications. In 2011, he tems and Data Processing, RWTH Aachen Univer-
became an Assistant Professor with Ruhr University sity. From April 1998 to March 1999, he was a
Bochum. In 2013, he became the Head of the Technology Consultant at the AT&T Speech and
Department of mm-wave Radar and High Frequency Image Processing Services Research Lab (Shannon
Sensors with the Fraunhofer FHR, Wachtberg, Germany. In 2016, he became Labs), Florham Park, NJ. From April 2002 until October 2003, he was
a Full Professor for Integrated Systems with Ruhr University Bochum. In a Professor of Digital Signal Processing at the Technische Universität
parallel, he is head of the Research group for Integrated Radar Sensors at Braunschweig, Braunschweig, Germany. Since October 2003, he has been a
Fraunhofer FHR. He has authored or coauthored more than 200 scientific Professor of Information Technology and Communication Acoustics at Ruhr-
papers and has issued several patents. His current research interests include Universität Bochum, Bochum, Germany, and from October 2007 to September
ultra-wideband mm-wave radar, design, and optimization of mm-wave inte- 2009, the Dean of the Electrical Engineering and Information Sciences
grated SiGe circuits and system concepts with frequencies up to 500 GHz Department. He is coauthor with P. Vary of Digital Speech Transmission
and above, as well as frequency synthesis and antennas. – Enhancement, Coding and Error Concealment (Wiley, 2006) and coeditor
Prof. Pohl is a member of IEEE, VDE, ITG, EUMA, and URSI. He was a with U. Heute and C. Antweiler of Advances in Digital Speech Transmission
co-recipient of the 2009 EEEfCom Innovation Award, and a recipient of the (Wiley, 2008). His main research interests are signal processing, estimation
Karl-Arnold Award of the North Rhine-Westphalian Academy of Sciences, and machine learning with applications in voice communication systems,
Humanities and the Arts in 2013 and the IEEE MTT Outstanding Young hearing instruments, human–machine interfaces, and sensor networks.
Engineer Award in 2018. Additionally, he was co-recipient of the best paper
award at EUMIC 2012, best demo award at RWW 2015, and best student
paper awards at RadarConf 2020, RWW 2021 and EUMIC 2021.

Authorized licensed use limited to: IMEC. Downloaded on June 23,2023 at 12:15:34 UTC from IEEE Xplore. Restrictions apply.