Deep Learning for Accurate Indoor Human
Tracking with a mm-Wave Radar
Jacopo Pegoraro
†∗
, Domenico Solimini
‡
, Federico Matteo
‡
,
Enver Bashirov
†
, Francesca Meneghello
†
and Michele Rossi
†
Abstract—We address the use of backscattered mm-wave radio
signals to track humans as they move within indoor environ-
ments. The common approach in the literature leverages the
extended Kalman filter (EKF) method, which however undergoes
a severe performance degradation when the system evolution
model is highly non-linear or presents long-term time dependen-
cies among the system states. In this work, we propose an original
model-free tracking procedure based on denoising autoencoders
and sequence-to-sequence neural networks, showing its superior
performance with respect to state-of-the-art methods. Our ar-
chitecture can be trained in either a supervised or unsupervised
manner, trading tracking accuracy for flexibility. The proposed
system is tested on our own measurements, obtained with a
77 GHz radar on single and multiple subjects simultaneously
moving in an indoor space. The results are compared against
the ground truth trajectories from a motion tracking system,
obtaining average tracking errors as low as 12 cm.
Index Terms—mm-wave radar, indoor sensing, human track-
ing, denoising autoencoders, sequence-to-sequence autoencoders
I. INTRODUCTION
R
ADAR devices for indoor environments are gaining a
growing interest. Recent studies have demonstrated the
possibility of exploiting the properties of the reflected radar
signal to infer the presence, position, and activity of human
targets in indoor spaces [1]–[3]. This approach is a sound
alternative to traditional camera-based sensing systems, as it
preserves the privacy of the users, i.e., no visual representation
of the scene is collected, and is robust to poor light condi-
tions [4]. The use of radio waves in the mm-wave frequency
band allows the estimation of the target distance with high
resolution, in the order of a few centimeters. Although this
increased sensitivity makes millimeter-wave (mm-wave) prone
to disturbances and clutter effects from the radio environment,
the use of data-driven deep learning methods has recently
emerged as a viable solution to these problems, enabling
person identification [2] and activity recognition [3] tasks.
This work has been supported, in part, by MIUR (Italian Ministry of
Education, University and Research) through the initiative ”Departments of
Excellence” (Law 232/2016) and by the EU MSCA ITN project MINTS
“MIllimeter-wave NeTworking and Sensing for Beyond 5G” (grant no.
861222).
† These authors are with the Department of Information Engineering, Uni-
versity of Padova.
‡ These authors are with the Department of Mathematics, University of
Padova.
∗
Corresponding author e-mail: pegoraroja@dei.unipd.it
In the present article, we focus on the problem of tracking
people as they move within an indoor environment, using the
backscattered signal from a mm-wave frequency-modulated
continuous-wave (FMCW) radar. Our aim is to obtain accurate
positioning information of the targets in the physical space. So
far, the few available solutions to this problem [1], [5], have
relied on a Bayesian approach using the extended Kalman filter
(EKF) method [6]. Kalman filter (KF), however, is suitable for
systems that follow a linear evolution model with Gaussian
noise. The extension to the non-linear and non-Gaussian case
(i.e., the EKF or the unscented Kalman filter (UKF), [6]) is
often problematic, especially in highly non-linear models.
In this work, we alternatively use deep neural network (NN)
architectures to sequentially estimate the location of human
targets in indoor spaces: we leverage denoising autoencoders
(DAE) [7] and sequence-to-sequence denoising autoencoders
(S2S) [8] to sequentially learn the best parameters from
the data, not requiring any preliminary assumptions on the
nature of the system evolution, nor on the noise process. S2S
architectures, moreover, are capable of modeling long time
dependencies.
Our main contributions are summarized next.
1) We propose two novel deep learning architectures for the
task of tracking human targets in indoor spaces with a
mm-wave FMCW radar, based on a DAE and a S2S,
respectively. The average tracking error is as low as
0.12 m for the single target case and 0.21 m for the
multi-target one.
2) We evaluate our position tracking system on a challeng-
ing and realistic dataset collected in a room including
furniture, metallic objects, and other people, emulating
real-life conditions.
3) We train the proposed tracking system in supervised
and unsupervised manners. For the former, the ground
truth positions of the targets are provided at training
time, while in the latter only the radar measurements are
used. In both cases, our approach outperforms Bayesian
methods such as EKF and UKF under several metrics.
The rest of the article is organized as follows. In Section II,
the FMCW radar signal model and the detection procedure
are described. In Section III we present the tracking problem
discussing the novelty of our approach. The signal processing
workflow is presented in Section IV. In Section V, experimen-
tal results are discussed, while concluding remarks are given
in Section VI.978-1-7281-8942-0/20/$31.00 ©2020 IEEE
2020 IEEE Radar Conference (RadarConf20)
Authorized licensed use limited to: Nanjing Univ of Post & Telecommunications. Downloaded on February 19,2021 at 16:36:48 UTC from IEEE Xplore. Restrictions apply.