Track 7: Channel Impulse Responses

8th IPIN Competition off-site Indoor Localization, version 3.0

Sebastian Kram1,2 Maximilian Stahlke1 Christopher Mutschler1

{firstname.lastname}@iis.fraunhofer.de
1
Fraunhofer Institute for Integrated Circuits IIS
2
Friedrich-Alexander-University Erlangen-Nürnberg (FAU)

1 Introduction and Scope

Radio-Frequency (RF) positioning in cluttered indoor environments is challeng-
ing. As signals travel through the environment along different paths it is difficult
to determine the correct time-of-flight (TOF) of the transmitted signals. Tra-
ditionally, fingerprinting-based solutions have been used to estimate a rough
position from narrow-band signals such as Wi-Fi or Bluetooth. However, with
modern ultra-wideband (UWB) technology signals can be transmitted at higher
bandwidths, enabling a much higher spatial resolution from which we can the
extract complex propagation conditions such as absorption, reflection, diffrac-
tion and scattering [1]. While UWB is not yet integrated in consumer devices,
current progress in development and standardization make it likely that they will
be ubiquitous in the near future. This allows for low-cost ad-hoc positioning.
To leverage the benefits of the high spatial solution we can make use of
the channel information (CI). For sufficiently high bandwidths the CI roughly
corresponds to the complex-valued channel impulse response (CIR). Recently,
these signals have been used for positioning in different ways [2]:
• Model Error Mitigation: The CI is used to classify propagation conditions
like non-line-of-sight (NLOS) or to estimate time-of-flight errors caused
by obstructed LOS (OLOS). This enhances classic tracking algorithms
by providing additional information on the channel states [3].
• Fingerprinting: The propagation conditions are assumed to cause signif-
icant differences in the spatial behavior of the CI, which can be exploited
by comparing them with previously recorded data (either using the CI or
extracted features). For Machine Learning (ML) and Deep Learning (DL)
approaches this constitutes a regression task [4].
• Multipath-SLAM : The CI (or extracted multipath components, MPCs) are
used to jointly estimate virtual anchors (i.e. characteristic reflection points
caused by reflecting surfaces) and the trajectory of the transmitter. The

1
 main challenge is to correctly associate the extracted MPCs with specific
surfaces or reflecting objects [5].
Apart from these main concepts, various different or hybrid approaches ex-
ist, each of them with its distinct advantages or disadvantages. We present a
dataset that contains a realistic indoor tracking scenario in an industrial setting
to allow for a fair comparison for practical application.

2 Environment and Measurement Setup

The environment consist of an area of approx 300m2 with a partial enclosure
by reflecting walls (consisting of the walls of the measurement wall, including
metal gates and artifically included reflector/absorber walls elements) and var-
ious metal objects that are typical for industrial indoor environments, like e.g.
industrial vehicles or metal shelves. Fig. 1 schematically sketches the environ-
ment, while Fig. 2 shows a picture of the actual setup: receiving anchors are
placed around the recording area at ∼ 1.5m height. The transmitter device is
carried by a human/worker and regularly transmits UWB signals received by
the anchors. The data is recorded using a platform based on the Decawave
DW1000 UWB chip.

20 m
15 m

recording area reflector walls

Receiver
obj setup 1 obj setup 2

Figure 1: Schematic environment setup, including exemplary object setups 1

and 2.

2
 Figure 2: Image of an environment similar to the one used for the challenge.

The ground truth of the transmitter position is collected using a millimeter-

accurate Qualisys motion tracking system. The data is collected and synchro-
nized by an NTP server and pre-processed (corrupted datapoints are removed
and RF and positioning reference data are synchronized).
This challenge contains two scenarios:

1. For the first scenario we provide training data with ground truth positional
information and the models submitted by the competitors will be evalu-
ated on a test set (a few trajectories) that originate from the same mea-
surement campaign, i.e., training and test data have been recorded on
the same environmental setup. Both training and test data contain com-
plete trajectories while the trajectories of the test data set are shorter.
The test set does not contain ground truth position labels.
2. The second scenario presents a modification of the first scenario. In this
setup we moved clutter elements within the environment (e.g. forklift, van,
etc.) which lead to a slightly different propagation scenario. We will not
provide training data for the modified version of the scenario but only test
data (without ground truth). The goal of this scenario is to test if the mod-
els submitted by the competitors over-fit to the previous environmental
setup and fail to generalize well to changes to the environment.

3
3 Dataset description
3.1 Main Dataset
The main dataset is provided as a HDF 5 file, that can be loaded by various en-
vironments. The main file (i.e. the training data) contains the CI and reference
positions. Each data instance (i.e. column of the dataframe) contains:

• rec time ([float]): the timestamp in s at which the CIR was received
at the receiver node. (This is the ”global time index” of the tracking prob-
lem)
• ci time (array[float]): the timestamps corresponding to the imagi-
nary and real parts of the CI in s. (This is the ”local time index” that can
be used to assign a distance to the CI values)
• burst id ([int]): the transmitter time index. This can be used for syn-
chronization. For clarity, at each of the burst IDs, the transmitter (i.e., the
mobile node) transmits an impulse that is received by a subset of the re-
ceivers (i.e, anchors). The complete set of CIRs from all anchors is not
available at all time steps (as at some receivers the detection was not
successful due to an insufficient channel and/or data corruption).
• ci real (array[int]) and ci imag (array[int]): the real and imagi-
nary parts of the CI as tuples. The CI is centered around the first distinct
peak and contains 366 samples each, which can be set in relation to
distance or time-of-flight by using cir time, as depicted in Fig. 3.

• anch id ([string]): the anchor id of the receiving anchor.

• The positions of the agent (i.e. the mobile tag, the transmitter) ref x,
ref y as float. The reference positions are corresponding to the re-
ceiver timestamp rec time.

An additional . TXT-file (anchors.txt) containing the anchor/receiver posi-

tions is also available. It contains:
• anch ID [string] the anchor IDs.
• p x, p y [float] positions of the anchors.

An additional . TXT-file (hallm easurements.txt)containingthecornercoordinatesof theabsorber/ref lectorw

x1,y1,x2,y2 [float] The corner coordinates of the walls.
description [float] A short description of the type of wall.
Note: We recommend using Python 3.7 or 3.8 and loading the data by using
the pandas library https: // pandas. pydata. org/ docs/ reference/ api/ pandas.
read_ hdf. html , if you have problems with the data format please do not hes-
itate to contact us.

4
Figure 3: Visualization of two exemplary recordings in a LOS and NLOS case:
the time labels of the x-axis are given in ci time, the corresponding mag-
nitudes on the y-axis are given by the complex numbered array defined by
ci real and ci imag
.

4 Challenge Objectives
The challenge is divided into two parts. In the first part the data that is used for
training and testing originate from the same environment setup. In the second
part, we made some changes to the environment setup (i.e., we moved mobile
metallic objects) in order to consider the robustness of the algorithms to envi-
ronmental changes. For the second scenario we do not provide training data
but only test data. The trajectories we use for testing in the second scenario
stay within a similar area as the one used in the first scenario.
For both scenarios, we provide a data frame that includes the same columns
as the training dataset (obviously, apart from the reference positions ref x,
ref y). Additionally, for each setup the initial position, perturbed by artificial
additive zero-mean white Gaussian noise of standard deviation 1m in x and
y-directions, is available. For the first test it is [27.19, 10.42]m, for the second
[29.82, 22.74]m. For each of the files, the competitors are to provide a file with
results. The overall duration of both test datasets is 190 s. Using these data,
the objective is to estimate position values at a regular time interval of 0.5s.
The participants must provide two . HDF- or . CSV files of length 380, labeled
results1 and results2 that include:

• the timestamps t est [float] of the position estimates in s.

• the corresponding estimated positions x est and y est [float].

Results should be provided for times [0.5, 1.0, ..., 189.0, 189.5, 190.0]s.

5
 UPLOAD: Please include the two files into a .zip-file with a name clearly
indicating your team and upload the file to https://owncloud.fraunhofer.
de/index.php/s/eIrdyMTXvLnpfV6. If a team submits multiple files, we will
only evaluate the last file.
Note: For both of these setups, predicted positions need to be provided by
all participants. It is up to the participants if they provide different solutions for
both scenarios.
Note: For clarity, the input data are not available at a perfectly regular sam-
pling interval and the complexity of CIRs does not allow for direct interpolation
to obtain regularly spaced data, so, depending on the applied methods, re-
sampling of the resulting position estimates might be necessary.

5 Exemplary approaches
The objective of the challenge is to use the presented sets of CI to estimate the
position of the tracked object. As mentioned in the introduction, different cate-
gories of positioning algorithms are possible for this task. For clarification, we
included a highly simplified description of a possible pipeline for each category.
Model Error Mitigation: An exemplary tracking pipeline could look like the
one depicted in Fig. 4: A ToF Estimation (Peak Tracking) algorithm is used to
identify the strongest peak in the CI implying the distance between transmitter
and receiver. An error mitigation algorithm, e.g. a machine learning approach,
trained on the available training data is also applied on the CI to estimate an
estimation error describing the difference in estimated an geometric difference
caused by environment interaction. The corrected distance estimates are then
processed in a tracking filter, producing a positioning result.

ToF Estimation / Peak

CI (Test Data)
Estimation

Tracking Filter Positioning Result

Error Mitigation

Training Data

Figure 4: Exemplary Pipeline for a model error mitigation based system.

Fingerprinting: An exemplary positioning approach is sketched in Fig. 5: The

positioning problem is seen as a regression task, where the input consists of

6
the complete CI and the labels are the 2D-positions of the tracking target. For
instance, a deep learning algorithm can be used for this regression task, pro-
ducing positioning estimates, which are then smoothed using e.g. a Kalman
filter.

CI (Test Data) Fingerprinting Smoothing Positioning Result

Training Data

Figure 5: Exemplary Pipeline for a fingerprinting based system.

Channel SLAM: The presented dataset is not ideal for channel SLAM as it can-
not directly benefit from the training data. To mitigate this, we included the
coordinates of the reflector walls in the environment to initialize virtual anchor
hypotheses. A typical pipeline for a channel SLAM is depicted in Fig. 6. Distinct
multipath components (MPCs) are extracted from the CI using a channel esti-
mation algorithm. The channel SLAM algorithm then processes these by data
association with existing virtual anchors (i.e., characteristic reflecting surfaces)
and new virtual anchor hypotheses and uses the associated spatial information
for tracking e.g. in a Rao-Blackwellized particle filter.

CI (Test Data) MPC Extraction Channel SLAM Positioning Result

Figure 6: Exemplary Pipeline for a channel SLAM system.

6 Evaluation metrics
The Euclidean distance between estimated and true results (each 2D-positions)
is the main evaluation metric. Specifically, third quartile is used as a perfor-
mance metric. The results from the known and unknown environment test sets
are weighted equally for performance.

7 Download
You can download the dataset at https://owncloud.fraunhofer.de/index.
php/s/RURgoPDou3PgF3U. Please don’t hesitate contact us for any questions

7
you might have.
Upload: Please include the two files into a .zip-file with a name clearly
indicating your team and upload the file to https://owncloud.fraunhofer.
de/index.php/s/eIrdyMTXvLnpfV6. If a team submits multiple files, we will
only evaluate the last file.

References
[1] A. Molisch, “Ultra-wide-band propagation channels,” Proceedings of the
IEEE, vol. 97, pp. 353 – 371, 03 2009.
[2] S. Aditya, A. F. Molisch, and H. M. Behairy, “A survey on the impact of
multipath on wideband time-of-arrival based localization,” Proceedings of
the IEEE, vol. 106, no. 7, pp. 1183–1203, 2018.
[3] H. Wymeersch, S. Maranò, W. M. Gifford, and M. Z. Win, “A machine learn-
ing approach to ranging error mitigation for uwb localization,” IEEE transac-
tions on communications, vol. 60, no. 6, pp. 1719–1728, 2012.
[4] A. Niitsoo, T. Edelhäußer, and C. Mutschler, “Convolutional neural networks
for position estimation in tdoa-based locating systems,” in Proc. 9th Intl.
Conf. Indoor Positioning and Indoor Navigation, Nantes, France, pp. 1–8,
2018.
[5] C. Gentner, T. Jost, W. Wang, S. Zhang, A. Dammann, and U.-C. Fiebig,
“Multipath assisted positioning with simultaneous localization and map-
ping,” IEEE Transactions on Wireless Communications, vol. 15, pp. 1–1,
09 2016.