IEEE Signal Processing - May 2018 - 68

Gr
ou
nd
-Tr
uth
Pa
th

ried in the participant's hand; the trajectory is inevitably affected by the human body's movement during walking, while the
distance estimator actually calculates the integration of speed
over all moving directions. When the distance is shorter, the
impact of this kind of error can be magnified. Nevertheless,
when the walking distance is large, the impact of the uncontrollable errors on the distance estimation is insignificant.
We also evaluated the statistical distribution of the TRITS
location error. In the experiment, a participant walked along a
path in the same office environment, denoted by the green line
in Figure 11. The person started from point A and stopped at
different path locations, which are marked by green double circles. The distances from starting point A to each end point were
5, 11, 21, 25, 30, 40, 55, 64, and 69 m. For each path length,
we repeated the experiment 25 times. The cumulative distribution function (CDF) of the location estimation error for all of
the paths is shown in Figure 12. The median of the estimation
error is around 0.33 m, and the 80th percentile is around 1 m.
Therefore, the TRITS can track a moving object in a complex
indoor environment at submeter accuracy.

1
Path
ted
a
ath 2
m
Esti
ated P
Estim

FIGURE 9. An illustration of map-based position correction.

Estimated Distance (m)

12
10
8

Wireless event detection

6
4
2
2

6
8
10
Actual Distance (m)

FIGURE 10. The results of walking distance estimation, with the four topto-bottom lines in each block representing the fifth, 25th, 75th, and 95th
percentiles of the estimated distance.

Inspired by the TR-based indoor positioning that associates a
location with a CSI fingerprint, we can also utilize the TR
technique to capture the variations in the multipath CSI due to
different indoor events, such as opening or closing a door or
window. By treating each path of the multipath channel in a
rich-scattering environment as a distributed virtual antenna, a
TR-based indoor event detection system (TRIEDS) can be
designed that takes the multipath CSI as the feature and determines the occurrence of an indoor event according to the current CSI in the propagation environment.

CSI as a feature for event detection
the positioning accuracy will be degraded by accumulated
error. Nevertheless, interior structures, such as doors, walls,
and corridors in the floor plan, can be embedded in the TRITS
and work as constraints and landmarks to alleviate the error.
For example, Figure 9 shows a T-shaped corridor with two
possible estimated paths. Compared to the ground-truth trajectory denoted by the dashed line, estimated paths 1 and 2 will
penetrate a wall. In these two cases, a reasoning procedure
can be performed that adjusts each path so that its estimated
trajectory can be fitted to the floor plan and all of the boundary constraints imposed by the floor plan can be satisfied. In
this way, the accumulated error from both the distance and the
direction estimation can be corrected.

The TRIEDS exploits the intrinsic property of the TR technique, i.e., capturing and fusing the information of the multipath propagation environment in the spatial resonance. To detect
the occurrence of an indoor event by TR spatial resonances, the
TRIEDS first learns through an offline training phase and then
works in an online testing phase. Specifically, in the training
phase, a training database is built by collecting the CSI, e.g., the
time-domain CIR h T , from each indoor event through channel
probing. After that, in the online testing phase, the TRIEDS first
estimates the instantaneous multipath CSI h T, test from the current indoor environment and then makes the detection based on
the one in the offline training database that generates the greatest strength of the generated spatial resonance.

Experiments

Phase 1: Offline training

The first set of experiments was to estimate the moving distance of a human walking inside an office. One experiment
participant carried the transmitter and walked, in turn, 2, 4, 6,
8, 10, and 12 m. For each distance, we repeated the experiment
20 times along different paths; the walking speed did not need
to be constant. The results are shown in Figure 10. The error in
distance estimation is mainly due to the transmitter being car-

For each indoor event S i ! D, with D being the event set,
the corresponding CSI is obtained through channel probing and forms a matrix H T, i as H T, i = [h T, i (t 0), h T, i (t 0), f,
h T, i (t N -1)], where N is the size of the CSI samples for a training event. The vector h T, i (t j) represents the estimated CSI vector of event S i at time t j, and H T, i becomes the CSI matrix for
event S i .

IEEE Signal Processing Magazine

May 2018

Table of Contents for the Digital Edition of IEEE Signal Processing - May 2018