Model-Based Localization and Tracking Using Blueto
Model-Based Localization and Tracking Using Blueto
Abstract: We introduce a high precision localization and tracking method that makes use of cheap
Bluetooth low-energy (BLE) beacons only. We track the position of a moving sensor by integrating
highly unreliable and noisy BLE observations streaming from multiple locations. A novel aspect of our
approach is the development of an observation model, specifically tailored for received signal strength
indicator (RSSI) fingerprints: a combination based on the optimal transport model of Wasserstein
distance. The tracking results of the entire system are compared with alternative baseline estimation
methods, such as nearest neighboring fingerprints and an artificial neural network. Our results show
that highly accurate estimation from noisy Bluetooth data is practically feasible with an observation
model based on Wasserstein distance interpolation combined with the sequential Monte Carlo (SMC)
method for tracking.
Keywords: Bluetooth low-energy localization; hidden Markov model; BLE tracking; observation
probability estimation; Wasserstein distance; Wasserstein interpolation; affine Wasserstein combination;
sequential Monte Carlo
1. Introduction
In this work we address the problem of positioning and tracking using beacons that transmit
data with Bluetooth low-energy (BLE) messages. In the most general setup, we distribute beacons
whose exact positions are not necessarily known in a closed indoor area, and we need to estimate and
track the position of a mobile sensor using BLE messages emitted by the beacons. As the underlying
technology uses short wavelength radio signals, data transfer is affected by a high number of factors,
such as scattering and reflections, all of which can not be exactly known. We attempt to model and
solve this issue using probabilistic methods.
BLE is a distinctive feature that was added to the Bluetooth technology standard at v4.0. It has
been introduced to facilitate short-range communication, which requires large amounts of data transfer.
However, this new feature in Bluetooth technology made it a good candidate as part of positioning
systems as it provides an efficient way of transferring messages with low energy, with the help
battery-powered mobile beacons to emit messages for durations in the order of years [1]. Moreover,
BLE technology provides the variables that can be used to predict the position information. One of
those variables is the received signal strength indicator (RSSI), which may give an estimate of the flight
distance of the signal from the beacon to the receiver [2]. Neburka et al. [3], after extensively analyzing
the performance of BLE technology in indoor environments, show that it is a promising technique for
indoor positioning, even though the RSSI values are inaccurate and highly depend on the BLE module
used. Its new abilities, like durability, mobility, and high reaction time have led to the Bluetooth BLE
technology replacing Wi-Fi for positioning purposes.
An accurate, lightweight, and easily deployable tracking system for indoor environments has
many practical applications. For example, in marketing and retail, a reliable indoor positioning and
tracking system (IPS) deployed in a closed area such as a mall or a shopping center, is used for
targeted advertising, campaign management and customer behavior monitoring [4–6]. In healthcare,
elderly people or patients with dementia can be tracked in indoor environments without needing any
other caretaking personnel or complex sensors [7]. In a factory or production environment, assets or
hardware can be continuously logged for security reasons or inventory analysis [8]. In mobile robotics,
localization is the basis for trajectory planning and smooth and secure navigation. Researchers attempt
to solve this problem with dead-reckoning using inertial sensors accompanied by visual/range-related
environmental sensors like cameras, range scanners, and sonars, or use a combination of these
methods [9,10]. BLE fits into this domain as another practical, simple, and economical sensory system.
There are mainly two basic techniques used in IPS: trilateration and fingerprinting: In trilateration,
estimated distances are used to calculate the most probable coordinates using the geometry of
triangles [11]. With BLE, the distances to the beacons are estimated via the RSSI readings which
are expected to map to the actual distances according to the radio frequency (RF) propagation model [12].
Chen et al. [13,14] combine the pedesterian dead reckoning (PDR) with a weighted path loss (WPL)
algorithm that bases on the log-distance path loss model between a router and a client, under an
extended Kalman filter. While the technique is based on a fast training phase, problems arise due
to the RSSI-to-distance mapping model, which is regarded unstable due to the unreliable distance
estimation in BLE [15,16]. On the other hand, fingerprinting relies on a prior scene analysis (or radio
map construction [17]), in which measurements are collected from the environment and calibrated
into specific features, called fingerprints, which describe a part of the environment [18]. The system
estimates the location of a sensor by comparing the online measurements with the fingerprints, and
the most similar fingerprint location is regarded as an estimate for the expected position. To make
the technique robust and precise, the fingerprint locations must be as dense as possible. Thus, there
arises a trade-off between the installation overhead versus the precision. Besides, the overhead may
also include the recalibration after the system is installed [19,20]. Previous work on IPS shows that
fingerprinting-based techniques outperform the trilateration technique with respect to reliability and
precision [16,21,22].
Different algorithms are employed over the above mentioned IPS techniques. The authors of [23]
study three typical algorithms in IPS domain: the k-nearest neighbors (kNN), neural networks and
support vector machines (SVM). They show that kNN methods yield better results for localization
than the other two. The authors of [24] use fingerprint technique and evaluate BLE localization
performances based on extensive indoor measurements, mapped by propagation modes. The authors
of [25] combine the propagation model (PM) with the extended Kalman filter (EKF), and reach an error
rate of 2.56 m, improving the localization accuracy with sparse beacon deployment. The authors of [26]
compare the fingerprinting technique via Wi-Fi with the current fingerprinting via BLE. They use a
Bayesian estimator as the main method for positioning with a grid size of 1 m, and achieve a tracking
accuracy of 2.6 m using a dense beacon distribution and 4.8 m using a sparse distribution, which,
as they claim, is a significant improvement over Wi-Fi even using sparse beacon deployment [27].
As another technique, Chen et al. [28] study the Bayesian fusion methods using Bluetooth fingerprints
and achieve a horizontal positioning accuracy of 4.7 m on average with respect to Bayesian static
estimation and Kalman filter.
Related literature guides us to use the fingerprinting technique. Nevertheless, we attempt to see
if the trilateration technique may be applied with the beacon set in hand. As previous research [16]
suggests, we ask if we can fit a logarithm-based attenuation model to the RSSI data so that we can map
the data to a distance value easily, but without even attempting to perform a fitting, we can see the
inconsistency in the histograms changing with distance. Figure 1 shows the points that move away
from two different beacons and their corresponding histograms of the RSSI data on each position.
While the expected behavior of the modes would be to move from lower values to higher values as the
Sensors 2017, 17, 2484 3 of 23
point gets farther away from the beacon, we cannot visualize this behavior consistently. The direction
of the blue arrow is selected to be aligned with almost linearly positioned fingerprint locations, so that
we expect a consistent attenuation of strength values of the corresponding beacon data. The histograms
of the blue beacon on Figure 1 do not display the expected behavior. Instead, we see more than one
peaked histograms as we get farther away. The same inconsistency can be visualized with the similar
setting for the red beacon.
As we cannot map the data to some distance metric, the trilateration method does not seem to be
an option in our case. Even when the beacon is in the line-of-sight (LoS) of the sensor, the data values
show interesting fluctuations. Moreover, a single reading, which may be a reflection, would lead to
incorrect distance mappings, but with fingerprinting the same reading can be explained by another
mode of a histogram.
Accordingly in this work, we rely on the fingerprinting technique and we compute the histograms
of RSSI readings on certain positions as the fingerprints (see Figure 2a), but we cannot use the
conventional tracking algorithms like the Kalman filter or its variations as the fingerprint histograms
do not display any Gaussian-like structure, possibly because of the signal reflections from various
objects and surfaces (see Figure 2b). Moreover, we could have discarded the data of the lower strengths
by using sliding windows to arrive at Gaussian-like histograms, but we have the tendency to keep the
multimodality in the histograms, believing that the multimodality has the location information within.
However, for fine localization, we require many fingerprints, the collection of which costs time
and thus is impractical. Vector interpolation techniques are considered to estimate the histograms
associated with arbitrary map positions. Such a technique is applied in the works of [29,30] with Wi-Fi
strength indicators. They approximate the likelihood of an observation via Nadaraya–Watson kernel
regression, which is a computationally complex model, but they achieve accuracies under one meter
with Wi-Fi.
We introduce a histogram interpolation technique, affine Wasserstein histogram interpolation, into
the IPS domain to approximate a radio map using lower number of fingerprints, which is obviously less
time consuming and more practical. With an intelligent interpolation, sparsely distributed fingerprints
are used to estimate the densities at any location. Our radio map model differs from the state-of-the-art
by its tendency to define the interpolated histograms by a combination of the surrounding histograms,
which makes the method totally data driven, and thus the output is expected to keep the original
modes of the histograms (because of scattering and reflections).
Sensors 2017, 17, 2484 4 of 23
(a) (b)
Figure 2. Sample fingerprints and their positions. (a) A dense fingerprinting example in an indoor
environment. Hollow, black and colored dots show measurement positions, sample fingerprint
positions (b) and beacon positions, respectively. (b) Sample fingerprints (histograms) on different
locations for different beacons. Color codes are the same with the beacon dot colors in (a).
The Wasserstein distance is originally a distance metric in optimal transport, defining the distance
between two distributions [31]. While optimally transporting measures between distributions, a
transport map is calculated and the measures are gradually moved according to this map [32,33],
which corresponds to an interpolation between two distributions. The authors of [34] use a similar
regression intuition in computer graphics under the name ‘Wasserstein Barycenter coordinates’.
One of their applications is recolorizing the images with a transition between the original and the
objective color histograms. Likewise, the authors of [35] introduce a class of algorithms for tractable
optimization problems to solve optimal transportation over geometric domains and apply the idea for
shape interpolation, Bidirectional Reflectance Distribution Function (BRDF) design, color histogram
manipulation, skeleton layout, and soft maps.
With a radio map approximation model in hand, we require a tracking method to estimate the
positions. For this purpose, the authors of [36] employ the article filter for indoor localization fusing
Wi-Fi- received signal strengths with accelerometer, gyroscope, map, and barometer data. They show
that map information provides a significant improvement on the positioning accuracy. The authors
of [37] use the particle filter over the iBeacon data. They achieve an error rate as low as 0.27 m using
the trilateration technique.
For the localization purposes, the SMC method can be employed if we can model the motion
and have a location related evaluation [38]. As we have both, we build an SMC method to tackle
the problem of indoor positioning and tracking. However, the main contribution of this work is the
application of Wasserstein interpolation to estimate the observation densities that map the positions to
the beacon RSSI densities. As side contributions, we also add the other regression methods, like the
k-nearest neighbor convex combination and neural networks to estimate the observation densities on
arbitrary map coordinates. The estimated observation densities are then used to in the update step of
the SMC algorithm.
Sensors 2017, 17, 2484 5 of 23
In Section 2, we give the details on the tracking algorithms and the accessories used in these
algorithms. Section 3 describes the test setup and the experiments. We give the experimental results in
Section 4 and finalize our report with a summary and a plan of the future of the research in Section 5.
2. Methodology
We are dealing with an indoor positioning and tracking problem. This section shows the model
of this problem, a hidden Markov model (HMM), followed by a simple transition density model.
We focus primarily on the estimation of the observation densities on arbitrary positions, so that
the likelihood of a single position can be computed. While we contribute to the domain with the
application of Wasserstein interpolation to estimate observation density, many other regression or
interpolation methods are also discussed. With transition and observation densities in hand, we show
how to construct an SMC filter for BLE localization and tracking at the end of the section.
In the most general case, bold uppercase letters (T, H) denote matrices, and bold lowercase letters
(x, y, α) are the vectors. Concordantly, h will always symbolize the histogram vectors. An element
of a vector will be denoted with the unbold version of the original vector and with an index in the
subscript. Index numbering is classically in lower case. Particularly, t always denotes the time points.
The superscript on the histogram (hP ) symbolizes the position it is generated on. If the histogram is
estimated by a specific method, it will be marked with a tilde (∼), the method will be written in the
subscript, and possible hyperparameters will be written as parameters to the method function (h̃NF P ( k )).
The positions are of two types: P, the arbitrary ones, and F, the fingerprint positions. Fingerprint
positions will be indexed in the subscript as they belong to a finite set. p( x |y, Θ) denote the conditional
probability distributions, and Θ are the problem specific hyperparameters that, in this work, depend
on the estimation algorithms.
Dtb ∼ p( Dtb | Pt , Θ)
P0 ∼ p( P0 ) (1)
Pt ∼ p( Pt | Pt−1 )
For the transition density, p( Pt | Pt−1 ), we use the motion model given in Section 2.2. In Section 2.3,
we construct the estimators for the observation density via multiple methods that use the fingerprints.
Sensors 2017, 17, 2484 6 of 23
With the estimated observation densities we compute the likelihood of a new observation sample,
p( Dtb | Pt , Θ). The two models are then combined in a tracking filter described in Section 2.4.
xt , yt
We assume that the robot is not fed by any internal sensory data that enables dead reckoning.
The diffusion motion model relies on the assumption that the robot will be in a close position to a
previous one, which is thus modeled as a Gaussian distribution with the previous position as the mean
value. The covariance of the distribution determines how far the robot can move in a time unit [10].
The next state, xt , can be modeled by:
! ! !
x x ex
= + (2)
y y ey
t t −1
where ei = N (0, R). We assume that distributions on individual dimensions are independent and
their variances are equal (R).
We use the corresponding histogram on the closest fingerprint location, hFi∗ , as the likelihood:
P
p( Db | P, Θ) = h̃NF = hFi∗ (4)
This method assumes that the area in which the position estimation will be handled is densely
sampled, so that for every position P, there happens to be a close fingerprint position.
Sensors 2017, 17, 2484 7 of 23
F3
d3
P
d1
F1 d2
F2
Figure 5. Visualization of a convex combination at the position P in the convex hull of some other
fingerprint positions.
Let Fj denote the fingerprint positions on which the histograms are previously computed, hFj ,
and let the position P reside in the affine hull of a subset of fingerprint positions. Then the histogram
on P, hP = {hi }, is defined as:
Fj
hi = ∑ λ j hi (5)
j
where λ = {λ j } are multipliers that are tuned with respect to the distances between the estimation
position P and the fingerprint positions Fj , and necessarily ∑ j λ j = 1. In particular, we use the softmax
exp(−k P− Fj k)
function to boost the effect of the close fingerprint positions: λ j = ∑k exp(−k P− Fk k)
. The observation
density of a specific position on the map can then be estimated as
P
p( Db | P, Θ) = h̃kNF (k) (6)
where k is a parameter that decides how many fingerprint positions closest to the estimation position
will be used. Note that the method NF is the special case of kNF with k = 1.
H→ h0 h1 ... hj ... hH
V→
+1
Z→ z0 z1 ... zi ... zN
W→
+1
X→ x0 x1
Layers are fully connected via nonlinear activation functions (hyperbolic tangent and sigmoid) of
the linear perceptron equations. We also add a bias for each node to make the structure learn an offset
value if possible. The forward neural equations in matrix notation are given in (7).
Z = tanh ([X 1] W)
(7)
H = sigm ([Z 1] V)
P
p( Db | P, Θ) = h̃ANN ( Nh ) (8)
one dimensional vector, y = {yi }. In our case, these vectors are actually normalized histograms, whose
bins are identical and each bin index corresponds to an integer between the given extremities.
x = { xi } 0 ≤ xi ≤ 1 ∑ x =1
where and i i
y = { yi } 0 ≤ yi ≤ 1 ∑i yi = 1
In the discrete case [41], the Wasserstein distance is the minimum cost of transporting one
histogram onto another:
where τi,j is the measure that would be transported from the location i to the location j and c(i, j) is the
cost multiplier that is usually related to the distance between the locations i, j.
If the cost function, c, is a linear function, the Wasserstein transport matrix is efficiently computed
by scanning the array and keeping track of the transported weights between bins.
Wasserstein Interpolation
We can perform the interpolation operation in multiple ways (Figure 7). In fact, we define two
parameters, α and β, both in the interval of [0, 1] [33]. The actual interpolation between densities is
controlled by the coefficient α, which decides how similar would the interpolation be to the original
densities. For the values of α near zero, the interpolation will be similar to the source histogram,
whereas for the higher values will make the interpolation resemble the destination histogram.
β controls the evolution of the interpolaton between a linear interpolation (β = 0) and a displacement
interpolation (β = 1) [33].
Figure 7. Interpolation ambiguity: We want to move the pile of red histogram onto the green histogram.
The figure on the left represents a direct jump of the measures onto the new location, whereas the one
on the right shows a gradual movement of the measures towards the new pile location.
Formally, given two histograms, namely h0 and h1 , we first compute the Wasserstein mapping,
Th0 ,h1 , which is merely a matrix that demonstrates the transport plan of the measures. According to
this transport plan, we calculate two intermediate plans, I0α,β = {ι0i,j } and I00α,β = {ι00i,j }, which represent
the destination indices. The measures, τi,j will be distributed to:
ι0i,j = i + dαβ( j − i )e
(9)
ι00i,j = j − d(1 − α) β( j − i )e
where d.e is the ceil operator as the indices should be integers. Indices of the final interpolation are
computed as follows:
Note that a two-position convex combination (kNF) of histograms is merely a special case of the
Wasserstein combination, in which β is set to 0.
|1 − α |
hk = ∑ ∑ |1 − α| + |α| τi,j δ(ι0i,j = k)
i j
(11)
|α|
+ τ δ(ι00 = k )
|1 − α| + |α| i,j i,j
The algorithm is given in Algorithm 1 and the combinations for different values of α and β can be
seen in Figure 8.
Figure 8. Affine combination of two histograms of Gaussian densities for different α and β values.
Yellow vertical lines show the original location of the modes.
Sensors 2017, 17, 2484 11 of 23
d⊥ (h Fi , Fj i, P) ≤ ρ
where d⊥ defines the orthogonal distance of a point to a line, h Fi , Fj i is the set of affine combinations or
simply the line passing through the points Fi and Fj , and ρ is the looseness parameter (See Figure 9).
ρ
P
Fj
Fi
d
gne
ali
Figure 9. A position P is loosely aligned with positions Fi and Fj if it falls in the green area.
Finally, if there exist predefined histograms at positions Fi and Fj and if P is aligned loosely with
these positions with respect to a previously defined ρ, we can apply an affine Wasserstein combination
of histograms between hFi and hFj at P.
P
p( Db | P, Θ) = h̃AWC-B (α, β) (12)
According to this extension, for the position of interest, P, we select all of the possible fingerprint
pairs that can generate a combination according to the selection model. For each pair we perform the
pairwise combination and obtain a histogram estimation on P with each pair. The final combination,
h̃AWC-E , is found by taking the affine combination of the available two-position affine Wasserstein
combinations, h̃AWC-B , and averaging them with the weights, λi , inverse exponentially proportional to
the pairwise average distances to P. We will denote the final histogram with h̃AWC-E :
P
p( Db | P, Θ) = h̃AWC-E = ∑ exp(−λi )h̃AWC-B
P
(13)
i
3. Setup
Figure 10. Bluetooth dongle on a tripod and beacons mounted on the walls.
Figure 11. Distribution of fingerprint locations with different sizes. Black dots: measurement locations,
colored dots: beacon locations.
As the complexity of the algorithms are also similar, we do not mention their complexities.
Only ANN takes too long to train, while it passes an actual training procedure compared to the others.
Moreover, we use a fine grid structured probabilistic radio map, on which histograms for every beacon
every grid center are previously computed, so that for the actual runs the likelihood estimation of the
particles is merely a lookup table operation with the trade-off of high memory usage.
Sensors 2017, 17, 2484 15 of 23
Observational data are sampled afterwards on the trajectory points. The details of the data
generation is given in Appendix B.
Sampling of
Estimation of
Synthetic
Trajectories with
Trajectories
SMC Filter
and BLE Data
(f) (g)
In a different track, (f) we sample a trajectory of a virtual agent that navigates in the test area
and generate RSSI data from the previously collected dense fingerprint locations. Finally, using the
estimate of the probabilistic radio map as the observation density, sampled RSSI data on the trajectory
points as the observations and the motion model as the prediction density, (g) we employ the SMC
filter to estimate the position estimations for the observations and measure the tracking performance
of the specific method.
Sensors 2017, 17, 2484 16 of 23
Figure 14. Running sequential Monte Carlo (SMC) filter: The yellow trajectory shows the
computer-generated positions, while the cyan trajectory shows the estimation.
The SMC filter was run for the same trajectory of 232 points, with 32 of them discarded as the
burn-in period, leaving us with 200 points of error for each run. We log the distances of the predicted
positions to the original positions as errors. Each combination (estimator and fingerprint set size) is
repeated 30 times, which gives 6000 position estimations or error measurements per combination.
The box plot of the statistics is given in Figure 15. We report both mean and median errors.
Figure 15. Trajectory estimation performances of different observation models with different sizes of
fingerprint sets with black lines denoting medians and red lines for means. NF: nearest fingerprint;
kNF: k-nearest fingerprint; ANN: artificial neural network; AWC-B: affine Wasserstein combination
with the best two positions; AWC-E: affine Wasserstein combination with multiple positions.
It can be seen from the results that high-resolution fingerprint information with 50 fingerprint
locations (FL) defines the map the best for the tracking purposes with any of the applied methods.
Without employing any complicated method, that is, using only NF in estimating the observation
Sensors 2017, 17, 2484 17 of 23
density, gives the best result: an error rate of 0.66 m, with the lowest variance. This is an expected
result but with an impractical setting. With more than 50 fingerprint locations, this error rate will
surely get lower, but making the scene analysis much more of a burden at the same time. Collecting
data from many locations is obviously not practical. A reduced number of fingerprints would facilitate
the installation and calibration procedure in BLE positioning and tracking infrastructure, so that the
results for such scenarios are substantial and realistic for us.
As the configurations lose a number of fingerprint locations, we see that NF is unable to keep
its success with the highest resolution, and Wasserstein interpolation-based techniques stand out.
Even though the error rates are over 1 m, AWC-B looks especially promising with an error rate of
1.29 m with 15 fingerprint locations. The same algorithm gives an error rate of 1.9 m with only
8 fingerprint locations.
The neural network approach (ANN) is both inconvenient as we do not have many samples
for training and impractical as the training iterations take too long. Moreover, the results show that
ANN yields consistently higher error rates compared to the other methods. These results are expected
whilst the neural networks require many samples beforehand, and we also try to reduce the required
fingerprint size, which makes ANN naturally inappropriate for our purposes.
Table 1 summarizes the mean localization errors for different fingerprint confugurations with
respect to the applied estimator algorithms. Concentrating on the results with lower number of
fingerprints (15 and 8), we see that Wasserstein-based methods race head to head with the kNF method.
We also run two t-tests for significance analysis on the last two configurations. In the significance test
tables we report the p-values for the one-sided hypothesis if the value on the row is less than the value
on the column. According to the p-values, for 15 fingerprint locations (see Table 2), AWC-B has the
best localization performance, but for the configuration of 8 fingerprint locations (see Table 3), there is
no best estimator, because AWC-B and kNF cannot outperform each other.
Table 2. p-values of the pairwise one-sided t-tests of the results for the set of 15 FLs.
Table 3. p-values of the pairwise one-sided t-tests of the results for the set of 8 FLs.
Even though the box plots in Figure 15 give a hint about error distributions, we also supply the
explicit distributions and corresponding cumulative error distributions. In Figure 16, we encode the
methods with the same color codes with Figure 15.
Figure 16. Localization error distributions and corresponding cumulative error distributions for two
fingerprint location configurations: 15 and 8 locations.
The figures clearly show that the error distributions are not normally distributed, which is
probably due to sudden incorrect far position estimations in the SMC filter. We believe that such results
are better shown with box plots and medians, as the errors have skewed distributions, where means
would be misleading. We also report the error medians in Table 4.
For the significance test, because the error results are not necessarily normally distributed,
we apply one-sided Wilcoxon signed-rank tests on the error rates of the configuration pairs. We report
the two low sized fingerprint sets’ values with the confidence value of α < 0.05.
Table 5 shows that for the configuration of 15 fingerprint locations, the proposed methods
AWC-B and AWC-E, perform significantly better than any other methods we compared. For a smaller
configuration of eight locations, we see almost the same results. Table 6 shows that the proposed
methods perform significantly better than the applied techniques, except against kNF, in which we can
say that our methods perform slightly better than kNF. This leads to the claim that, for the small-sized
fingerprint sets (8 and 15), Wasserstein-based interpolation techniques reduce the errors significantly,
(except for only one case) compared to the applied techniques so far. Moreover, as the complexity of the
methods is similar, Wasserstein interpolation techniques are preferable for small-sized fingerprint sets.
Amongst the proposed methods, AWC-B is seen to perform slightly better than its counterpart AWC-E.
Sensors 2017, 17, 2484 19 of 23
Table 5. p-values of the pairwise one-sided Wilcoxon tests of the results for the set of 15 FLs.
Table 6. p-Values of the pairwise one-sided Wilcoxon tests of the results for the set of 8 FLs.
Thus, we conclude that that two methods AWC-B and AWC-E are the two candidates to be used
in observation density estimation in the SMC filter for tracking purposes with BLE beacon information
for lower number of fingerprints if the positions of these fingerprints are scattered evenly in order to
perform better histogram interpolations.
5. Conclusions
In this work, we have developed a method to render indoor localization and tracking practical
using only BLE sensors. Our model-based approach relies critically on the accurate estimation
of a probabilistic radio map—a distribution of RSSI values for every position inside the region of
interest—from a few RSSI fingerprint measurements, obtained only at a few locations. The estimated
radio map is subsequently used as the observation model of a dynamical system where we do target
tracking by a sequential Monte Carlo algorithm.
Not surprisingly, there is a direct relationship between the accuracy of the radiomap and accuracy
of localization and tracking. We show first that when RSSI fingerprints are collected on a very dense
grid, a radio map can be accurately estimated and a satisfactory tracking performance can be obtained
despite the high variability of the actual measurements. However, dense fingerprint sampling is often
not practical or at least not desired in real applications as this requires careful measurement and a long
data collection process. In this paper, we develop methods to obtain accurate radio maps from far fewer
and sparsely sampled fingerprints and show empirically that tracking accuracy is still acceptable.
We cast the radio map estimation as a fingerprint interpolation method, that we cast as a regression
problem. We have proposed two variations of the Wasserstein interpolation method, which is also
originally derived from the Wasserstein distance computation method. The first one uses two points as
the pivot points to find an affine combination on an unknown coordinate aligned with the pivot points.
The second one fuses multiple affine combinations to estimate a histogram on an unknown coordinate.
The results of these Wasserstein variations are compared with the results of the well known regression
methods, namely the nearest neighbor approach, the affine combination and the neural networks
We can generalize from the results that we can estimate the unknown radio information on the
arbitrary positions by an interpolation using the radio strength densities on previously measured
positions. Moreover, Wasserstein combination variations are good candidates for histogram estimation
purposes, as the error rates of these metods increase consistently while the number of fingerprints are
reduced, but they perform better compared to the other applied techniques.
Sensors 2017, 17, 2484 20 of 23
An expected result is that with such small data, neural networks are unable to learn the nonlinear
histogram on a planar surface without overlearning. The domain can benefit from the autoregressive
models to train the histograms against the positions.
In fingerprinting techniques, initial scene analysis is a must, but it is not practical to collect data
from high number of locations. For future work, an immediate study could to perform an analysis on
the appropriate fingerprint locations. Having the practicality in mind, the researcher can also find the
number of fingerprint locations with respect to the area size and the required fineness. Additionally,
if a method can give confidence for a position in the area, it may urge the researcher to gather data
from locations with low confidence values. A proper Gaussian process regression for multiple outputs
would be a good future study using this data, as the algorithm by nature provides the confidence
values.
Author Contributions: F. Serhan Daniş and Ali Taylan Cemgil conceived and designed the algorithms and the
experiments. F. Serhan Daniş collected the data, performed experiments and analyzed the data. Both authors
wrote and revised the document.
Conflicts of Interest: The authors declare no conflict of interest.
τ0 = 0
x0 , y0 ∼ U (A)
θ0 ∼ U ([0, 2π ])
δ̃tτ ∼ G( aτ , bτ )
δ̃tl ∼ G( al , bl )
(A1)
δ̃tθ ∼ N (0, s)
τt = τt−1 + δ̃tτ
xt = xt−1 + δ̃tl cos(θt−1 + δ̃tθ )
yt = yt−1 + δ̃tl sin(θt−1 + δ̃tθ )
θt = θt−1 + δ̃tθ
We sample an initial location in A and an initial orientation between 0 and 2π, and set the
initial time stamp to 0. At each step, a time interval, δ̃tτ , and a distance value, δ̃tl , are sampled from
gamma distributions, and a rotation value, δ̃tθ , is sampled from a normal distribution. The parameters
of the distributions, aτ , bτ , al , bl , s are experimentally set to generate values similar with the real
data. To prevent our virtual robot from leaving the area, sampled rotation values are deliberately
manipulated by an additional value of π8 according to the current orientation. This helps the robot
return smoothly into the area when it approaches the borders. Sample trajectories can be seen
in Figure 12.
bt ∼ M(b, Pt )
i∗ = arg min d( Pt , Fi ) (A2)
i
Fi∗
Dt ∼ M(hbt )
Sensors 2017, 17, 2484 21 of 23
We first sample a beacon label from the set of all available beacons b, according to a probability
distribution Pt , which reshapes at each time step inversely proportional to the use counts of the
beacons, in order to let the beacons be sampled uniformly in small time windows. We then find the
closest fingerprint index (i∗ ), and sample an RSSI value, Dt , using the corresponding beacon histogram.
References
1. Cinefra, N. An Adaptive Indoor Positioning System Based on Bluetooth Low Energy RSSI. Master’s Thesis,
Politecnico di Milano, Milano, Italy, 24 April 2013.
2. Cabarkapa, D.; Grujic, I.; Pavlovic, P. Comparative Analysis of the Bluetooth Low-Energy Indoor Positioning
systems. In Proceedings of the 12th International Conference on Telecommunication in Modern Satellite,
Cable and Broadcasting Services (TELSIKS), Nis, Serbia, 14–17 October 2015; pp. 76–79.
3. Neburka, J.; Tlamsa, Z.; Benes, V.; Polak, L.; Kaller, O.; Bolecek, L.; Sebesta, J.; Kratochvil, T. Study of the
Performance of RSSI Based Bluetooth Smart Indoor Positioning. In Proceedings of the 26th International
Conference Radioelektronika (RADIOELEKTRONIKA), Kosice, Slovakia, 19–20 April 2016; pp. 121–125.
4. Li, F.; Zhao, C.; Ding, G.; Gong, J.; Liu, C.; Zhao, F. A Reliable and Accurate Indoor Localization Method
Using Phone Inertial Sensors. In Proceedings of the 2012 ACM Conference on Ubiquitous Computing, UbiComp ’12;
ACM: New York, NY, USA, 2012; pp. 421–430.
5. Kuo, Y.S.; Pannuto, P.; Hsiao, K.J.; Dutta, P. Luxapose: Indoor Positioning with Mobile Phones and
Visible Light. In Proceedings of the 20th Annual International Conference on Mobile Computing and Networking,
MobiCom ’14; ACM: New York, NY, USA, 2014; pp. 447–458.
6. Dickinson, P.; Cielniak, G.; Szymanezyk, O.; Mannion, M. Indoor Positioning of Shoppers Using a Network
of Bluetooth Low Energy Beacons. In Proceedings of the International Conference on Indoor Positioning
and Indoor Navigation (IPIN), Alcala de Henares, Madrid, Spain, 4–7 October 2016; pp. 1–8.
7. Komai, K.; Fujimoto, M.; Arakawa, Y.; Suwa, H.; Kashimoto, Y.; Yasumoto, K. Elderly Person Monitoring
in Day Care Center Using Bluetooth Low Energy. In Proceedings of the 10th International Symposium on
Medical Information and Communication Technology (ISMICT), Worcester, MA, USA, 20–23 March 2016;
pp. 1–5.
8. Bisio, I.; Sciarrone, A.; Zappatore, S. A new asset tracking architecture integrating RFID, Bluetooth Low
Energy tags and ad hoc smartphone applications. Pervasive Mob. Comput. 2016, 31, 79–93.
9. Siegwart, R.; Nourbakhsh, I.R. Introduction to Autonomous Mobile Robots; Bradford Company: Scituate, MA,
USA, 2004.
10. Thrun, S.; Burgard, W.; Fox, D. Probabilistic Robotics (Intelligent Robotics and Autonomous Agents); The MIT
Press: Cambridge, MA, USA, 2005.
11. Martin, P.; Ho, B.J.; Grupen, N.; Muñoz, S.; Srivastava, M. An iBeacon Primer for Indoor Localization:
Demo Abstract. In Proceedings of the 1st ACM Conference on Embedded Systems for Energy-Efficient Buildings,
BuildSys ’14; ACM: New York, NY, USA, 2014; pp. 190–191.
12. Boukerche, A.; Oliveira, H.A.B.F.; Nakamura, E.F.; Loureiro, A.A.F. Localization systems for wireless sensor
networks. IEEE Wirel. Commun. 2007, 14, 6–12.
13. Chen, Z.; Zou, H.; Jiang, H.; Zhu, Q.; Soh, Y.C.; Xie, L. Fusion of WiFi, Smartphone Sensors and Landmarks
Using the Kalman Filter for Indoor Localization. Sensors 2015, 15, 715–732.
14. Chen, Z.; Zhu, Q.; Soh, Y.C. Smartphone Inertial Sensor-Based Indoor Localization and Tracking With iBeacon
Corrections. IEEE Trans. Ind. Inform. 2016, 12, 1540–1549.
15. Mahiddin, N.A.; Safie, N.; Nadia, E.; Safei, S.; Fadzli, E. Indoor Position Detection Using WiFi and
Trilateration Technique. In Proceedings of the International Conference on Informatics and Applications
(ICIA2012), Kuala Terengganu, Malaysia, 2–5 June 2012; pp. 362–366.
16. Dahlgren, E.; Mahmood, H. Evaluation of Indoor Positioning Based on Bluetooth Smart technology.
Master’s Thesis, Chalmers University of Technology, Gothenburg, Sweden, 10 June 2014.
17. Muñoz-Salinas, R.; Marín-Jiménez, M.J.; Yeguas-Bolivar, E.; Carnicer, R.M. Mapping and Localization from
Planar Markers. arXiv 2016, arXiv:abs/1606.00151.
18. Pritt, N. Indoor Location with Wi-Fi Fingerprinting. In Proceedings of the Applied Imagery Pattern
Recognition Workshop (AIPR), Washington, DC, USA, 23–25 October 2013; pp. 1–8.
Sensors 2017, 17, 2484 22 of 23
19. Liu, H.; Darabi, H.; Banerjee, P.; Liu, J. Survey of Wireless Indoor Positioning Techniques and Systems.
Trans. Syst. Man Cybern. Part C 2007, 37, 1067–1080.
20. Farid, Z.; Nordin, R.; Ismail, M. Recent Advances in Wireless Indoor Localization Techniques and System.
J. Comput. Netw. Commun. 2013, 2013, 12.
21. Toplan, E. RFID Based Indoor Location Tracking. Master’s Thesis, Bogazici University, Istanbul, Turkey, 2011.
22. Bisio, I.; Lavagetto, F.; Marchese, M.; Sciarrone, A. Smart probabilistic fingerprinting for WiFi-based indoor
positioning with mobile devices. Pervasive Mob. Comput. 2016, 31, 107–123.
23. Zhang, L.; Liu, X.; Song, J.; Gurrin, C.; Zhu, Z. A Comprehensive Study of Bluetooth Fingerprinting-Based
Algorithms for Localization. In Proceedings of the 27th International Conference on Advanced Information
Networking and Applications Workshops (WAINA), Barcelona, Spain, 25–28 March 2013; pp. 300–305.
24. Pelant, J.; Tlamsa, Z.; Benes, V.; Polak, L.; Kaller, O.; Bolecek, L.; Kufa, J.; Sebesta, J.; Kratochvil, T.
BLE device indoor localization based on RSS fingerprinting mapped by propagation modes. In Proceedings
of the 27th International Conference Radioelektronika (RADIOELEKTRONIKA), Brno, Czech Republic,
19–20 April 2017; pp. 1–5.
25. Zhuang, Y.; Yang, J.; Li, Y.; Qi, L.; El-Sheimy, N. Smartphone-Based Indoor Localization with Bluetooth Low
Energy Beacons. Sensors 2016, 16, 596.
26. Faragher, R.; Harle, R. An analysis of the accuracy of bluetooth low energy for indoor positioning applications.
In Proceedings of the 27th International Technical Meeting of The Satellite Division of the Institute of
Navigation (ION GNSS+ 2014), Tampa, FL, USA, 8–12 September 2014; Volume 812.
27. Faragher, R.; Harle, R. Location Fingerprinting With Bluetooth Low Energy Beacons. IEEE J. Sel.
Areas Commun. 2015, 33, 2418–2428.
28. Chen, L.; Pei, L.; Kuusniemi, H.; Chen, Y.; Kröger, T.; Chen, R. Bayesian Fusion for Indoor Positioning Using
Bluetooth Fingerprints. Wirel. Pers. Commun. 2013, 70, 1735–1745.
29. Zhang, V.Y.; Wong, A.K.S. Kernel-based Particle Filtering for Indoor Tracking in WLANs. J. Netw.
Comput. Appl. 2012, 35, 1807–1817.
30. Zhang, V.Y.; Wong, A.K.S.; Woo, K.T. Histogram Based Particle Filtering with Online Adaptation for Indoor
Tracking in WLANs. Int. J. Wirel. Inf. Netw. 2012, 19, 239–253.
31. Villani, C., The Wasserstein distances. In Optimal Transport: Old and New; Springer: Berlin/Heidelberg,
Germany, 2009; pp. 93–111.
32. Benamou, J.D.; Brenier, Y. A computational fluid mechanics solution to the Monge-Kantorovich mass transfer
problem. Numer. Math. 2000, 84, 375–393.
33. Papadakis, N.; Peyré, G.; Oudet, E. Optimal Transport with Proximal Splitting. SIAM J. Imaging Sci. 2014,
7, 212–238.
34. Bonneel, N.; Peyré, G.; Cuturi, M. Wasserstein Barycentric Coordinates: Histogram Regression Using
Optimal Transport. ACM Trans. Graph. 2016, 35. doi:10.1145/2897824.2925918.
35. Solomon, J.; de Goes, F.; Peyré, G.; Cuturi, M.; Butscher, A.; Nguyen, A.; Du, T.; Guibas, L. Convolutional
Wasserstein Distances: Efficient Optimal Transportation on Geometric Domains. ACM Trans. Graph. 2015,
34, 66.
36. Nurminen, H.; Ristimäki, A.; Ali-Löytty, S.; Piché, R. Particle Filter and Smoother for Indoor Localization.
In Proceedings of the International Conference on Indoor Positioning and Indoor Navigation,
Montbeliard-Belfort, France, 28–31 October 2013; pp. 1–10.
37. Zafari, F.; Papapanagiotou, I. Enhancing iBeacon Based Micro-Location with Particle Filtering. In Proceedings
of the IEEE Global Communications Conference (GLOBECOM), San Diego, CA, USA, 6–10 December 2015;
pp. 1–7.
38. Gustafsson, F.; Gunnarsson, F.; Bergman, N.; Forssell, U.; Jansson, J.; Karlsson, R.; Nordlund, P.J.
Particle filters for positioning, navigation, and tracking. IEEE Trans. Signal Process. 2002, 50, 425–437.
39. Chollet, F. Keras. 2015. Available online: https://github.com/fchollet/keras (accessed on 27 July 2017).
40. Tieleman, T.; Hinton, G. Lecture 6.5-rmsprop: Divide the gradient by a running average of its recent
magnitude. In COURSERA: Neural Networks for Machine Learning; University of Toronto: Toronto, ON,
Canada, 2012.
41. Ferradans, S.; Papadakis, N.; Rabin, J.; Peyré, G.; Aujol, J.F. Regularized Discrete Optimal Transport.
In SSVM 2013—International Conference on Scale Space and Variational Methods in Computer Vision; Kuijper, A.,
Bredies, K., Pock, T., Bischof, H., Eds.; Springer: Leibnitz, Austria, 2013; Volume 7893, pp. 428–439.
Sensors 2017, 17, 2484 23 of 23
42. Gordon, N.; Salmond, D.; Smith, A. Novel approach to nonlinear/non-Gaussian Bayesian state estimation.
IEEE Proc. F Radar Signal Process. 1993, 140, 107–113.
43. Doucet, A.; de Freitas, N.; Gordon, N. An Introduction to Sequential Monte Carlo Methods. In Sequential
Monte Carlo Methods in Practice; Springer: New York, NY, USA, 2001; pp. 3–14.
44. Alpaydin, E. Introduction to Machine Learning, 2nd ed.; The MIT Press: Cambridge, MA, USA, 2010.
45. Arlot, S.; Celisse, A. A survey of cross-validation procedures for model selection. Statist. Surv. 2010, 4, 40–79.
c 2017 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).