Download as pdf or txt
Download as pdf or txt
You are on page 1of 6

Large-Scale Indoor Positioning using Geomagnetic

Field with Deep Neural Networks


Han Jun Bae and Lynn Choi
School of Electrical Engineering
Korea University
Seoul, Korea
{qwerty2901, lchoi}@korea.ac.kr

Abstract— The existing RF signal based indoor localization build IPSs with higher positioning accuracy, the infrastructure
techniques such as BLE or Wi-Fi fingerprinting are hard to apply system to generate and receive such signals is too expensive to
to large scale indoor environment such as airport and department be used with smartphones or IoT devices. Furthermore, they
stores since the localization error grows as the physical dimension often require higher power consumption [8, 9, 10] or range
of the indoor space increases. This can be attributed to unstable limitation [8]. Therefore, we need a fundamentally different
received signal strengths (RSS) of the underlying RF signal, which approach to build high accuracy yet cost effective IPS.
is enlarged with the increased physical scale and the complexity of An alternative signal that can be used for IPS is the
the indoor space. In this paper, instead of RF signal we use the geomagnetic field signal generated by the earth movement. The
geomagnetic sensor signal for indoor localization, whose signal geomagnetic field signal has several distinct advantages over
strength is more stable than RF RSS. Our approach using the other signals. First, IPS based on the geomagnetic field does not
geomagnetic field is as follows. Although similar geomagnetic field require infrastructure such as access points or beacons since the
values exist in indoor space, an object movement would experience geomagnetic field is created by nature unlike RF, ultrasonic, or
a unique sequence of the geomagnetic field signals as the laser signals. Second, most of smart IoT devices and
movement continues. We can locate the position of the object by smartphones already include IMU (Inertial Measurement Unit)
tracking the geomagnetic field signal sequence sensed with the sensors, a.k.a. motion sensors, which include the magnetic field
object movement by using a deep neural network model called sensor in addition to the accelerometers and gyroscope sensors.
recurrent neural network (RNN), which is good at recognizing Thus, it is easy to sense the geomagnetic field signal with these
time varying sequence of sensor data. We use two different smart devices. These two characteristics enable inexpensive IPS
versions of RNN model: basic RNN and Long Short-Term implementation based on the geomagnetic field. Third, the
Memory (LSTM). We have trained RNNs to learn the magnetic signal strength of the geomagnetic field is much more stable
field maps of both medium scale (about 94m x 26m) and large scale over time than those of RF or ultrasonic signals, which would
(about 608m x 50m area) indoor testbeds and analyze both potentially enable higher accuracy IPS [11].
training and test set results by tuning several training
hyperparameters. For comparison, we have also implemented Despite these advantages, it is not easy to achieve high
both Bluetooth Low Energy (BLE) and Wi-Fi based fingerprinting positioning accuracy even with the geomagnetic field signal in
localization techniques and measured their localization accuracies the existing IPSs. Most of the existing geomagnetic field based
for the testbeds. By using Google TensorFlow 1.6 and Nvdia CUDA IPSs rely on fingerprinting with K-NN (K-Nearest Neighbor)
Toolkit v9.0 with cuDNN v7.1 library as a deep learning mapping, which is borrowed from RF fingerprinting.
framework, we could achieve the average localization accuracy of Fingerprinting based IPS has the problem of sensing the same
0.51 and 1.04 meters for the medium and the large-scale testbeds received signal strength in multiple indoor locations. Although
respectively with LSTM model, substantially improving the RF-based fingerprinting is also vulnerable to the distribution of
localization performance compared to the existing RF based the same received signal strength in multiple indoor locations, it
fingerprinting techniques. can alleviate the problem with multiple SSIDs since an RF signal
can reach a certain range depending on its transmission power.
Keywords—indoor localization; geomagnetic field; recurrent As opposed to the RF-based fingerprinting [2, 3], techniques
neural network; machine learning, long short-term memory model; using the geomagnetic field vector could not resolve the problem
of differentiating the same geomagnetic field value registered at
I. INTRODUCTION
multiple locations, which could potentially degrade the
In the past few decades, a diverse range of indoor performance of the geomagnetic field based indoor localization
localization schemes have been proposed. These schemes can be system.
classified by the type of sensor signals used, which includes LF,
RF, acoustic, IR, light, laser and etc. Among these the indoor In this work we propose a fundamentally different approach
to build the geomagnetic field based IPS. We first construct a
positioning system (IPS) based on RF signals such as Bluetooth geomagnetic field map of an indoor space by collecting the
or Wi-Fi is the most common since such RF technology is
widely adopted by smart devices and signals are often available geomagnetic field vectors at reference points as done by the
existing geomagnetic field based IPSs. However, instead of K-
in indoor environment. The RF based IPSs may measure the NN mapping we employ machine learning techniques based on
received signal strength (RSS) [1, 2, 3], time of arrival (ToA) [4]
or time difference of arrival (TDoA) [5, 6] for the positioning. artificial neural networks. We train a neural network model to
memorize all feasible movement paths and their corresponding
However, ToA and TDoA are not common due to their large geomagnetic field vector sequences of a moving object in the
clock synchronization overhead compared to the fast speed of
RF signal propagation. Therefore, most of RF based IPSs rely space. Then, we can locate the position of a moving object by
using the sequence of consecutive geomagnetic field vectors
on sensing RSS. collected during the movement as an input to the neural network.
However, it is not easy to build high accuracy IPS even with Although a single geomagnetic field vector can lead to multiple
RF’s RSS. It is because of the time varying and unstable nature locations of the same value, a sequence of the geomagnetic field
of RF signal, which is caused by multi-path propagation, vectors can uniquely identify the current position of the object
diffraction, and scattering of RF signals incurred by the indoor as the sequence grows longer.
structures such as walls, furniture, humans and elevators. Over
In order to memorize the contiguously changing patterns of
10 to 20 meters of positioning errors are often reported in large- geomagnetic vector sequences, we use a recurrent neural
scale indoor environment such as airports and department stores
[7]. Although LIDAR, UWB or ultrasonic signals can be used to network (RNN) model which is good for recognizing the time-

978-1-5386-8088-9/19/$31.00 ©2019 IEEE


varying sequence of data. The reason why RNN is good for and by almost a factor of four the large-scale testbed. In addition,
recognizing the time-sequence data can be seen from its both RNN models substantially improve the localization
architecture. RNN cannot only process the current input data but performance of the existing RF-based localization schemes,
it can also remember the past sequence of input data and use which could achieve the positioning errors of 3.56 meters and
them to yield a different result from its recurrent network model. 7.26 meters for the two testbeds even with their extra
In our earlier work [11] we have shown the preliminary results infrastructure HW.
of our geomagnetic field based IPS implementation for a small- The rest of this paper is organized as follows. Section II
scale testbed by using a simple basic RNN model. surveys recent studies related to our work. Section III presents
However, the basic RNN has the vanishing gradient problem. RNN models and the optimization of hyperparameters for our
In [12], the vanishing gradient problem was analyzed in detail. proposed IPS. Section IV presents the generation of the training
Whenever the gradient of the error function of the neural and test data sets for the RNN models including the construction
network is propagated back through a node of a neural network, of the geomagnetic field maps for our testbeds. Section V
it gets scaled by a certain factor. For nearly all practically discusses the experimental results of the proposed IPS for both
relevant cases, this factor is either greater than one or smaller training and inference by varying the experimental parameters
than one. As a result, in the basic RNN model the gradient blows such as the number of hidden nodes and the step count. We also
up or decays exponentially over time. Thus, the gradient either compare the localization performance of LSTM model against
dominates the next weight adaptation step or effectively gets lost. that of the basic RNN model as well as those of BLE and Wi-Fi
To avoid this scaling effect, we employ a more advanced version fingerprinting for the testbeds. Finally, Section VI concludes the
of RNN model called long short-term memory (LSTM) model paper.
[13], addressing the vanishing gradient problem of basic RNN.
To evaluate our geomagnetic-field IPS based on our RNN II. RELATED WORKS
models for large-scale indoor environment, we use two testbeds: Positioning algorithms are generally classified into range-
a medium-scale campus testbed called Hana Square at Korea based and range-free algorithms depending on whether they
University and a large-scale off-campus testbed at the Incheon rely on distance measurement or not. The range-based
International Airport. The dimension of the medium-scale algorithms are further classified into trilateration and
testbed is 94.4 meters by 26 meters while the dimension of the triangularization techniques. Multilateration techniques
large-scale testbed is 608.6 meters by 49.3 meters, which is including trilateration usually employ ToA or TDoA to locate
about 12 times larger than the medium-scale testbed. We the position while triangularization techniques use AoA (Angle
construct geomagnetic field maps for both testbeds where the of Arrival). While these are popular in outdoor environment,
smaller one has 3177 reference points and the larger one has they are usually not used in indoor space either due to their high
20088 reference points, each of which are spaced every about clock synchronization overhead or due to line of sight (LOS)
60cm interval. requirement.
For comparative evaluation, we have also implemented the A. Indoor localization using RF RSS
existing RF-based localization schemes. We use BLE Instead of conventional trilateration or triangularization
fingerprinting for the medium-scale testbed while we apply Wi- techniques based on ToA, TDoA, or AoA, signal strength
Fi fingerprinting for the large-scale testbed since its scale attenuation model [23] can be used to measure the distance in
requires too many BLE beacons yet public Wi-Fi signals are indoor space. SpotOn [1] is a kind of RF-based IPS that use RF
available at the airport. For the RF-based fingerprinting, we also signal strength. A central server in this system collects the RF
construct radio maps for both testbeds which have the same signals to locate the precise position of an object by using an
number of reference points as the geomagnetic field maps. aggregation algorithm based on trilateration.
For the geomagnetic-field IPS, we use two RNN models, a More popular approach for indoor localization is RF
basic RNN model and LSTM model. From the geomagnetic fingerprinting based on Wi-Fi or BLE signal strengths. This is
field maps, we generate millions of traces that simulate various an example of range-free positioning algorithm since the
pedestrian movement paths. 60% of the traces are used for positioning no longer depends on distance measurement. The
training, 20% are used for validation, and the remaining 20% are fingerprinting method was first introduced in RADAR [2] where
used for evaluation. With supervised training, we construct a they first collect Wi-Fi RSSs at pre-determined reference points.
basic RNN model and an LSTM model for the testbeds by Then, they measure the RSSs of an unknown location and
varying the number of hidden nodes and the number of hidden compare them with the previously collected RSSs to determine
layers through repeated experiments. We also vary several the position. RADAR employs K-Nearest Neighbor (K-NN)
training hyperparameters such as activation functions and algorithm for the positioning, which predicts the current location
optimization algorithms for the network parameter update to by calculating the weighted sum of points of the closely
generate the models. For the RF-based IPS, we use K-NN matching K reference points.
mapping which estimates the position from the RF RSS vector
sensed at the current location by computing the weighted sum of B. Indoor localization using geomagnetic field RSS
the first K closely matching reference points. A few recent studies [14, 21] proposed geomagnetic field
We compare the localization performance of the LSTM based IPSs for pedestrians [14] and robots [21] that use the
model against that of the basic RNN model as well as that of the patterns of a geomagnetic field values as fingerprints for
existing RF fingerprinting techniques based on BLE and Wi-Fi. locations. They survey the magnetic field pattern of a
From the evaluation, with LSTM we could achieve average predetermined path and estimate the position of an object by
positioning errors of 0.80 meters and 2.59 meters for medium calculating similarity between the stored magnetic field pattern
and large scale testbeds. However, this includes start-up errors and the measured magnetic field pattern by using the maximum
since the first few steps may not be enough for LSTM to likelihood function [14, 21]. However, both of previous works
uniquely identify the magnetic field vector sequence. Excluding focused on the geomagnetic field pattern of a single main path
the start-up errors for the first few steps, we could achieve and may not cover the complexity of all possible movement
average positioning errors of 0.51 meters and 1.04 meters for the paths of an object as the size of an indoor space increases.
two testbeds. In contrast, the basic RNN model archives the
average positioning errors of 1.20 meters and 4.10 meters for C. Indoor localization using artificial neural networks
those two testbeds. By addressing the vanishing gradient Several existing studies [15, 16] have applied artificial neural
problem, LSTM can remember the geomagnetic vector networks (ANN) to RF-based IPS. In [15] they use a simple 3-
sequence deeper in its memory and can outperform the basic layer feedforward neural network with one input layer, one
RNN by more than a factor of two for the medium-scale testbed hidden layer and one output layer. With RSS values of a single
location as an input and 2-D coordinate value as output, they
demonstrate that they can outperform the localization time t as the output of the LSTM model. The 2-dimensional
performance of the existing RF-based localization scheme coordinate output represents the position of a moving object in
which simply computes the nearest neighbor instead of K-NN. the 2-dimensional space.
CiFi [16] uses a deep convolutional neural network (DCNN) to
process AoA images. They estimate AoA values from channel B. Configuration of training hyperparameters
state information (CSI), which are extracted from 960 Wi-Fi There are several hyperparameters which should be tuned
packets. Then, they construct 16 images from 960 AoA values. for the neural network training. They include those which
Then, AoA images are used as inputs to the DCNN to estimate specify the structure of the network such as the number of
the location. They could achieve better localization accuracy hidden nodes and the number of hidden layers, as well as those
compared to three existing localization schemes based on CSI. which determine how the network is trained such as mini-batch
However, both of the existing ANN based IPS studies [15, 16] size, the choice of activation function for hidden and output
using RF signals show localization errors of only about 3 meters layer, the selection of the loss function, the choice of an
on average even in very small testbeds with less than or equal to optimizer for backpropagation and learning rate for an
100 meter square dimension. optimizer, etc. After setting these values, we should set the
maximum number of epochs for the entire training. One epoch
III. RNN MODELS FOR THE PROPOSED IPS means that the training is performed once for the entire data set.
A. Architecting ANN for IPS based on the geomagnetic field In our experimentation, we set the maximum number of epochs
to 500 while we vary the number of hidden nodes from 300 to
There are many types of artificial neural networks including 600 for medium-scale testbed and from 900 to 1200 for large-
feedforward neural networks (FNN), convolutional neural scale testbed. We also vary the number of hidden layers from 1
networks (CNN) and recurrent neural networks (RNN). Among to 3 for both testbeds.
these networks, we choose RNN. RNN uses present and past Batch normalization is generally used for more effective
input data to deduce an output. Because of this feature, RNN is
good at recognizing continuous time-varying sequence data such ANN training. Training ANN with a single input data item often
leads to convergence problems while using the entire data set
as voice, audio and video data stream. Our indoor localization as a single batch for training would make ANN forget earlier
also needs to recognize continuous time-varying sequence data.
There are several variations of RNNs such as basic RNN, bi- data’s features. To prevent this problem, we use mini-batch for
training. Since too large or small batch size could lead to the
directional RNN, GRU [17] and LSTM [13]. Our previous study same problem, we have come up with appropriate batch sizes
[11] used basic RNN, but in this paper we employ LSTM. It is
due to the vanishing gradient problem that is associated with the through several training sessions.
basic RNN. As discussed in Section I, the gradient in the basic Activation functions are used for modifying the output of a
RNN model blows up or decays exponentially over time during hidden or an output node and the selection of the activation
back-propagation. This causes RNN to lose information from function can affect the training results. We select hyperbolic
the past input data as the depth of the RNN layers increases, tangent as the activation function for the hidden nodes since it
substantially degrading its train performance. LSTM solve this can generally alleviate gradient explosion problem of RNN
problem by adding the cell state. model. In contrast, for the activation function of output nodes
we choose the ReLU function since our neural network’s outputs
are greater than or equal to 0.
After input values are fed forward through the neural
network, we need to compare the outputs of the neural network
with the desired outputs, i.e. labels for supervised training. For
comparing outputs with labels, we need to select a loss function.
Because our outputs are 2D coordinates, we need a loss function
which is similar to the Euclidean distance between outputs and
labels. For this purpose, the mean squared error shown in the
equation below is chosen for the loss function.
𝑛
Figure 1. LSTM architecture 1
Mean Squared Error = ∑(output 𝑖 − label𝑖 )2
Figure 1 shows the LSTM cell architecture and how each 𝑛
𝑖=1
gate in the cell is connected with other gates and how each cell
is connected with cells in the previous and the next layer. Given The optimizer is an update method of neural network
an input sequence x = (x1, . . . , xT ), LSTM computes the hidden parameters such as weights and biases during back-propagation.
vector sequence h = (h1, . . . , hT ) and output vector sequence y After calculating the loss function, the optimizer determines
= (y1, . . . , yT ) by iterating the following equations from t = 1 to how to update the network parameters by using several
T: hyperparameters such as the gradient of the loss function,
learning rate, and momentum. The learning rate determines how
it = σ ( Wxi xt + Whi ht-1 + bi ) much of the gradient should be applied to the ANN weight
ft = σ ( Wxf xt + Whf ht-1 + bf ) adjustment while the momentum is used to speed up the training
ot = σ ( Wxo xt + Who ht-1 + bo ) by accumulating the previous update history. There are several
optimizers such as stochastic gradient descent (SGD), Adam
ct = ft ⊙ct-1 + it ⊙ tanh ( Wxc xt + Whc ht-1 + bc ) [18], ADADELTA [19], and Adaptive subgradient method [20],
ht = ot ⊙ tanh(ct) in which they differ by how they reflect those hyperparameters
yt = A ( Why ht + by ) on the network parameter update. Among these optimizers, we
use Adam optimizer for our training, which is one of the best
where W terms denote weight matrices (e.g. Wxh is the input- optimizers known to be quite effective. In our experimentation,
hidden weight matrix), b terms denote bias vectors (e.g. b h is it also shows the best training results compared to other
hidden bias vector), σ is the logistic sigmoid function, ⊙ denotes optimizers.
the inner product operation, A denotes the activation function of
the output layer, and i, f, o and c are respectively the input gate, IV. METHODOLOGY
forget gate, output gate and cell activation vectors, all of which
are the same size as the hidden vector h. A. Experimental testbeds
In order to use LSTM for our geomagnetic field based IPS, We use two testbeds, Hana Square Basement Level 1 Floor
we use the 3-dimensional geomagnetic field vector x1t, x2t, x3t at of Korea University Science and Engineering campus and the
time t as the input and the 2-dimensional coordinate y1t, y2t at 1st floor (Gate 9 to 14) of Incheon International Airport. The
physical dimension of Hana Square testbed is 94.4 meters by 26 geomagnetic field vectors and its associated coordinate values
meters while that of Incheon Airport testbed is 608.6 meters by for supervised training. The geomagnetic field vector sequence
49.3 meters. We call Hana Square testbed as the medium-scale is used as an input sequence for RNN while the sequence of the
testbed and Incheon Airport testbed as the large-scale testbed. corresponding coordinates is used for the desired output for
Figure 2 and 3 show the geomagnetic field maps of our testbeds. RNN supervised training. Each of the path data set is generated
The large-scale testbed is approximately 12 times larger than from the geomagnetic field map as follows:
the medium-scale testbed. Both testbeds include some empty Step 1: Select a random start point among the coordinates in
spaces, which are actually outdoor or closed space shown in the magnetic field map.
dark purple color. This makes both testbeds more difficult to
learn by a neural network model because the model should Step 2: We choose a direction from that start point. Since an
recognize these blank areas as empty spaces for its localization. object can move any of 360 degree direction, we select a
random direction out of 360 degree options in order to give
B. Measuring the geomagnetic field vectors diversity of directions.
In order to measure the geomagnetic field values, we use the Step 3: Go a random distance to that direction. If the path
IMU sensor of a Samsung Galaxy S8 smartphone. The IMU reaches a wall or a blank area that cannot be entered, at this
sensor integrates 3 different sensors: accelerometer, gyroscope, point, go to Step 2
and magnetic sensor, each of which measures its sensor value Step 4: If the path length reaches the maximum, i.e. 100
in the 3-dimensional X, Y, and Z axes. Using the magnetic field coordinate movement, go to the first step and start again until
sensor, we can measure the 3-dimensional components of the the number of data sets reaches our demand.
geomagnetic field vector. Because the geomagnetic field vector
depends on the smartphone orientation, if we turn the This path generation actually follows a simplified version of
smartphone, the axis is changed, which makes the geomagnetic a mobility model called the random way point model [22].
field values change. It can be fixed by using the orientation Through this path generation process, we generate 20 million
value. The orientation value shows us how much the coordinate movements for Hana Square testbed and 120 million
smartphone has been rotated by the absolute 3-dimensional coordinate movements for Incheon Airport testbed. A single
cardinal dimension, namely north, east, south, west, up and movement path consists of 100 coordinate movements and it is
down. This orientation value can be estimated by the the unit of training and test data. Therefore, there are 200,000
smartphone’s IMU sensor. data sets for Hana Square testbed and 1,200,000 data sets for
Incheon Airport testbed. 60% of these data sets are used for
training, 20% of the data sets are used for validation and the rest
20% are used for test.
E. Radio map construction and RF fingerprinting
For comparative evaluation, we also implement and test RF
fingerprinting techniques. We use BLE RSS for Hana Square
testbed and Wi-Fi RSS for Incheon Airport testbed. For BLE
fingerprinting we deploy 30 BLE beacons in the Hana Square
Figure 2. Geomagnetic field map of medium-scale (Hana Square) testbed testbed while we use the existing Wi-Fi access points available
C. Construction of the geomagnetic field maps at the airport since the testbed is too large to cover with our BLE
beacons. First, we collect RF RSSs at pre-determined reference
Similar to radio map construction in RF-based points which are same with those of the geomagnetic field maps
fingerprinting, we need to construct the geomagnetic field maps discussed in Section Ⅳ.D. Then, we measure the RSSs using the
for our testbeds. However, geomagnetic field maps are easier to sensors of the smartphone and compare them with the previously
construct than radio maps. This can be attributed to the collected RSSs to determine the position. We calculate Euclidian
following reasons. First, we only need to collect one distance between RSSI vectors to find a reference point with the
geomagnetic field signal instead of multiple beacon or AP closest RSSI vector. We K-NN algorithm for the positioning,
signals. Second, the geomagnetic field signal is more stable which predicts the current location by calculating the weighted
over time compared to unstable time-varying RF signals. Thus, sum of points of the closely matching K reference points.
signal sampling and calibration is generally faster. Hana Square Table 1. System Environment
testbed has 3177 reference points that are spaced every 60.5cm Parameters Tools
interval, while Incheon Airport testbed has 20088 reference CPU Intel i7-6900K 3.2GHz
points spaced every 57.5cm interval excluding the blank area. GPU NVIDIA GeForce GTX1080 Ti
Using the smartphone’s IMU sensor, we collect the RAM DDR3 32GB
geomagnetic field vector signals at the reference points and also OS Ubuntu Server 16.04 LTS
records the physical coordinates of the reference points. Figure Language Python 3.5
2 and 3 illustrate the resultant geomagnetic field maps of our Google TensorFlow 1.6
testbeds using contour lines. Note that the sizes of two maps Library Nvidia CUDA Toolkit v9.0
shown in the figure do not reflect the relative sizes of the actual Nvidia cuDNN v7.1
maps. Incheon Airport map should be drawn much larger V. EVALUATION
compared to Hana Square map.
A. HW and SW system framework for deep learning
D. Path data generation for basic RNN and LSTM
We use Google TensorFlow 1.6 as our deep learning
To train our RNN models, we should generate a large framework for RNN models. To accelerate RNN training, we
number of path data sets, each of which is a sequence of the

Figure 3. Geomagnetic field map of large-scale (Incheon Airport) testbed.


also use NVDIA GPU with its CUDA Toolkit v9.0 and cuDNN Table 3. Test set result
v7.1 library. Table 1 shows hardware and software system Basic RNN LSTM
environments and deep learning frameworks used for our 1 layer 3 layers 1 layer 3 layers
experimentation. 300 600 300 600 300 600 300 600
Max 28.08 30.51 28.81 28.12 61.42 68.83 10.67 36.01
B. Results Hana
(m)
Min 1.471 0.888 0.745 0.611 0.556 0.301 0.337 0.182
Table 2 summarizes RNN training hyperparameter settings squar
(m)
e
used for our testbeds. Avg. 3.590 2.216 1.860 1.625 1.581 1.350 1.189 0.795
(m)
Table 2. RNN hyperparameters
900 1200 900 1200 900 1200 900 1200
Incheon Max 185.6 110.3 168.8 101.6 247.7 248.7 91.59 130.1
Type Hana Square
Airport Inche (m)
Hidden nodes 300, 600 900,1200 on Min 3.372 3.209 4.083 2.208 1.718 1.085 0.873 0.657
Airp (m)
Hidden layers 1,3
ort Avg. 15.62 8.008 9.836 5.510 4.642 4.518 3.136 2.594
Batch size 300 (m)
Learning rate 0.0005 0.0001
Loss function Mean squared error
optimizer Adam
Activation function tanh/ReLU
Train : validation : test 60% : 20% : 20%
Epochs 500
Path length 100

Figure 6. CDF of the test errors of Hana Square and Incheon Airport
Table 3 shows the test set results of two testbeds. With 600
hidden nodes and 3 hidden layers, we could achieve the average
localization error of 0.80 meters and the maximum error of 36.01
meters for LSTM model for Hana Square testbed. In contrast,
we achieve the average localization error of 2.59 meters and the
maximum error of 130.1 meters for Incheon Airport testbed with
Figure 4. Training errors of basic RNN (RNN) and LSTM models in 1200 hidden nodes and 3 hidden layers. This is due to the fact
the Hana Square testbed. Figure legend shows the model, the number of that the amount of information that LSTM should learn in
hidden nodes, and the number of hidden layers. Incheon Airport testbed is much larger than that in Hana Square
testbed. Even though LSTM was trained by 6 times more
training data, there is some loss of information in the LSTM
model, which leads to higher localization errors.
Figure 6 shows the cumulative distribution function (CDF)
of the test errors for two testbeds. From now on we assume 600
hidden nodes and 3 hidden layers for the Hana Square testbed
and 1200 hidden nodes and 3 hidden layers for the Incheon
Airport testbed. As shown in the figure, 97.5% of the test data
has errors of less than 2 meters for the Hana Square testbed and
97.2 % of the test data has errors of less than 5.0 meters for the
Incheon Airport testbed.

Figure 5. Training errors of basic RNN (RNN) and LSTM models in


the Incheon Airport testbed. Figure legend shows the model, the number
of hidden nodes, and the number of hidden layers.
Figure 4 and Figure 5 show the training error by varying the
number of hidden nodes and the number of hidden layers with
basic RNN and LSTM models. As we increase the number of
hidden nodes and the number of hidden layers, we could achieve
higher localization accuracy for both testbeds. A larger number
of hidden nodes are required for the large-scale Incheon Airport
testbed since much larger size of the testbed requires a larger
network capacity. Also, even if we assume the same number of Figure 7. Average localization errors in terms of step count
hidden nodes or the same number of hidden layers, LSTM
models generally converge to much smaller training errors than Figure 7 shows the average localization errors of test data
basic RNN models since LSTM can memorize longer history as sets in terms of step count for two testbeds. As shown in the
we have discussed in Section Ⅰ. Overall, we could achieve the figure, localization errors of the first few steps are much higher
final training errors of 0.58 meters for the medium-scale testbed than the remaining steps. As we have discussed in Section I,
with 600 hidden nodes and 1.92 meters for the large-scale RNN models use the past input sequence to predict the output.
testbed with 1200 hidden nodes on average assuming 3 hidden This causes RNN models to take time to produce the correct
layers for LSTM model. output initially. Therefore, after the first 10 steps for Hana
Square testbed and after the first 20 steps for Incheon Airport
testbed the localization errors converge to a small value for both
testbeds. Since we calculate localization error in every ACKNOWLEDGEMENT
coordinate of a path rather than the final coordinate, these start- This work was supported by the National Research Found
up errors contribute to the larger localization errors. ation of Korea(NRF) grant funded by the Korea government
Table 4. Test set result without the start-up errors (MSIP) (NRF- 2017R1A2B2009641) and by the MSIT(Ministry
Basic RNN LSTM of Science and ICT), Korea, under the ITRC(Information
1 layer 3 layers 1 layer 3 layers Technology Research Center) support program (IITP-2018-
300 600 300 600 300 600 300 600 2015-0-00363) supervised by the IITP (Institute for Information
Max 26.89 29.38 26.96 27.89 60.23 65.70 8.816 35.79 & communications Technology Promotion).
(m)
Hana
squar
Min 1.314 0.793 0.691 0.568 0.399 0.206 0.283 0.139 REFERENCES
(m)
e [1] Hightower, Jeffrey, Roy Want, and Gaetano Borriello. "SpotON: An indoor
Avg. 2.858 1.678 1.429 1.197 0.874 0.712 0.624 0.512
(m) 3D location sensing technology based on RF signal strength." (2000).
900 1200 900 1200 900 1200 900 1200 [2] Bahl, P., and Padmanabhan, V. N., “RADAR: An in-building RF-based user
Max location and tracking system,” in Proceedings of the IEEE INFOCOM, Vol. 2,
183.6 109.1 129.5 92.5 224.7 242.5 84.49 129.6
Inche (m) 2000, pp. 775-784
on Min
2.486 2.719 2.371 1.948 1.111 0.705 0.612 0.477
[3] An, J. H., and Choi, L., “Inverse fingerprinting: server side indoor localization
Airpo (m) with Bluetooth low energy,” in IEEE PIMRC, 2016, pp. 1-6.
rt Avg. [4] Chen, Xuechen, Shupeng Song, and Jihong Xing. "A ToA/IMU indoor
9.801 5.754 4.901 4.104 2.316 2.442 1.445 1.040
(m) positioning system by extended Kalman filter, particle filter and MAP
algorithms." Personal, Indoor, and Mobile Radio Communications (PIMRC),
Table 4 shows the localization performance of RNN models 2016 IEEE 27th Annual International Symposium on. IEEE, 2016.
with and without the start-up errors. By excluding the data for [5] Harter, A., Hopper, A., Steggles, P., Ward, A., & Webster, P. (2002). The
anatomy of a context-aware application. Wireless Networks, 8(2/3), 187-197.
the first 10 steps for the Hana Square testbed and the first 20
[6] Priyantha, N. B., Miu, A. K., Balakrishnan, H., & Teller, S. (2001, July). The
steps for the Incheon Airport testbed, we could substantially cricket compass for context-aware mobile applications. In Proceedings of the
reduce the average localization error from 0.58 to 0.51 meters 7th annual international conference on Mobile computing and networking (pp.
for the Hana Square testbed and from 1.92 to 1.04 meters for the 1-14). ACM.
Incheon Airport testbed, achieving one or sub-meter level of [7] Cheng, Y. C., Chawathe, Y., LaMarca, A., and Krumm, J., “Accuracy
localization accuracy for both testbeds with LSTM model. characterization for metropolitan-scale Wi-Fi localization,” in Proceedings of
the 3rd ACM MOBISYS, 2005, pp. 233-245.
C. Comparison against RF fingerprinting techniques [8] Alarifi, A., Al-Salman, A., Alsaleh, M., Alnafessah, A., Al-Hadhrami, S., Al-
We also test the localization performance of RF Ammar, M. A., & Al-Khalifa, H. S. (2016). Ultra wideband indoor positioning
fingerprinting techniques for both testbeds. We use K = 4 for technologies: Analysis and recent advances. Sensors, 16(5), 707.
the K-NN mapping. With BLE fingerprinting, we could achieve [9] Chen, Y., Liu, J., Jaakkola, A., Hyyppa, J., Chen, L., Hyyppa, H., ... & Chen,
R. (2014, May). Knowledge-based indoor positioning based on LiDAR aided
the average, the minimum, and the maximum localization errors multiple sensors system for UGVs. In Position, Location and Navigation
of 3.56 meters, 0.71 meters and 6.81 meters for the medium- Symposium-PLANS 2014, 2014 IEEE/ION (pp. 109-114). IEEE.
scale testbed. For the large-scale testbed, we achieved the [10] Ijaz, F., Yang, H. K., Ahmad, A. W., & Lee, C. (2013, January). Indoor
average, the minimum, and the maximum localization errors of positioning: A review of indoor ultrasonic positioning systems. In Advanced
7.26 meters, 1.78 meters and 13.64 meters with Wi-Fi Communication Technology (ICACT), 2013 15th International Conference on
fingerprinting. Compared to these RF fingerprinting, our (pp. 1146-1150). IEEE.
proposed geomagnetic field IPS based on LSTM model [11] Jang, H. J., Shin, J. M., and Choi, L., “ Geomagnetic Field Based Indoor
achieved the average, the minimum, and the maximum errors of Localization Using Recurrent Neural Network,” in IEEE GLOBECOM, 2017,
pp. 1-6
0.51 meters, 0.14 meters, and 35.8 meters for the medium-scale
[12] Hochreiter, Sepp. "The vanishing gradient problem during learning recurrent
testbed and 1.04 meters, 0.48 meters, and 129.6 meters for the neural nets and problem solutions." International Journal of Uncertainty,
large-scale testbed. Considering only the average positioning Fuzziness and Knowledge-Based Systems 6.02 (1998): 107-116.
errors, our geomagnetic field IPS shows quite impressive results [13] Gers, Felix A., Jürgen Schmidhuber, and Fred Cummins. "Learning to forget:
since we can reduce the localization errors by a factor of 7 Continual prediction with LSTM." (1999): 850-855.
without using extra infrastructure such as access points or [14] Shahidi, S., and Valaee, S., “GIPSy: geomagnetic indoor positioning system
beacons. However, in terms of the maximum localization errors, for smartphones,” in IEEE Indoor Positioning and Indoor Navigation, 2015,
RF fingerprinting shows much better results against our pp. 1-7.
proposed IPS. This is due to the fact that APs or beacons can act [15] Tuncer, S., and Tuncer, T. “Indoor localization with bluetooth technology
using artificial neural networks,” in IEEE Intelligent Engineering Systems,
as a ranging device in the fingerprinting techniques since a
2015, pp. 213-217
smartphone user cannot receive BLE or Wi-Fi signals if it is out [16] Wang, X., Wang, X., and Mao, S., “CiFi: Deep convolutional neural networks
of reception range, whereas the magnetic field signal can be for indoor localization with 5 GHz Wi-Fi,” in IEEE Communications, 2017,
measured at all area of the map. In our future work, we will pp. 1-6.
investigate how we can exploit available RF signals such as [17] Cho, K., Van Merriënboer, B., Gulcehre, C., Bahdanau, D., Bougares, F.,
public Wi-Fi signals by using them as ranging model to reduce Schwenk, H., and Bengio, Y. “Learning phrase representations using RNN
the maximum localization errors of our geomagnetic field IPS. encoder-decoder for statistical machine translation.” arXiv preprint
arXiv:1406.1078, 2014
VI. CONCLUSION [18] Kingma, Diederik, and Jimmy Ba. "Adam: A method for stochastic
optimization." arXiv preprint arXiv:1412.6980, 2014.
In this paper we present the geomagnetic field based indoor [19] Zeiler, Matthew D. "ADADELTA: an adaptive learning rate method." arXiv
positioning scheme using the RNN models of deep learning preprint arXiv:1212.5701, 2012.
technology. By training LSTM to memorize the geomagnetic [20] Duchi, J., Hazan, E., and Singer, Y., "Adaptive subgradient methods for online
field maps of two indoor testbeds, we could achieve the learning and stochastic optimization," in Journal of Machine Learning
localization accuracy of 0.51 meters for the medium-scale Research, 2011, pp. 2121-2159.
testbed and 1.04 meters for the large-scale testbed respectively, [21] Storms, William, Jeremiah Shockley, and John Raquet. "Magnetic field
excluding the start-up initial errors. We show that our LSTM navigation in an indoor environment." Ubiquitous Positioning Indoor
based IPS using the geomagnetic field signal substantially Navigation and Location Based Service (UPINLBS), 2010. IEEE, 2010.
outperforms the basic RNN model which archives the average [22] Camp, Tracy, Jeff Boleng, and Vanessa Davies. "A survey of mobility models
for ad hoc network research." Wireless communications and mobile computing
localization errors of 1.20 meters and 4.10 meters and also the 2.5 (2002): 483-502.
existing RF-based fingerprinting techniques which archive the [23] J. S. Seybold, Introduction to RF propagation. United Kingdom:
average localization errors of 3.56 meters and 7.26 meters for WileyBlackwell Ney jersy, USA, 2005.
the medium-scale and the large-scale testbeds.

You might also like