=Paper=
{{Paper
|id=Vol-2498/short19
|storemode=property
|title=Indoor navigation for a library
|pdfUrl=https://ceur-ws.org/Vol-2498/short19.pdf
|volume=Vol-2498
|authors=Janek Stoeck,Harald Sternberg
|dblpUrl=https://dblp.org/rec/conf/ipin/StoeckS19
}}
==Indoor navigation for a library==
https://ceur-ws.org/Vol-2498/short19.pdf
Indoornavigation for a library
Janek Stoeck and Harald Sternberg
HafenCity University Hamburg
Abstract. To find the optimal way to a destination without the help of
appropriate tools is sometimes not particularly easy or even impossible.
This is also the case in the HafenCity University Hamburg (HCU) library
in Hamburg. Due to the inhomogeneous signatures of the books, it is not
possible to make clear assignments to specific subject areas here. This
is due to the fact that the current collection has been brought together
from several old collections.
To simplify this situation, an indoor navigation system based on smartphone
sensors is currently being developed. Acceleration sensors are used to
realize a step counter. In combination with gyroscope data a dead
reckoning can be carried out. The height information is derived of
barometric measurements. For additional support, routing graphs are
used to constrain the walking direction.
Keywords: indoor navigation, indoor localization, dead reckoning, routing,
topological support
1 Introduction
Automatically locating people in real time inside buildings is a challenging task.
There are many technologies that can be used to enable real-time localization
(WLAN, Bluetooth, Ultra-Wide-Band etc.). In order to realize complex applications,
custom developments with infrastructure-based positioning methods are more
complicated to realize, especially when it comes to linking several disciplines. This
is where systems that the user already owns, such as the widely used low-cost
multi-sensor system smartphone, are ideal. It combines sensor technology for
tracking objects, as well as many advantages of modern computers in a portable
handy device.
Such a complex application is also presented in the following project. In this
project a navigation from the current location to a searched book in the library
of HCU shall be offered. The reason for this is an inhomogeneous signature
indication as well as a not intuitive description of the location of the book.
For this purpose, the book search is to take place on the smartphone with the
website offered by the library. Afterwards the way to the book is to be represented
on a map. Here, the presentation of a website, extraction of desired data, linking to
a database containing information between semantic and topological information
of the book, localization and tracking of the smartphone and presentation of the
route and the environment on a map are used.
2 Janek Stoeck and Harald Sternberg
It is much easier for the user to use his own smartphone because he is familiar
with it and does not have to find his way into a custom development. In addition,
a custom development would result in a limited number of simultaneous users
due to limited quantities of the system.
In the following section 2 related works are shown. Afterwards in section 3
the application with regard to localization and in section 4 the work-flow are
described. In section 5 first results are shown, whereupon in section 6 a summary
and an outlook into the following work is given.
2 Related Work
In [5] a algorithm based on micro electro mechanical systems (MEMS) was
presented. A fusion of sensor data and map data for position estimation was
performed using particle filters. The used sensors are accelerometer, gyroscope
and barometer of a smartphone. The accuracy of this algorithm depends strongly
on the number of particles. The more particles are used the more accurate the
estimation is, so each particle influences the accuracy but also the computation
time which is why it is not particularly suitable for a runtime application on the
smartphone with today’s technology.
Reference [3] presents an algorithm that uses smartphone MEMS sensors
to estimate the position of a smartphone in a building using a particle filter.
Inertial measuring units (IMU) and barometers are also used here. Furthermore,
a possibility is presented to use additional sensors to limit the search range of
the particles. Magnetic signals are used for this purpose. Since particle density
is a critical factor as mentioned before and additional infrastructure has to be
provided, this form is not suitable for the planned project.
Another system is presented in [1]. In addition to the smartphone, a smartwatch
is also used, which is located at the foot ankle. The smartphone is used for a
dead reckoning (DR), which is similar to the one presented in this project.
The smartwatch is used to get the trajectory over double integration of the
accelerations to completed path. By the property that there are phases during
the step where the IMU records accelerations close to zero, zero velocity updates
can be performed to minimize sensor drift. The results of both approaches are
combined into a single trajectory. Although the data is promising, the use of
additional hardware is not practical.
Another unpublished project at HCU also deals with indoor navigation, only
on a larger scale. In this case, an indoor navigation for the building of the HCU
has to be set up. The purpose of this is to make it easier for the user to find the
room on the basis of the room number or the person sitting there. Here, too, there
are problems in finding the right room by grouping the rooms in combination
zones, without displaying the room numbers on signs.
Indoornavigation for a library 3
3 System setup
The following section provides an overview of the prerequisites and an insight
into the algorithms used to implement the indoor navigation system.
3.1 Prerequisites
For an indoor navigation system to certain rooms, an accuracy in the class
between 2-5 m is usually sufficient to reach the desired destination. In the case
of this project, this is not enough, because the shelves in which the books are
located are about 0.5 m apart. The shelves are marked at the front with the
signature area, so that an accuracy of 1-2 m is sufficient to detect the correct
shelf. Since the shelves are loaded from both sides, the correct side of the shelf
should also be displayed.
3.2 Sensors and algorithms for indoor navigation
The algorithms used were already developed in the above-mentioned project and
were adopted for this project. The algorithms were also strongly based on [5].
Fig. 1 shows the principle of the inertial navigation using smartphones,
which is realized by the IMU (accelerometer, gyroscope) and a barometer. The
accelerometer is used to implement a step counter like in [5]. In Fig. 1 this and
the step length estimation are grouped as virtual sensors, as they are based on
another sensors data and further are no physical entities. The step counter and
the step length estimation are used together with integrated angular velocity
of the gyroscope to perform a step based DR. The barometer is used to derive
height changes to identify the floor where the user is located. It is assumed that
the user is holding the smartphone in his hand within an angle of roughly 45 to
the ground.
First of all, the axes of the acceleration and angular rate sensors are planed in
the Z direction. This is done by calculating the inclination angles from acceleration
using (1) and rotate the acceleration and angular rates by them.
α arctanp q
ay
az
(1)
β arctanp q
ax
az
The step counter is based on the accelerometer z-axis and has two conditions
to be fulfilled. First, an initial maximal threshold has to be passed followed by
passing a lower threshold. If these conditions are true, a step is detected.
If the actual accelerometer z-reading passes the thresholds value including a
scaling factor, the threshold value increases to the actual z-reading.
This is the advantage of this step counter, because it ensures, that it adopts
the walking behavior of the user. Further the biases of the accelerometer have no
impact on this method, because the thresholds are dependent to the mean value
of the acceleration in z.
4 Janek Stoeck and Harald Sternberg
Fig. 1: System setup for inertial navigation using smartphone sensors
The step length is estimated using the accelerations as well. For this the
algorithm of [4] is used with a K of 0.54. Amax is the highest value in the z
direction and Amin is the lowest value during a recognized step.
a
li K 4 Amax Amin (2)
r i r i1 ωi ∆t (3)
The angular rates are integrated over time with (3). The actual angle ri depends
on the previous angle ri1 and the actual angular rate ωi multiplied with the
time difference ∆t between actual and previous reading.
Combining the information that a step is taken, the integrated gyroscope
readings and the step length DR can be performed. The formula can be seen in
(4), where xi is the actual position, xi1 is the previous position and R and t
are the rotation and translation between both. R contains the angular reading
around the z-axis and t includes the estimated step length. For a more detailed
described implementation of the algorithm for DR and the step counter see [5].
xi xi1 Rt (4)
b pphq
288.15K 288.15K 5.255 1013.25hP
hb
a
(5)
0.0065 K
m
The height information is obtained by deriving the height from barometer data
using the general height formula (5). Inaccuracies based on changing air conditions
do not have a big impact, because the height information is used to identify the
Indoornavigation for a library 5
floor, where the differences between them are more than 4 m. Together with the
data of the DR a 2D+1D trajectory is formed.
As the integration of a angular velocities leads to an time dependent drift, the
system is supported by the addition of routing data. In doing so the algorithm
1 is used. If the estimated position is close to the routing graph ?P and the
orientation Ψ is about 7.5°, the orientation Ψ is set to the orientation of the
graph ?P and the position is set to the base point BP , which is calculated from
the triangle between start and end node of the routing graph and the estimated
position. In this way, the user moves on the graph. If the orientation of the
smartphone deviates over 7.5°, the routing graph is left again.
Algorithm 1 Snap Position to Graph
if P px, y q, BP px, y q 0.3m then
if |Ψ ?graph | 7.5 then
Ψ Ð ?graph
P Ð BP
end if
end if
Another topological support is implemented through the delimitation of the
areas in which the stride length changes to a given measure. This happens here
when stairs are used. It is necessary because climbing stairs does not differ
significantly from normal walking and therefore no correct step length can be
estimated. In addition, such topological area checks can also be used to determine
whether the user is at the lower or upper end of the staircase. This is particularly
useful if the user’s smartphone does not have a barometer. Here, the height can
be estimated so that the height change corresponds to a stair step height, if the
user is on the stairs.
4 Workflow
The user searches for the desired book via a web interface which is integrated in
the application. Using a JavaScript code, the signatures of the borrowable books
are now selected. Once a signature has been identified, the appropriate routing
node is searched for using an SQL query. The user is then asked to initialize his
starting point. This is done by scanning a QR code with a camera. This contains
the location information (position and elevation coordinates and orientation of
the smartphone).
A routing is now calculated with the help of the two nodes. For this case the
Dijkstra algorithm is used, where each following node after the current get a
weighting, based on its distance to the current node. The next node picked is
the one with the lowest weight. This continues until the destination is found. As
soon as the routing is known, the user is shown the corresponding route in a map
6 Janek Stoeck and Harald Sternberg
Fig. 2: Workflow of the developed android application
view. Now the position of the user is estimated using the algorithm described in
section 3. Each new position updates the map so that the progress is visible.
5 Experimental Results
To investigate the inertial navigation system, several trajectories were recorded
in the building of the HCU, each of which took the same course with minor
deviations. In this early phase of the project the tests were done with only one
person and used a LSM6DL from STMicroelectronics which is built in a Samsung
Galaxy S8.
The start and end points are identical so that a loop-closure error can be
calculated. This should provide information about the accuracy of the position
estimation. In addition, the trajectories run over three floors, with the changes
being made via stairs and elevator rides. The distance covered is 340 m. The
results are shown in Fig. 3. Similar to the work flow described in the previous
section, the test person had to scan a QR-Code to initialize the position and
orientation in order to get an absolute position on the maps coordinate system.
The upper part of the figure shows the course of a total of four trajectories on
the fourth floor of HCU. In the lower part the heights calculated from barometer
data are displayed. The mean error of the four trajectories is 1.80 0.25 m and
thus represents a solid result of less than 1% of the traveled distance. The sensor
drift could clearly get rid of with the described algorithms in section 3. Since
the respective floor data are selected over the height, an accuracy of the height
estimation to one meter is required. This could also be achieved.
Another test should show that the algorithms also work in smaller areas, such
as the library. The library is a room of HCU, which has the dimensions of about
40 m x 20 m and extends over three floors.
For this purpose the application was used as described in chapter 4, whereby
the same book was searched for in each of the thirteen tests carried out. The
start was on the ground floor in front of the stairs, where a QR code was located.
The destination was a book on the first floor. The test person always went from
Indoornavigation for a library 7
Fig. 3: Several trajectories estimated with the described setup
the starting node to the routing node in the passage between the shelves in order
to determine the deviation from target to actual coordinate. As in the test before
this was done by the same person with the same smartphone.
In Fig. 4 the results are shown visually. There the target node of the routing
graph and the shelf in which the book is located are shown in red. The last
estimated position is colored blue and the ground truth trajectory is colored
green. To get the accuracy of the system the deviation of each last estimated
Fig. 4: Results of the position estimated in the library
position to the targeted routing node were calculated. It can be seen that the
8 Janek Stoeck and Harald Sternberg
results scatter within a range of two meters. The mean deviation is 1.35 0.5
m. This is a accuracy of 4 % of traveled distance. The effect, that all deviations
from last estimated position to the targeted routing node are negative, indicates
that the step counter did not recognize the correct amount of steps.
6 Conclusion and Outlook
This paper presents an application for finding a book in a non-homogeneously
structured library. The stock of this library was brought together from different
stocks, whereby due to the expenditure a new cataloguing was not possible.
Algorithms are used which allow for the spatially accurate estimation of the
position (2-5 m). By adding topological support, an attempt was made to achieve
a more accurate result, since the shelves wanted are approximately 0.5 m apart.
The results are all in the range of less than two meters, which is a good result
for the localization of people indoors. This also satisfies the requirement for an
accuracy of less than 2 m for this project. However, reaching the right side of the
shelf has not yet been achieved on the basis of the results. To achieve this, the
parameters of the step counter can be adjusted or additional sensors can be used.
For example, the camera could be used again to identify the correct shelf on the
basis of the signature on the front of the shelf and determine the deviation from
the current position.
While particle filters are out of the question due to the runtime of the
algorithms, the implementation of a Kalman filter for support could be tested in
the future. However, in contrast to older IMUs, the IMUs used here show a low
drift. Accordingly, it will hardly be very significant in such short distances as in
this application.
One can claim that a step length estimation according to [4] does not make
sense for widespread use, since the K factor is a value that depends on the user.
In order to avoid this we will implement and test algorithms in the future, as
they are presented in [2], since they rely on the estimation of the K-factor and
show a very small error according to own statements.
References
1. Gunia, M., Wu, Y., Joram, N., Ellinger, F.: Building up an inertial navigation system
using standard mobile devices. Journal of Electrical Engineering 5, 299–320 (2017)
2. Ho, N.H., Truong, P., Jeong, G.M.: Step-detection and adaptive step-length estimation
for pedestrian dead-reckoning at various walking speeds using a smartphone. Sensors
16(9), 1423 (2016)
3. Real Ehrlich, C., Blankenbach, J., Sieprath, A.: Towards a robust smartphone-based 2,
5d pedestrian localization. In: 2016 International Conference on Indoor Positioning
and Indoor Navigation (IPIN). pp. 1–8. IEEE (2016)
4. Weinberg, H.: Using the adxl202 in pedometer and personal navigation applications.
Analog Devices AN-602 application note 2(2), 1–6 (2002)
5. Willemsen, T.: Fusionsalgorithmus zur autonomen Positionsschätzung im Gebäude,
basierend auf MEMS-Inertialsensoren im Smartphone (2016)