Printer Friendly

A study of devising neural network based indoor localization using beacons: first results.


The problem of object localization and navigation has been existing for a long time. Even before the era of modern electronic devices, people used maps, compasses and starry skies. For shorter distances it was enough to remember a set of important objects in the area, due to which the orientation in space was easier. While, for the human perception system, information like "go over the brook and turn left behind a large old oak tree" is useful, it is not a suitable representation for machines (there are systems that process even image inputs. Such an approach, however, is computationally much more demanding and complicated). A common feature for people and machines remains that for localization it is necessary to know the position of any reference points. Their position has to be constant or at least calculable. These reference points may be for example satellites. Outdoor localization works thanks to them. Several satellite-based navigation systems are already deployed worldwide, for example the global positioning system (GPS), with its European variation Galileo and Russian variation GLONASS. They can operate and supply users with precise locations anywhere in the world provided there is an unobstructed line of sight to three or more satellites (Hofmann-Wellenhof 2012, Samper 2008). Indoor localization is however an unsolvable task with satellite-based technology due to the characteristic of the satellite signal which cannot penetrate dense objects, such as roofs, walls or terrain. Even indoor localization still needs a set of reference objects. Usually they are signal transmitting devices with a static position. This position as well as the way and the strength or time of signal transmission are fundamental and indispensable data that localization module needs in order to calculate the object position. These devices operate with different types of signals; in recent times it is usually

* Wi-Fi (Bahl 2001, Wang 2011, Ahmad 2006, Ladd 2002),

* Bluetooh (Wang 2011,Oksar 2014),

* GSM (LaMarca 2005) or

* other types of wireless signals (Curran 2011).

Since this is still the same problem, all types of signals (or in combination) can be processed using the same methods. Researchers have developed many different methods and approaches on how to effectively and accurately localize objects (Curran 2011, Hightower 2001, Wang 2011). Unfortunately, there is no objective way of how to compare the various algorithms of various works with each other, because the accuracy of location method depends mainly on

* the deployment of equipment (the distance between the transmitters, the coverage area of the transmitters),

* the characteristics of the transmitted signal (the signal strength, signal range, frequency of signal transmission, etc.) and

* the number of samples per location--for one run of calculation (preprocessing can also be included).

The rest of the paper is organized into 5 sections. Section 2 presents basic concepts of Bluetooth beacons. Section 3 mentions related work in the area of indoor localization. In Section 4, we report on the measurements that we conducted with beacons to analyze their properties. In Section 5, design and experiments with our neural network indoor localizator based on beacons are presented. Finally, Section 6 briefly discusses conclusions and future work.


Bluetooth beacons are in general low power consumption and low cost transmitters which notify other devices of their presence. These beacons utilize Bluetooth protocol to periodically send short messages --advertisements to the surroundings. These messages contain the beacon identification data and can send also additional data like temperature, humidity or data from other types of sensors.

We can define three basic attributes of beacons transmitting power, advertisement interval and battery lifetime. Transmitting power is measured in logarithmic scale dBm. The more power the transmitter gets, the stronger the signal is and the further it can reach. According to Bluetooth specification documents, there are three classes of transmitting power--class 1 (20 dBm), class 2 (4 dBm) and class 3 (0 dBm). However, in commercially available beacons the user can usually set the power according to his needs, not necessarily to the power of the mentioned classes. Transmitting power has a significant influence on battery lifetime.

Advertising interval can be set in the range from few milliseconds to few seconds. Every time this interval is met, the beacon advertises its message to the surroundings via its antenna. The advertising interval has huge influence on battery lifetime. However with shorter intervals the devices can receive beacon's signal more often and more reliably. This means faster data gathering which leads to faster localization as a tradeoff of shorter battery lifetime.

Beacons can be powered by one of the two sources either directly from electric grid or from built-in battery. Of course, various types of batteries exist, which do have significant influence on the lifetime of the beacon. We note that on average the manufacturers claim the beacons to have the lifetime of few months to 2-3 years of continuous interval Bluetooth signal advertising.

The beacon advertisement message is represented as a stream of bytes, which can be translated at least into the following properties:

* 6 bytes long MAC address

* 16 bytes long Universally Unique Identifier (UUID)

* 2 bytes long major value

* 2 bytes long minor value

* 1 byte of calibrated received signal strength indicator (RSSI), which the receiver should measure 1 meter from the beacon

The advertisement message can contain also additional properties, including the battery status, temperature, humidity or data from other sensors. In commercially available Bluetooth beacons the user can usually change the UUID, major and minor values. This is done purely for the ability of identification of the beacon and aligning it with the physical placement of the beacons (e.g. major value can represent the floor the beacon is on and the minor value can represent beacon ordinal number on that floor).


In their well-informed overview of indoor localization techniques Hightower and Borriello claim there are three main approaches to this problem (Hightower 2001).

The first one is triangulation, which utilizes the geometric properties of triangles to compute object location. This approach can be divided into two subcategories--lateration and angulation. Lateration is based on distance measurements (or estimations) from multiple reference points. Calculating the position in n dimensional space requires n+1 reference points. Angulation approach is similar to lateration, but it uses angles instead of distances to locate an object in space.

The second approach is based on scene analysis utilizing computer vision methods. Static scene analysis relies on detecting the features in the static scene and comparing them with the database, whereas the differential approach tracks the difference between successive scenes.

The final approach, which is also the topic of this paper relies on proximity location sensing. This is done either with physical contact or by making use of wireless signals.

Bahl et al. (2000) used signal from Wi-Fi access points to estimate user's location in the building. They exploited the fact that the signal strength from the access points did not vary significantly in one location. They developed the first fingerprinting approach to in-building localization and achieved an average error of 2-3 meters by using k-nearest neighbors clustering algorithm onto the gathered data.

LaMarca et al. (2005) describe in their paper the possibility to combine the information about GSM, Wi-Fi and Bluetooth (BT) transmitters into the database. They used Bayesian particle filter on the collected data to estimate user's location. Having created the sufficiently dense network of transmitters, by using Wi-Fi transmitter data only, they achieved a median error of 20-30 meters, whereas by using GSM data only they achieved accuracy of 100-200 meters. By combining Wi-Fi and GSM data they could localize the user with the error of 20 meters. They experimented in three different environments--urban, residential and suburban, which yielded the worst results because of low density of transmitters.

Wang et al. (Wang 2011) describe the use of particle filtering (Hightower 2004) and fingerprinting (Bahl 2000) algorithm to combine data from Wi-Fi and BT transmitters inside of a building. The fingerprint data was gathered on a 3x3 meters grid. They could achieve an average error of 2.9 meters with maximum error of 8.9 meters. By using solely Wi-Fi data, the average error was 3 meters with maximum of 9.4 meters.

Ahmad et al. in their work (2006) on indoor localization employed Modular Multi-Layer Perceptron (MMLP) technique to provide better location estimates than other approaches. The authors collected 300 samples of Wi-Fi signal strengths for each of the reference points in their building, which were approximately 2-3 meters apart. Then the data was divided into training and testing datasets and used to train a classification neural network in various configurations. The best result they could achieve was by using a 3-8-8-1 structure, by using logsig and tan transfer functions and Levenberg-Marquardt training algorithm. Average error in this configuration was only 0.12 meters with maximum error of 2.16 meters. Although these results are extremely good for Wi-Fi signal, we cannot assume the same for Bluetooth beacon signal. Due to some inferior properties mainly much lower transmitting power, we expect to get similar or worse results.


Bluetooth operates in the unlicensed 2.4 GHz band which has become useful due to the possibility of high data transfer rates. However, this part of the spectrum is used by many devices including Wi-Fi networks, car alarms, Bluetooth devices and even microwave ovens. This clearly poses a problem, because of their possible interference (Wysocki 2000).

We decided to do an experiment to find out how Bluetooth signal from different beacons behaves in different environments. At the time of the experiment we possessed two types of beacons shown in Table 1.

We placed the beacons one at the time in two different environments--an open hall and a narrow hallway. We believe that this setup is completely different in terms of signal propagation. In the open hall the signal can be radiated to the whole space, but in the corridor it will interact with walls and other possible obstacles. Before the measurement, we made sure that in the areas there are no people or any foreign objects. Then we proceeded to measure RSSI with a handheld smartphone device at 1 meter and gradually stepped up to 15 meters from the beacon. At each point of measurement we collected 100 samples. Then, according to Oksar (2014) we noted the maximum RSSI and plotted the graphs. We chose this high number of samples because we wanted to find out whether the same model of signal propagation can be used for different environments. In real user localization scenario, we would have to use lower sample count due to the time needed to collect them.

The results of the measurements are shown in Figures 1 and 2. As we can observe, the RSSI decreases with the distance. The theoretical RSSI at a given distance is calculated from the calibrated RSSI at 1 meter given by the beacon manufacturer by the use of inversesquare law which applies to electro-magnetic radiation. However, according to the measured data, the signal of USB-powered beacon fluctuates heavily and is influenced by the environment it is in.

One more factor which may influence the strength of transmitted signal is the battery charge status. It might be possible that the beacons with almost depleted batteries will transmit less powerful signal which will influence the accuracy of localization. Unfortunately, we did not conclude any experiments so far to prove or disprove this theory.


The goal of our work is to estimate the location of the user based on the RSSI of beacons in the vicinity by using a properly devised artificial neural network. We chose this approach because of the promising results with the Wi-Fi signal in the paper (Ahmad 2006).

The design of our neural network reflects the nature of the task. It is quite straightforward to use the architecture of a multilayered perceptron neural network. The input layer is obviously determined by the data registered from beacons. The output layer consists of two neurons determining normalized x and y coordinates in the building. We started with a single hidden layer composed of a fixed number of perceptrons. Determining that number to be optimal will be one of the objectives of the next experiments to be described subsequently.

5.1 Four beacon experiment

We decided to conclude the first experiment on the south wing of the 2rd floor of our university building with four Bluetooth beacons. Three beacons were USB-powered beacons and one battery-powered. We designed a 4x4 meters grid on which we performed fingerprinting of RSSI from all four beacons using a handheld smartphone device. This way we gathered 10 samples from each location. Each sample consisted of 10 measurements from which the maximum value of RSSI was taken. We chose 10 measurements because of the similarity to the real user scenario. Ten measurements take about 1-2 seconds to collect and process, which means we can use this approach to localize semi-static target in a real environment.

The beacon locations and the associated heat maps are shown in the Fig. 3. The gathered data was preprocessed for the use of neural network training. We used Encog Machine Learning Framework for the learning phase.

The input data consisted of four real RSSI values from all of the beacons, and the output data represented x and y coordinates on the building floor. Point zero was set in the middle of the hall opening. All values were linearly normalized to the interval from 0 to 1, which means that upon receiving the results from output neurons, we will have to denormalize them to find out the actual x and y coordinates. The whole dataset was divided into the training set (75%) and the testing set (25%).

We trained more neural network configurations. The training algorithm was resilient backpropagation in all the cases and the target error goal was set to 0.1 %, which was reached in all the cases. We primarily used two transfer functions inspired by Ahmad (2006). The number of neurons in the hidden layer was recommended by the Encog Framework, but we plan to conclude more testing in this aspect according to the best practices. The results are shown in Table 2.

The average error in all the configurations is in the interval of 2-3 meters. One of the possible reasons for this result may be the characteristic of the three USBpowered beacons.

In the next experiment we wanted to simplify the experiment area, use more and better quality beacons from a well-known manufacturer, increase their transmitting power and advertising intervals.

5.2 Nine beacon experiment

The second experiment was held in a 36 meters long corridor. The y coordinate was set to zero. This time, we used 9 battery-powered beacons from a wellknown manufacturer which were placed approximately 4 meters aside behind glass panels over the doors. The beacons were set to the highest possible transmission power level of 4 dBm and according to the datasheet, the calibrated RSSI at 1 meter from the beacon is -59 dBm. The advertising interval was set to 100 milliseconds.

This time we gathered 20 samples from each of the 36 locations 1 meter aside. Each sample, as in the previous experiment, consisted of 10 measurements of RSSI of all 9 beacons, from which only the maximum RSSI values were noted.

As in the previous experiment, the data was preprocessed--normalized and segregated. The training algorithm was resilient backpropagation in all the cases and target error goal of 0.1 % which was reached. The results are shown in Table 3.

We also tried to reduce the dataset by one half--we selected only those even coordinated. Results of reduced dataset are shown in Table 4, which are very similar to the results of non-reduced dataset.

In this experiment we could achieve the average error of 1.2 to 1.5 meters across all the configurations. However, as we can observe, the hyperbolic tangent (tanh) transfer function performed much worse in terms of the maximal error--up to 14 meters.

However, if our approach would be used in a real life situation of a semi-static object, the accuracy could be slightly worse due to the movement of the object, which would cause signal variance in the samples. We cannot evaluate this situation yet because we do not possess a device capable of accurately showing the actual position of it in time and that way we cannot compare our results with real ones.


We defined the problem of accurate indoor localization and analyzed the possibilities of using Bluetooth beacons. We performed basic Bluetooth beacon signal analysis and compared it to the theoretical model. We devised a neural network to yield coordinates for indoor localization. As a part of our design methodology, we performed two experiments to determine suitable neural network parameters and to evaluate its performance. First results of our experiments show that our approach has an acceptable localization accuracy compared to the other Bluetooth based solutions. We will continue with further research of this topic and try to achieve even more accurate localization.


This work was partially supported by the Scientific Grant Agency of Slovakia, grants No. VG 1/0752/14 and VG 1/0646/15.


Ahmad, U., Gavrilov, A. and Lee, S. (2006) Inbuilding localization using neural networks. IEEE International Conference on Engineering of Intelligent Systems, IEEE, pp. 1-6.

Bahl, P. and Padmanabhan, V. N. (2000) RADAR: An in-building RF-based user location and tracking system, Nineteenth Annual Joint Conference of the IEEE Computer and Communications Societies. Proceedings (INFOCOM 2000). IEEE. Vol. 2, pp. 775-784.

Curran, K., Furey, E., Lunney, T., Santos, J., Woods, D. and McCaughey A. (2011) An evaluation of indoor location determination technologies. Journal of Location Based Services Vol. 5, No. 2, pp. 61-78.

Hightower, J. and Borriello, G. (2001) Location sensing techniques, IEEE Computer, Vol. 34, No. 8, pp. 57-66

Hightower, J. and Borriello, G. (2004) Particle filters for location estimation in ubiquitous computing: A case study. UbiComp 2004: Ubiquitous Computing. Springer Berlin Heidelberg: pp. 88-106.

Hofmann-Wellenhof, B., Lichtenegger H. and Collins, J. (2001) Global positioning system: theory and practice, Springer-Verlag Wien.

Ladd, A. M. et al. (2002) Using wireless ethernet for localization. International Conference on Intelligent Robots and Systems (IEEE/RSJ), IEEE, Vol. 1, pp. 402-408

LaMarca, A., et al. (2005) Place lab: Device positioning using radio beacons in the wild.

Proceedings of the Third international conference on Pervasive Computing (PERVASIVE'05). Springer-Verlag Berlin, Heidelberg: pp. 116-133.

Oksar, I. (2014) A Bluetooth signal strength based indoor localization method. International Conference on Systems, Signals and Image Processing (IWSSIP), IEEE, pp. 251-254

Samper, J. M., Lagunilla, J. M. and Perez R.B. (2008) GPS and Galileo: Dual RF Front-end receiver and Design, Fabrication, & Test. McGraw-Hill Education.

Wang, R. et al. (2011) Fusion of wi-fi and bluetooth for indoor localization. Proceedings of the 1st international workshop on Mobile location-based service (MLBS '11). ACM,.USA. pp. 63-66.

Wysocki, T. A. and Zepernick, H-J. (2000) Characterization of the indoor radio propagation channel at 2.4 GHz. Journal of telecommunications and information technology. pp. 84-90.

Filip Mazan, Alena Kovarova

Faculty of Informatics and Information Technologies Slovak University of Technology in Bratislava, Slovakia and

Table 1. Two beacon types used in experiment.

                       Beacon 1      Beacon 2

Power                  USB-powered   Battery-powered
                                     CR2032 battery
RSSI at 1 meter        -66 dBm       -59 dBm
Advertising interval   100 ms        500 ms

Table 2. Results of various configurations of neural networks in
the first experiment.

Network     Transfer   Average   Maximal
structure   function   error     error

4-7-2       Tanh       3.04 m    5.95 m
4-7-2       Log        2.62 m    8.84 m
4-7-7-2     Tanh       2.76 m    10.20 m
4-7-7-2     Log        2.48 m    8.61 m

Table 3. Results of various configurations of neural networks in
the second experiment.

Network     Transfer   Average   Maximal
structure   function   error     error

9-15-2      Log        1.42 m    7.41 m
9-15-2      Tanh       1.37 m    14.60 m
9-15-15-2   Log        1.21 m    5.98 m
9-15-15-2   Tanh       1.25 m    6.52 m

Table 4. Results of various configurations of neural networks in
the second experiment using reduced dataset.

Network     Transfer   Average   Maximal
structure   function   error     error

9-15-2      Log        1.24 m    4.24 m
9-15-2      Tanh       1.52 m    9.66 m
9-15-15-2   Log        1.48 m    5.71 m
9-15-15-2   Tanh       1.53 m    6.15 m

Table 5. Comparison of localization accuracy.

Solution       Trans-   Number    Algorithm        Samples    Average
               mitter   of                         per        error
               type     trans-                     location   m

RADAR          Wi-Fi    3         k-nearest        40         2.65
(Bahl 2000)                       neighbors
Rice                              Bayesian
University     Wi-Fi    9         inference        100        1.50
(Ladd 2002)                       localization
Kyung Hee                         Classification
University     Wi-Fi    3         neural           300        0.13
(Ahmad 2006)                      network
Fusion         Wi-Fi    ?         Bayesian         700        3.03
(Wang 2011)                       filtering
Fusion         Wi-Fi              Bayesian         700        2.91
(Wang 2011)    + BT               filtering
ASELSAN        BT       6         RMSE             ?          2.31
(Oksar 2014)
our best       BT       9         Regression       20         1.21
attempt                           neural
COPYRIGHT 2015 University of the West of Scotland, School of Computing
No portion of this article can be reproduced without the express written permission from the copyright holder.
Copyright 2015 Gale, Cengage Learning. All rights reserved.

Article Details
Printer friendly Cite/link Email Feedback
Author:Mazan, Filip; Kovarova, Alena
Publication:Computing and Information Systems
Article Type:Report
Date:Feb 1, 2015
Previous Article:Factors affecting the effectiveness of web 2.0 as a mobile learning tool in the workplace: a conceptual view.
Next Article:A fuzzy logic expert system for the prediction of breast cancer.

Terms of use | Privacy policy | Copyright © 2022 Farlex, Inc. | Feedback | For webmasters |