In the Ambient Assisted Living (AAL) scenario, indoor localization represents one of the main pillars for the development of contextaware applications. In this context, comparing and testing indoor positioning system is a hot topic in the indoor localization research community. In fact, after several years algorithms and methods have been developed and matured, no general frameworks exist yet to reliably compare them. The scarcity of common datasets for o -line test of emerging indoor positioning systems, together with the lack of available frameworks for real-time comparison and evaluation of indoor localization solutions, is one of the main barriers to their standardization. The lack of a common usable software framework for implementing and testing new algorithms, on a fair basis, is an additional barrier. In this work, we address this research challenge by proposing a free software framework enabling the development of indoor localization applications on the Android platform. It is composed of two applications: PrettyIndoor is a positioning app, FingerFood is a ngerprint-building app. We show that the framework's modular architecture can be exploited to easily develop many data fusion strategies, in order to easily compare and improve indoor positioning systems.
Indoor Positioning Systems (IPSs) can be used where the standard GPS
technology is not available. Nevertheless, no standard solutions exist yet for the indoor
scenario. A standard and e cient solution, like the GPS is for outdoor
positioning, would be very useful as the base for innovative location-based service.
Possible use cases include advertising in big malls, navigating to a speci c place
in wide public areas, assisting users in emergency situations and many more.
Among the possible target users, people with motion or cognitive impairments
could perceive large crowded environments as intimidating [
The research community has spent big e orts in the last years to improve the current technology of indoor positioning systems. Several approaches have been proposed, including infrared light, ultrasonic sensors, WLAN, RFID, Bluetooth Low Energy, Ultra Wide Band, ZigBee and computer vision, among others [7{ 11]. The literature shows that researchers try to improve previous solutions using several systems, mathematical methods, and signal processing techniques. In this paper, we are speci cally interested in smartphone-based methods that do not require instrumenting the environment; in other words, we consider localization systems that run on a smartphone and require no dedicated infrastructure to be deployed in the area of interest.
The main reason to avoid a dedicated infrastructure is that its deployment
may be impossible for reasons of costs, security, limits on installable devices,
especially in a public area, hospital, university, cultural heritage site, where
regulations may be limiting the available options. As far as why we limit our
interests to smartphone, the reasons are again convenience and cost. In recent years,
smartphone-based solutions are emerging, as nowadays the sensors embedded in
smartphones make them a solid support for positioning and navigation purposes.
In practice, the state of the art in terms of on-board sensors, computation power
and other hardware speci cations makes smartphones mature enough for this
kind of application. In this regard, the experience of the IPIN [
With such a wide range of methods and sensors used, comparison of IPSs
performance is a major issue. The problems are mainly the lack of a common
dataset for o -line comparison of methods [
The main purpose of this paper follows this line of thought: we propose an
open framework to compare and evaluate IPSs. Despite the maturity of
mathematical models and data-fusion algorithms, very few similar open framework
solutions have been proposed. A public and common software platform allows
to move towards the ambitious goal of testing the reliability and robustness
of IPSs, especially when fusing di erent techniques. A notable related work,
in terms of open source indoor localization application, is represented by
AnyPlace [
In this work, we propose two free software tools running on the Android OS, for indoor localization purposes. The rst tool, called PrettyIndoor, is the positioning application, while the second one, called FingerFood, is used to acquire WiFi and magnetic ngerprint maps. Both are modular, allowing to include and compose di erent algorithms and methods. We believe that modularity can be useful to the research community for in-depth research and service composition.
Both tools are freely distributed under the Apache Software License (ASL),
version 2.0, and are publicly available on GitHub 6. We chose to use a free
software license because we consider the use of free software in research activities
a de nite plus, from both a practical and philosophical point of view [
The rest of this paper is organized as follows: Section 2 describes the design of the proposed framework in details, Section 3 presents the experiments performed
in order to test the framework and the obtained results, while Section 4 draws the conclusions. 2
The system is composed of two applications, namely PrettyIndoor and FingerFood. The former is the position engine which implements all the algorithms and the data structures required for getting things done. It comes with a front-end, thought for researchers testing operations in many possible strategies both existing and coming in the future. The latter is a utility application that allows the user to capture WiFi and magnetic ngerprints, save them into a le, and make a textual ngerprint map that can be used by the other application.
Both applications access the phone sensors through a library which extends the Android native methods for sensor access. 2.1
The purpose of FingerFood is to ease the tasks related to creating ngerprint maps. The front-end is shown in Figure 1. It is composed by three Android Activities: one for survey management, one for ngerprint acquisition, and one for map completion.
To use FingerFood, the user has to select the survey he wants to work on or create a new one. During ngerprint acquisition, coordinates can be chosen by either directly writing them or moving step by step using the arrows. Step length can be customized. When pressing the buttons \Start WiFi acquisition" and \Start magnetic eld acquisition" the application locally registers the coordinates and the received data for the speci ed duration. Once the user has nished the survey on a given oor, data is merged directly on the device by averaging the registered measurements and so the map is created. It can be exported into a text le, but there is no need to do it when using PrettyIndoor, which communicates directly with the FingerFood 's Content Provider. 2.2
The positioning application is composed of an Android service implementing the back-end and an essential graphical user interface shown in Figure 2. These two main components provide all the utilities for testing indoor navigation solutions. Currently, the user can choose among PDR-driven, ngerprint-driven or mixed strategies but more techniques are planned to be implemented.
PrettyIndoor requires a starting position, to be speci ed in the text box. After doing this, the localization service can be started by pressing the green Play button. This action switches the front-end to the online mode: the bottom oating action buttons change and the toggle buttons corresponding to sensors used by the chosen strategy are switched on. These buttons allow to enable and disable any data source during run-time, which is a useful function for doing a deeper testing. The indoor positioning service runs in the background updating its saved current position in real-time, using the user-selected method. The current position can be saved into a log le, which is automatically created by the application, by pressing the bottom-right button in online-mode GUI. The format of the position log is simple: timestamp,x; y; z, where timestamp is the time when the button is pressed, x; y is the 2D position on a oor, the position moves on the z-axis by using an integer value representing the oor, where 0 is the ground oor. Lines are newline-terminated, so the log is a standard a CSV text le where every row contains a di erent time-position relation. To save the current position the ag button must be pressed, but the separation of the backend logic into an Android service allows to easily expose this function to a future extension of the application and even to third-party applications. 2.3
Figure 3 shows the architectural concept of PrettyIndoor. The main goal was to develop an three-tier architecture with a logical separation between native raw data, data abstraction layer and core logic layer. The strength of this model is the easy implementation of further modules and strategies into the core logic layer and more speci cally, inside the localization strategy sub-layer. In fact, handlers which manage the native raw data, coming from the physical sensors on the smartphone board, are o ered using an adapter. It allows to implement new algorithm, or to enhance the previous one, with no any a priori knowledge of how the operating system manages sensor data. The nal output is a local coordinates triple x; y; z useful for rendering, navigation, and mapping.
Native Data Handler
Data Adapter Sensor Handler
Wi-fi Handler Data Layer
Data Refiner Compass Step detection Preprocessing Sub-Layer Core logic Layer
Localisation Strategy State prediction filter Kalman Filter
Particle Filter Localisation strategy
Sub-Layer
As shown in Figure 4, the current version of the PrettyIndoor 's service implements ve di erent strategies for solving the indoor location problem: Pedestrian Dead Reckoning (PDR), k-Nearest Neighbours (K-NN) WiFi based ngerprinting, k-Nearest Neighbours (K-NN) geomagnetic based ngerprinting, Kalman lter PDR- ngerprinting fusion, and particle ltering fusion.
SimpleIndoor Service Fingerprint Strategy Particle Filter Strategy
Kalman Filter Strategy
The rst strategy is based on PDR techniques. It relies only on accelerometer, gyroscope and magnetometer using sensors for determining orientation and step events. The conceptual work- ow is shown in Figure 5. This implementation of the technique consider an average step length of 0.6 m and simply elaborates the variation on the x; y coordinates and adds it to the previous saved position.
Strategies that use a ngerprint map compare either the measured WiFi RSSI or magnetic eld with the values in the database. The position is then found by operating on the results of the k-Nearest Neighbours (k-NN). A lot of solutions in literature don't limit themselves to the usage of a single technique, instead they usually combine more of them. In order to do this, one of the location strategies contained in the application uses a Kalman lter. It always keep the positions found by both PDR and ngerprint, whose di erence is then corrected combining it with a prede ned covariance. This re ned variation is then added to the position and this output is assumed as the new coordinates.
Accel.
Magnet Gyro.
AccMagOrientation
AccMagOrientation
GyroOrientation
Time interval
Fig. 5. The work- ow of the orientation algorithm.
Another strategy that uses a state estimation lter is based on a particle lter. In contrast with the previous one, this lter directly operates on the position. In fact, during the initialization a number of particles representing the possible positions are generated on the start point. When a step is detected, they are moved by the variation detected by the PDR plus a random error. This error comes from a model represented by a zero-mean Gaussian distribution with set to 0.15 m. For each particle, the algorithm then calculates a distance dparticle = n P di and constraints the particles to lie inside the map. Picking a random i=1 ri;particle number from 0 to an experimentally tunable maximum, if it lies between zero and dparticle, the particle is removed. Eventually, a number of particles is resampled in order to restore their original number. All these strategies take account the map topology in order to assure the correctness of the found positions. 2.4
Since portability is a ubiquitous requirement in recent software production, PrettyIndoor, FingerFood and their libraries accomplish this by encapsulating data in a proper type for each kind of source.
android.hardware Sensor Event Listener indoornavigation android.handlers Sensor Data Emitter
Adapts Android’s Sensor Events types RawData
Figure 6 represents how the listener and the adapter are organized and work. For example, when an Android SensorEvent is sent to its listener, the array containing its oating point values is read by the adapter for making an Acceleration object which is then sent to the classes that are waiting for it.
The actual data type hierarchy is represented in Figure 7. For this, the Android service that implements navigation makes a new data object when an Android sensor event arrives through an adapter class.
RawData In order to test the capabilities o ered by the proposed open framework, di erent experimental campaigns have been performed at the Italian National Council of Research (CNR), located in Pisa. The map of the experimental region is characterized by a straight corridor with o ces located on both sides, a small hall between the two corridor and two o ces of 10 m2, as shown in Figure 8.
Using FingerFoood, WiFi and magnetic ngerprints were acquired by standing still for 5 s in each reference point, in order to create two ngerprint maps. The points are equally spaced by 60 cm in both directions in order to uniformly cover the interested area.
Figure 9 shows the two di erent paths used in the experimental campaign, represented as green lines. Paths are composed by 13 and 8 points, respectively. Points were placed on the oor, using circle markers, used as ground truth. An actor, who held the smartphone in his right hand, used the PrettyIndoor application as explained in Section 2.2. We tested the application using two di erent smartphones: a Xiaomi Mi3w with Android 4.3, and an Lg G3 with Android 6.4. Finally, for each path and for each smartphone, di erent runs were performed using the di erent algorithms and strategies currently implemented on PrettyIndoor.
In order to test the proposed framework, we evaluate the results on di erent paths calculating min error, max error, mean error, and third quartile error, the latter according to the EvAAL competition metric. The overall localization performance is measured on two paths for ve di erent strategies: corridor 3 office 1 corridor 1 hall corridor 2
Fig. 9. The two evaluation paths. { PDR - only using the inertial sensors; { WiFi - only using the WiFi ngerprint database; { GeoMag - only using the magnetic database; { PF - using the rst three fused in a particle lter; { KF - using the rst three fused in a Kalman lter.
Performance is generally better for the second path. By looking at the results, for the rst three simple strategies, we can observe that a signi cant part of path 1 has bad WiFi performance in a speci c area, which probably means that the ngerprint database should be improved in that area. Similar observations can be done for the magnetic ngerprint database. Additionally, one can observe that the step detection implementation is far from perfect, and works reasonably well only if there are bends in the path. In contrast, steps are lost in long rectilinear paths. These problems are all concentrated in the second path, which explains why particle lter performance is much better for the rst path.
It is worth noting that the above analysis is much eased by the modular nature of the tools, which allow to enable, disable and fuse modules together.
All in all, the results obtained in the second path suggest that the framework can produce good results once the algorithms are optimized, and its purpose is ful lled, that is, creating a exible, extensible and modular indoor localization suite for Android that can be useful for researcher thanks to the free software license used for distribution. 4
In this work, we propose a free software framework to develop indoor localization applications. We show that its modular architecture can be exploited to easily develop diverse data fusion strategies and to analyze their relative strengths and weaknesses.
The PrettyIndoor application is currently available and usable on Android smartphones and can easily be used to generate logs in order to validate algorithms and enhancement. It is a modular development environment for implementing existing and future IPS solutions, algorithm and fusion strategy implementations, and testing them on di erent paths into di erent indoor environments in a fair way.
FingerFood can currently be used to collect and store ngerprint maps for WiFi and magnetic eld. The modular and expendable architecture of both tools makes them a basis for the research community working on indoor localization.
Both applications are still under active development. Future work includes re ning the already implemented methods, introducing an adaptive motion module for a better understanding of the human walking pace and more.
The free software Apache software license used for distribution allows anyone to use, modify and redistribute the software, whether modi ed or not. It also avoids the risk that developers include algorithms on which they own a patent, with the future purpose of asking royalties on them.
Acknowledgments. This work was carried out in the framework of the INTESA project, co-funded by the Tuscany Region (Italy) under the Regional Implementation Programme for Underutilized Areas Fund (PAR FAS 2007-2013) and the Research Facilitation Fund (FAR) of the Ministry of Education, University and Research (MIUR).