cations A Survey of Calibration-free Indoor Positioning Systems

cations A Survey of Calibration-free Indoor Positioning Systems
Computer Communications 00 (2015) 1–18
A Survey of Calibration-free Indoor Positioning Systems
A. K. M. Mahtab Hossaina,∗, Wee-Seng Sohb
a Mobile
Internet and System Laboratory, Department of Computer Science, University College Cork, Ireland
b Department of Electrical and Computer Engineering, National University of Singapore
Last decade observed a significant research effort directed towards indoor localization utilizing location fingerprinting techniques.
Fingerprinting solutions generally require a pre-deployment site survey procedure during which a radio-map is constructed by
laboriously collecting signal strength samples (e.g., Wi-Fi) over the whole localization area. However, such localization efforts have
certain shortcomings. For example, it is time consuming, labor intensive, vulnerable to environmental changes, and the process
requires certain pedigree on the surveyor that may deem the fingerprinting techniques impractical to be deployed over large areas
(e.g., shopping malls, multi-storey offices/residences, etc.). Newer emerging techniques try to bypass this expensive pre-deployment
effort of fingerprinting solutions altogether. They may build the radio-map through the implicit participation of the building
occupants, office employee, shoppers, visitors, etc. Apart from the traditional performance comparison criteria like accuracy,
precision, robustness, scalability, algorithmic complexity based on which the localization techniques were evaluated, these newer
approaches warrant some additional ones. For example, whether they require an actual geographical map of the localization area,
the percentage of occasional location fix to ensure reasonable accuracy, the usage of explicit/implicit user participation to construct
the radio map, the usage of building landmarks (e.g., entrance, conference room, elevator, escalator, etc.) or additional sensors
(e.g., accelerometer, gyroscope, compass, etc.), whether they address device heterogeneity, etc. In this article, we survey the
newer emerging fingerprinting solutions that try to relieve the pre-deployment woes. We also identify some newer performance
comparison criteria based on these solutions’ inherent characteristics, and apply them together with the traditional ones in order to
evaluate a number of such proposed systems.
c 2014 Published by Elsevier Ltd.
Keywords: Indoor Positioning System, Location Fingerprint, Calibration-free, Site-survey-free, Localization.
1. Introduction
Global Positioning System (GPS) is the most popular
outdoor localization system, it has certain limitations
when applied inside the indoor environment. Indoor
environment requires finer granularity and precision of
localization accuracy. For example, while a 5 ∼ 10m
accuracy may well be quite acceptable outdoor, when
inside a building, the target could be in different rooms.
From operational perspective, GPS’s signals are not designed to penetrate most construction materials, and
generally require line-of-sight (LOS) transmission between receivers and satellites.
Accurate real-time indoor location determination is
an indispensable part to enable various context-aware
services and protocols [1, 2, 3]. Its applications range
from a hospital’s inventory and medical resource tracking, navigation tools for fire-fighters inside an unknown
indoor environment, to various commercial locationbased services (e.g., finding the cheapest store, sales, or
electronic coupons inside a shopping mall, etc.). While
An indoor positioning system (IPS) is a framework
consisting of a network of devices (both customized or
off-the-shelf) that are used to wirelessly locate objects
∗ Corresponding
Email addresses: (A. K. M. Mahtab
Hossain), (Wee-Seng Soh)
Hossain et al. / Computer Communications 00 (2015) 1–18
whole localization area. However, such localization efforts have certain shortcomings. For example, it is time
consuming, labor intensive, vulnerable to environmental changes, and the process requires certain pedigree on
the part of the surveyor that may deem the fingerprinting
techniques impractical to be deployed over large areas
(e.g., shopping malls, multi-storey offices/residences,
There is a breed of newer emerging calibration-free
techniques that try to relieve the pre-deployment woes
of the fingerprinting solutions discussed above. In this
article, we provide a survey of such calibration-free systems. We also identify some newer performance comparison criteria based on these solutions’ inherent characteristics, and apply them together with the traditional
ones in order to evaluate a number of such proposed
systems. Our main target is to provide a qualitative
overview of such systems, and offer a quantitative comparison among them whenever possible. To the best of
our knowledge, our article is the first to provide such a
survey of the calibration-free indoor localization techniques, which will become increasingly important.
The survey paper is organized as follows. We explain
the basic fingerprinting approach in detail, and its apparent shortcomings in Section 2. In Section 3, we outline
both the traditional and newer performance comparison
criteria that we identify, considering the inherent characteristics of the emerging approaches that try to get rid
of the pre-deployment woes typically seen in the fingerprinting based solutions. Next, we discuss a few emerging calibration-free techniques in Section 4, and provide
a qualitative comparison of the systems in Section 5. Finally, we conclude in Section 6 by pointing out some of
the future research directions.
or people carrying handheld devices inside a building.
The research efforts for such systems over the years can
largely be divided into two main categories:
• Those that rely on specialized hardware (e.g., IR
or RF tags, ultrasound receiver, etc.) and require
extensive deployment of infrastructure throughout
the service area solely for localization purpose [4,
5, 6, 7]. They may also require customized tags
attached to an object or specialized client devices
to be carried by the person to be tracked.
• Those that are built on top of existing infrastructure (e.g., Wi-Fi or Bluetooth communication networks) and use off-the-shelf wireless networking
hardware found in handheld devices.
The family of various location fingerprinting [8, 9,
10, 11, 12, 13] or model based [8] techniques where
RSS measurement is used to estimate the distance between a transmitter and a receiver based on some radio propagation model, and other proximity based approaches [14, 15, 16, 17, 18, 19] found in wireless sensor networks fall into the second category above. Location fingerprinting techniques hold more potential in an
IPS because they offer finer accuracy (≈ 2 to 3m) and
precision compared to both the model based or proximity based approaches. However, it is much lower than
many of the infrastructure-based scheme’s cm-level accuracy. There is currently no de-facto standard for an
IPS design. Nevertheless, there are several commercial systems in the market. Firefly [20] (IR), OPTOTRAK [21] (Camera + IR), Sonitor [22] (Ultrasound)
are client-tag based infrastructure oriented systems that
require extensive deployment effort. Ubisense [23]
(UWB), Topaz [24] (Bluetooth), Apple’s iBeacon [25]
(Bluetooth Low Power), MotionStar [26] (Magnetic),
Easy Living [27] (Camera) are a few other commercial systems that require infrastructure support. Among
the fingerprinting class, there are Ekahau [28] and GiPStech [29] which operate by building the Wi-Fi and
magnetic fingerprints’ map over the localization area,
In this article, we mainly concentrate on the fingerprinting techniques because of their promise towards
solving the localization problem inexpensively. We first
provide an overview of the traditional fingerprinting research, and its drawbacks. Then we point out the current research trends which try to address and resolve
these shortcomings. Fingerprinting solutions generally
require a pre-deployment site survey procedure during
which a radio-map is constructed by laboriously collecting signal strength samples (e.g., Wi-Fi) over the
2. Location Fingerprinting Techniques
Over the past decade, there had been a growing interest in indoor localization techniques that rely on
in-building communications infrastructure (e.g., Wi-Fi,
Bluetooth, etc.) mainly because it allows the design
of an easily deployable low-cost positioning system.
RADAR [8] opened the door to utilize Wi-Fi communications infrastructure widely ubiquitous in the residential or commercial buildings by using location fingerprinting techniques. Even though these fingerprinting
techniques offer coarser accuracy and precision compared to its infrastructure based counterparts (e.g., ActiveBAT [5], Cricket [6], etc.), its advantage lies in using
the existing infrastructure and off-the-shelf hardware in
providing the positioning service. Subsequently, a lot of
similar research followed suit in order to achieve finer
Hossain et al. / Computer Communications 00 (2015) 1–18
AP 1
AP 2
gerprinting based localization algorithms. For example,
Nearest Neighbor (NN) is used for the above example,
and its variants kNN and weighted kNN have been utilized in [8] and [30], respectively. Among other algorithms, the probabilistic Maximum Likelihood Estimator (MLE) [9, 10, 11, 30, 31, 32] is quite popular, while
factor graphs [33], kernel estimation [34, 35], neural
network [36, 37], support vector machine (SVM) [30,
38], and extreme learning machine [39] have been applied as well.
The fingerprint’s representation has also attracted significant attention as well. From the simple average
RSS [8] which has been used in our example in Table 1, to Gaussian model [32, 40] or other complex
distributions [10, 41], and even histogram representation [11, 13] of RSS at a particular location – all have
been investigated. Moreover, because of RSS’s variability among different client devices under the same
wireless condition, other fingerprints like SSD [42],
HLF [43], DIFF [44], ordered sequence of RSSs [19]
have also emerged.
Based on the operating principle of a fingerprinting
technique, its training phase calibration of radio map
comes with various challenges, e.g.,
Training Device
Middle (Room)
Left (Room)
Right (Room)
(a) Training phase.
AP 1
AP 2
Client Device
Middle (Room)
Left (Room)
Right (Room)
(b) Location estimation phase.
Figure 1: Schematic diagram of location fingerprinting
technique inside a 3 room space of interest with 2 APs.
accuracy and precision, but at the same time retaining
the inherent benefits of such techniques.
Most of these approaches utilize location fingerprinting techniques [8, 10, 9, 28], that makes use of the
already available infrastructure but entail a laborious
training phase in order to construct the radio-map. Fingerprinting based systems generally have two phases
– offline training phase and online location estimation
phase. During the offline phase, the location fingerprints
(i.e., signal strength samples) at the selected locations of
interest are collected, yielding the so-called radio map.
During the online location estimation phase, the signal
strength samples (e.g., received signal strength (RSS)
perceived at the access points (APs) from the client device, or vice versa, will be sent to a positioning engine.
This engine will then compare the observed fingerprint
with the previously collected radio-map, and will return
the corresponding location with which the fingerprint
produces the best match. A simplistic diagram explaining the fingerprinting technique is shown in Fig. 1. After conducting the site survey inside the space of interest
(i.e., the three rooms) of Fig. 1, its radio map may look
Left (Room)
Middle (Room)
Right (Room)
• It is time consuming, labor intensive, and vulnerable to environmental changes. The process also requires certain pedigree on the part of the surveyor.
• The accuracy and precision offered by existing fingerprinting solutions need to be robust in the face
of unforeseen environmental circumstances. For
example, an average localization error of 2 ∼ 3 m
may locate a user on either side of a dividing wall
depending on the time or its surroundings. Therefore, apart from achieving reasonable localization
error, a 100% room-level accuracy is also desired
for indoor environment at all times or settings.
RSS (in dBm)
AP 1
AP 2
• An implicit assumption of similar wireless conditions is held during both the training and online location estimation phases (even in our example in
Fig. 1), which may not hold in real scenarios due to
movement of furniture, adding or deleting of APs,
density of crowds and the associated interference
caused by their devices, etc. Therefore, an online
update phase is necessary to reflect the real settings
into the training radio-map database.
Table 1: An example radio map for the scenario of Fig. 1
During the online phase of Fig. 1b, if the client device perceives RSSs −92 and −51 dBm from AP 1
and 2, respectively, then its location can be resolved
as the “Right” room since it has the closet match (see
the radio map of Table 1). This is the main operating
principle of the fingerprinting based solutions. Generally, pattern matching techniques are utilized for fin-
• The labor intensive training phase may not be scalable inside big buildings or commercial indoor environment.
Hossain et al. / Computer Communications 00 (2015) 1–18
• Finally, device heterogeneity has been observed to
cause significant anomalies in some fingerprints
(especially RSS) across different client devices
even under the same wireless conditions [32, 42,
43, 44, 45]. Since a positioning system cannot
assume that its users carry the same device with
which the training phase has been carried out, this
issue needs to be addressed as well.
a numeric value, precision gives a measure how consistently the accuracy can be achieved. Cumulative distribution function (CDF) [48, 49, 42] is usually utilized
to show the overall performance of a positioning system. For example, a median of 3m in the CDF graph
indicates that 50% of the occasions, the system’s localization error (or accuracy) is within 3m.
3.1.2. Scalability
In general, a positioning system needs to scale w.r.t.
two parameters: i) geography and ii) density. Most
works in the literature report their experimental results
within a limited scope both in terms of geographical
area, and the number of mobile devices [1]. A system
will be termed as scalable only when it is able to perform the same when deployed inside a large testbed offering its service to a lot of client devices as it may have
been shown to perform in its limited scope.
A newer solution is thus required that is free from
the aforementioned drawbacks. Consequently, seeking
a calibration-free system has become the recent research
trend among fingerprinting based techniques that retain
the advantage of using existing infrastructure and offthe-shelf hardware.
3. Performance Comparison Criteria
In this section, we introduce the performance comparison criteria in order to evaluate the various calibration free IPSs found in the literature. These performance
comparison properties have largely been divided into
two main categories, namely the traditional ones, and
the calibration-free ones. Apart from the traditional performance comparison criteria like accuracy, precision,
robustness, scalability, cost, complexity, and latency,
based on which the localization techniques were evaluated, the newer approaches warrant some additional
ones. For example, whether they require an actual geographical map of the localization area, the percentage of occasional location fix needed to ensure reasonable accuracy, the need for explicit/implicit user participation to construct the radio map, the usage of building landmarks (e.g., assumption of knowledge of some
known locations) or additional sensors (e.g., accelerometer, gyroscope, compass, etc.), whether they address
device heterogeneity, etc. These comparison criteria are
quite important in evaluating the calibration-free newer
techniques since the traditional ones alone cannot completely characterize the approaches’ performances. In
the following, we elaborately discuss both types of performance comparison criteria.
3.1.3. Robustness
Robustness is the property that ensures the positioning system to offer location services in the face of unforeseen circumstances, e.g., malfunctioning of APs or
mobile devices, surrounding changes, inclusion or exclusion of newer components inside the positioning system (e.g., shutdown of an active AP), etc [47]. In other
words, the positioning system should still be operative,
however, it may provide coarser accuracy compared to
the previous ideal scenario.
3.1.4. Security and Privacy
A secure positioning system is not vulnerable to attacks from adversaries, and privacy ensures the confidentiality of location data [1, 47]. As indicated in [50],
security and privacy can be maintained from the system
architecture side, while a client-based positioning system (i.e., the client software computes its own location)
can also easily ensure privacy of location data [51].
3.1.5. Cost and Complexity
The infrastructure based positioning systems require
complex and expensive sensors or transceivers to be installed over the whole localization area [4, 5, 6]. The
fingerprinting solutions are generally cost-effective because they reuse the existing communication infrastructure such as WLAN [9, 8, 10, 28]. However, they may
require professional engineers to carry out the training phase to build the radio-map. Another aspect of
complexity is the running time of the positioning algorithm. A computationally fast and feasible algorithm
will be more attractive to serve many location queries
3.1. Traditional Performance Comparison Criteria
3.1.1. Accuracy and Precision
Accuracy (or localization error) and precision are the
two most important performance metrics of a positioning system. Localization error is generally defined as
the Euclidean distance between the actual and estimated
location [1, 2, 46, 47]. While accuracy is represented by
Hossain et al. / Computer Communications 00 (2015) 1–18
from the clients simultaneously. This is important for
client based positioning system as well where the client
computes its own location because of its limited processing and battery capability.
attractive compared to a time-intensive one especially
for client-based IPSs with limited resources.
Latency also comprises of another important component for fingerprinting solutions, i.e., how fast the training part of the positioning algorithm can be completed
using the collected database. Machine learning techniques, such as neural networks [36, 64, 65] or support
vector machines (SVMs) [30, 38] or genetic algorithms
(GAs) [49] may take several hours to even days to converge to solution depending on the training database
size. This has a significant impact on the performance
of the IPS because it has the risk of running the system
built upon a stale database for some period.
3.1.6. Technology
The technology upon which the positioning algorithm is designed plays an important role in the system’s
widespread availability or deployability. For example,
a WLAN (e.g., RADAR [8], Horus [10], Ekahau [28],
etc.) or Bluetooth (e.g., [52, 42]) based system is more
attractive from commissioning purpose. This is because
they can be tested easily, the positioning service can be
made available to a greater user-base, and overall the
system could be deployed seamlessly utilizing the already existing infrastructure. On the contrary, IR [4],
Ultrasound [22], UWB [23, 53], Audible sound [54],
Camera [27] generally require some degree of infrastructure to be setup over the localization area, and customized badges/sensors to be worn by the personnel to
be located, which ultimately might make their deployment less attractive. Many systems even try to combine two or three technologies together to build a hybrid
positioning system, thereby merging their advantages.
For example, Radianse [55] uses RFID+Wi-Fi, Active
BAT [5], Cricket [6] and HP Lab’s Smart-LOCUS [56]
use RF+ultrasound signals, Place Lab [57] and Skyhook [58] use GPS+GSM+Wi-Fi, SurroundSense [59]
utilizes ambient features such as acceleration, sound,
light, color together with Wi-Fi signals, etc. The enhanced localization solution (ELS) [60] uses inertial
sensors integrated inside smartphone and applies human
mobility modelling (HMM), and machine learning techniques. It falls back to using standard location tracking techniques (e.g., GPS+GSM+Wi-Fi) only when the
HMM and machine learning approach fails to provide
location information.
3.2. Calibration-free Performance Comparison Criteria
3.2.1. Map Requirement
A typical fingerprint based localization scheme so far
assumed the availability of an indoor map [9, 8, 10, 28].
During site survey, the laboriously captured signal signatures are annotated with their recorded locations inside the map, and then stored inside the database as a
<location, signal fingerprint> tuple by the surveyor.
It is a reasonable assumption for the traditional
schemes to assume the availability of the map because
of the collection procedure being carried out inside
a controlled environment. However, this assumption
does not hold anymore for the calibration-free schemes
where any layman with a wireless device may contribute
to the training phase being oblivious of the whole procedure [49, 66]. This layman may not even be a regular occupant (may just be a visitor) of this indoor environment. Therefore, those calibration-free solutions
that can operate without the map requirement assumption offer more practicality than the others.
3.2.2. Acquiring Location Fix
Some calibration-free schemes may require the occasional location fix from user devices collected via GPS
or other means in order to offer reasonable accuracy [49]
while others do not [67]. Initial accurate location fixes
are quite important for the overall performance of the
system. An IPS that is capable of providing comparable
accuracy similar to other approaches without the need
of any location fix is certainly more attractive.
3.1.7. Latency
Latency generally quantifies the responsiveness of an
IPS. The faster an IPS can provide a user with location
information, the better is the responsiveness of the system. For example, an inquiry based Bluetooth IPS introduced a delay of at least 10.24 seconds which was a
standard period to discover the classical Bluetooth devices in range [61]. Improvements to the classical approach [62] and newer Bluetooth Low Power (BLE)
technology offer faster discovery phase; however it may
still require a relatively large inquiry window to discover all the devices in the vicinity [63]. A computationally faster positioning algorithm will also be more
3.2.3. Seamless User Participation
The key idea behind the calibration-free schemes is
to involve users to participate implicitly in order to construct the training database. Any user carrying a wireless device may be expected to contribute to the radiomap construction after agreeing to some initial terms
Hossain et al. / Computer Communications 00 (2015) 1–18
and conditions, or without even being informed about
it. This seamless user participation is more attractive
compared to the scenario where the user explicitly inputs location fingerprint data as a feedback to the system
(e.g., [68, 69, 70, 71]). In other words, the lesser the requirement of knowledge on part of the user in building
the system, the more desirable the system is.
AP 1
AP 2
AP 3
inter−AP measurements
Client perceived RSS vector = [ −58, −88, −74 ]
Figure 2: In TIX and SDM, inter-AP RSS measurements are used in order to infer distances between the
target and each AP.
4. Site-survey-free Indoor Positioning Systems
3.2.4. Usage of Indoor Landmarks
Many of the previous fingerprinting techniques assumed the knowledge of Wi-Fi APs’ locations [31] or
even placed sniffers at known positions [72, 45]. These
locations termed as indoor landmarks are required to be
known a priori for the positioning system to operate. For
example, a model-based localization scheme first estimates the propagation parameters (e.g., path-loss exponent) during the training phase [8, 6]. Subsequently, it
approximates the distance of the target during localization from at least three known landmarks, and applies
trilateration to find its position. One of the motivation
of the newer approaches is the ability to seamlessly integrate the system preferably anywhere without any a priori knowledge about the layout of the deployment area
or its components (e.g., APs’ locations).
In this article, we provide an insightful summary of
the state-of-the-art IPS that does not require any site
survey process. We select some interesting research
in this section that comes close to achieving this property, and also touch upon their relative advantages and
disadvantages. Finally, the performance of each discussed IPS is outlined according to both traditional and
calibration-free performance comparison criteria inside
Table 3 and 4, respectively. A list of common notations
that are used inside the description appears in Table 2.
di j
i, j, k
3.2.5. Need for Additional Sensors
SurroundSense [59] opened the door for various ambient sensors to be collectively used for localization, and
subsequently many works followed suit [73, 74, 66].
Most of these works assume the availability of these
sensors in the mobile phones carried by the crowdsourcing users. Even if the newer smartphones come with
most of these sensors (e.g., accelerometer, gyroscope,
compass, etc.) built-in, an approach that has a wide
range of sensor requirements may be restricted in its applicability, or may result in lower user-base.
number of training locations
number of APs
geographical distance between location i and j
observed RSS at location i from AP k
observed RSS vector at location i, i.e., [si1 , si2 . . . sim ]
used as index variables
Table 2: A list of the common notations used in the
state-of-the-art literature description in Section 4.
4.1. TIX
TIX [75] works only with on-line Wi-Fi RSS measurements, thereby getting rid of the laborious sitesurvey process. After receiving the location query from
a client, TIX’s location server sends request to both the
client, and all the APs to retrieve the RSSs. For example, if there are m APs, then each AP will report the
other (m − 1) APs’ RSSs, and the client will report m
RSS measurements where all APs have coverage over
the whole localization area. With the assumption that
RSS (log-scale) decays linearly with distance, a linear
RSS-to-distance mapping can be obtained at each AP
for the other (m − 1) APs. Then the client uses the approximated mapping curves of the AP from which it
is receiving the strongest signal to obtain its distance
to the other (m − 1) APs. For its distance approximation to the strongest AP, it uses the mapping curve from
the second strongest AP. For the example scenario of
Fig. 2, AP1 ’s constructed mapping curves of AP2 and
AP3 will be used to approximate the distances between
the client and the two APs whereas AP3 ’s constructed
mapping curve of AP1 will be used for inferring the distance from AP1 . Thereafter, the final location estimate
3.2.6. Addressing Device Heterogeneity
Device heterogeneity has been identified as a cause
for affecting the localization accuracy in [45, 32, 43,
42, 44]. For example, some RF fingerprints tend to
vary significantly with the device’s hardware under the
same wireless conditions. Consequently, fingerprints
collected by different devices tend to be quite different from one another. The newer calibration-free techniques are expected to incorporate all the users possibly
carrying heterogeneous devices even during the training
phase. Therefore, the positioning system should be designed in such a way that it accounts for the fingerprint
discrepancy caused by the heterogeneous devices that
the user may carry.
Hossain et al. / Computer Communications 00 (2015) 1–18
is obtained using the approximated distances between
the client and APs, and the APs’ location coordinates.
For this, TIX utilizes Triangular Interpolation and eXtrapolation (TIX) algorithm [75]; other lateration based
algorithms (e.g., triangulation) could also be used.
TIX achieved 5.4m localization accuracy with zero
calibration effort inside an office environment with an
area of 1020m2 . TIX does not require the floor plan information to be operational, but it requires AP’s location
information. TIX may also warrant modifications of a
typical AP’s operations in the form of receiving and replying RSS queries from/to the location server.
During the location determination phase, a client j
first measures the RSSs from its neighboring APs, and
retrieves the coefficients bi ’s associated with each AP
i. It then computes the geographical distances to the
neighboring APs as d j = exp(Bs j ). Thereafter, lateration technique is used to compute the client’s final location estimate using the distance approximations from
the APs.
SDM does not require the floor plan map to operate, but it needs to know the locations of the APs. Its
localization accuracy is within 3m inside a small academic departmental building, and performs better than
TIX [75]. It may require a little pre-deployment effort
though unlike TIX in the form of AP’s RSS to distance
mapping construction phase. It may also warrant modifications of a typical AP’s operations, or additional deployments of monitor/sniffer elements.
4.2. SDM
Signal-distance map technique, SDM [76] develops a
localization algorithm that only utilizes the on-line RSS
measurements to locate the client device. No laborious site survey process to construct the radio map is
necessary. Inter-AP RSS measurements are made periodically to automatically calibrate RSSs in the spatiotemporal domain. A truncated singular value decomposition (SVD) technique is then applied to map the relationship of the RSS measurements and the geographical
distances to the APs where the APs’ locations are assumed to be known. The goal of this on-line mapping
is to mitigate the adverse effect of the measurement error, and retain as much recent environmental information as possible by getting rid of the training phase. The
least square method is used with the following objective
function to minimize:
ei =
4.3. OIL
The organic indoor location (OIL) [77] merges the
“training” and “use” phases of a typical fingerprint
based localization. In other words, it involves the users
of the location system to also contribute to build the radio map simultaneously. Since OIL runs as a daemon
process on the client providing localization service, it
determines on its own to prompt the user for giving
feedback. It may also discard the feedback based on a
filtering mechanism. For location determination, it uses
the probabilistic maximum likelihood estimator (MLE).
OIL uses voronoi region to map spatial uncertainty of
its location estimate, and prompts the user for explicit
location feedback if the estimation confidence falls below a threshold. The user will only be prompted if
his/her feedback is likely to either increase the system’s coverage or improves the input location’s accuracy. During the prompt, only the voronoi regions which
meet a certain similarity metric w.r.t. observed APs
(e.g., the number of matched MACs between the candidate and the client that provided fingerprint) are included in the feedback choices. The partial map with
those associated voronoi regions only need to be made
available to the client, thereby not overwhelming it in
terms of computation and storage resources. Additionally, it handles the erroneous user feedback by outlier detection in signal space, i.e., the observed RSSs
for each AP. It uses agglomerative hierarchical clustering approach to group the user feedback by similarity.
The distance between inter-cluster (Ci and C j ) feedback
pairs is represented as:
dS (si , s j )
DS (Ci , C j ) =
|Ci ||C j | (s ,s )∈(C ,C )
(log(dik ) − bi sik ),
where dik is the distance between ith AP and each AP
k (dii = 0), and sik ∈ S, k = 1, 2, . . . , m, are the perceived RSSs of the ith AP from every other AP. Consequently, the coefficient vector can be solved as, bi =
log(di T )ST (SST ) . The SDM can then be expressed
as, B = log(D)ST (SST ) where D = [d1 , d2 , . . . , dm ]
and di = [di1 , di2 , . . . , dim ]T are the geographic distance
matrix and vector, respectively. Its working principle
is almost similar to TIX discussed above which also
uses inter-AP measurements to obtain mapping functions. However, its algorithmic operations is quite different compared to TIX. In SDM, the coefficient bi associated with a particular AP i is computed by taking
into account all of ith AP’s measurements collected inside the rest (m − 1) APs. On the contrary, the ith AP’s
mapping curve co-efficient will be different at each AP
considering only its collected measurements at that particular AP for TIX.
Hossain et al. / Computer Communications 00 (2015) 1–18
2 2
where dS (si , s j ) = [ M1 m
k=1 (sik − s jk ) ] is the dissimilarity metric between two feedback RSSs si and s j .
M ≤ m is the subset of total m APs for which its RSS
appears in either si or s j . If only one of them appears,
the missing value is replaced by −100 dBm. A correct
cluster Cl∗ given a feedback at location l is identified as:
Cl∗ = arg min
DS (C, Ck∗ )
works according to the principle that given enough RSS
measurements (i.e., mn > 4m + 2n), the system will be
uniquely solvable. EZ uses Genetic Algorithm (GA) to
come up with the solution, i.e., a vector of all the unknown values to be solved in the LDPL equations. Gradient Descent (GD) is also used inside GA for refinement purpose. As their GA progresses, solutions with
higher fitness evolve, and it terminates if there is no improvement for ten consecutive generations.
During the execution of the GA, first the parameters
of the APs are determined with as many known locations as possible. Then the unknown locations from
which at least three of these determined APs can be
heard are solved via triangulation. These determined
locations in turn result in more APs’ parameters to be
solved. This process continues until the vector of all the
unknowns of APs and locations have been determined.
Various techniques are also adopted in order to reduce
the search space of the GA in achieving this purpose,
i.e., if an AP is heard from five or more determined locations, its parameters can be uniquely solved. Again,
if it is heard only from four determined locations, there
are two possible solutions for it, and so on. The indoor
RSS model is built offline using the GA that may even
take hours to converge depending on the specifics of the
indoor space and the amount of measured data. Subsequently, the location queries from the EZ clients are
answered by the EZ server in real time using this built
EZ also addresses some additional inherent research
challenges that come with this automated radio-map
building process, e.g., choosing the right set of APs
that tries to achieve the maximum entropy of the selected APs’ collective information, or the right subset
of measurement data that are picked in a similar fashion. EZ also tackles the device heterogeneity issue that
might arise because of the involvement of different users
(thereby different mobile devices) in building the radiomap. EZ introduces a receiver gain parameter in the
LDPL equation (1), that varies with mobile devices’
hardware. It eliminates the effect of receiver gains by
subtracting one mobile device’s RSS from another corresponding to an AP. It tries to cluster the similar difference readings with the assumption that probably they
are taken at the same location or at locations close to
each other. Thereafter, the relative receiver gains w.r.t. a
random device’s gain can be computed using these measurements.
EZ’s no pre-deployment site survey effort comes at
the cost of some accuracy loss compared to traditional
fingerprinting techniques such as RADAR [8] or Horus [10]. EZ does not demand the knowledge of physi-
where N(l) is the neighboring location set of l, and Ck∗ is
the correct cluster for location k during the time of computation. The localization algorithm will only incorporate the feedback if the feedback location corresponds
to the correct cluster location.
OIL requires the availability of a map, and active participation on the part of the users. In order to build a
responsive system for user feedback, the client maintains a local cache of partial fingerprint database which
affects OIL’s scalability.
4.4. EZ
EZ [49] relaxes the requirement of prior knowledge
of indoor RF environment, that would have been laboriously captured during the site survey process of a traditional fingerprinting technique. The received signal
strength (RSS) measurements are implicitly reported by
EZ clients running inside a user’s mobile device without any human intervention. Such automated operation brings forth a number of challenges that have been
addressed in [49], e.g., extracting only the useful measurements, and efficiently building the indoor RF model
based on these measurements.
EZ’s built RF model’s core lies in the log-distance
path loss (LDPL) formula:
where dik
si0 − 10γi log dik ,
(xk − ci )T (xk − ci ).
Eq. (1) represents the RSS of the kth mobile user’s device located at a distance dik from the ith AP. Suppose
xk and ci denotes the mobile user and AP’s location in
2D, and γi is the path loss exponent concerning ith AP’s
signals. With the RSS measurements collected implicitly from user devices, EZ tries to build the radio-map of
the indoor environment. Each RSS measurement from
an AP at a particular location is fitted into (1) where
the AP’s (γi , ci , si0 ), and the location, xk are unknown
parameters. If there are m APs and n locations where
each AP is seen, then the total number of LDPL equations will be mn, and the total number of unknowns is
4m + 2n where the location is represented in 2D. EZ
Hossain et al. / Computer Communications 00 (2015) 1–18
cal layout (i.e., map) of the indoor environment or even
the location and transmit power information of the APs
to be known. This particular feature together with no
pre-deployment effort make it a feasible localization solution inside indoor settings such as malls, and multitenant commercial buildings where different APs might
be deployed/managed by different entities. However,
EZ works upon the assumption that the users carry WiFi equipped gadgets over the indoor environment that
has excellent Wi-Fi coverage throughout, and also the
availability of occasional location fix, e.g., GPS lock at
the edges of the indoor setting.
cation estimate:
j∗ = arg max
p(x j = 1, zk = 1|sobs )
WiGEM follows traditional model-based fingerprinting techniques in its working, but takes into account the
client device heterogeneity and their different transmit
power level issues. It is an important aspect for WiGEM
to consider because of its infrastructure based approach.
Furthermore, WiGEM assumes the RSS to follow Gaussian distribution which is received with mixed reactions
in the literature. It also requires the availability of a
map, and the sniffers’ locations to be known.
4.5. WiGEM
4.6. WILL
WiGEM [78], a wireless localization algorithm using Gaussian mixture model (GMM) with expectation
maximization utilizes model based parameter estimation, thereby relieving the site survey process. It then
applies maximum a posteriori estimation algorithm on
the created model in order to locate the client.
This infrastructure-based learning oriented approach
requires sniffers at known positions in order to collect
Wi-Fi RSS measurements from the client. The RSSs at
the sniffers are then represented by GMM:
p(s) =
WILL [67], a wireless indoor logical localization approach relieves the site survey process in constructing
the training database while providing comparable accuracy and precision. WILL utilizes the commonly seen
property that the RSS goes through significant change
through a wall, and an accelerometer’s reading can be
used to detect whether a user is moving or stationary.
WILL captures the user traces as a series of < F, A >
values, where F and A indicate the Wi-Fi signal fingerprint and accelerometer values, respectively. Depending
on the similarity of the fingerprints, they are grouped
inside the same virtual room or a different one. Subsequently, a logical floor plan is conceived which is a
diagram depicting the reachability among virtual rooms
utilizing the user traces. Some filtering techniques are
also adopted where traces with few steps or APs are
discarded. The actual physical floor plan is also converted to a graph where an edge between two nodes (i.e.,
rooms) indicates the reachability between them. The
logical floor plan is then mapped onto the physical one
utilizing some vertex matching techniques between two
graphs, e.g., skeleton mapping, and branch-knot mapping [67, 79]. Redundant information like neighboring information of the mapped vertices are also utilized
to correct the mapping errors of the mapping process.
Once completed, the location queries can then be answered by the localization engine which uses the fingerprint database to first localize the virtual room, and
then obtains the corresponding physical room from the
WILL provides 86% room-level accuracy without the
requirement of measurement feedback from known locations, or the knowledge of APs’ locations. However,
it requires the map of the space of interest for the virtual
to physical room mapping process.
v j τk N(s|µ( j,k) , σ2(j,k) ),
j=1 k=1
where v j is the probability of the client being at location
j, and τk is the probability of transmit power level being
k, and s is the RSS vector measured at the deployed sniffers. µ( j,k) and σ( j,k) are the mean and standard deviation
of the measured RSSs given the client is stationed at location j with transmit power level k. The parameters of
the GMM are θ = (v, τ, µ, σ).
The area of interest is first divided into J grid locations, and v and τ are assumed to be uniformly distributed over all possible locations, and power levels,
respectively. Initial µ is computed using Eq. (1) for each
j ∈ J. This is possible since the sniffers’ locations are
known. σ’s value is fixed at 5 for all sniffers. Given one
RSS vector, s, the GMM model parameters are then updated using expectation maximization technique which
will eventually converge by associating the RSS vector
with a particular location of the client, j.
Real-time localization is then performed by first finding the probability for each (location, power-level) pair
given an observation RSS sobs , and then marginalizing
it over the power levels. The location with the highest
probability will then be returned as the client’s final lo9
Hossain et al. / Computer Communications 00 (2015) 1–18
4.7. UnLoc
4.8. Zee
Zee [73] utilizes inertial sensor measurements of
a user’s smartphone (e.g., gyroscope, compass, accelerometer, etc.) to track his/her movement/direction,
and subsequently annotates the locations traversed with
the Wi-Fi RSS measurements. In other words, Zee uses
Wi-Fi and inertial sensor readings crowdsourced from
the users’ handhelds to construct the Wi-Fi training set
which was otherwise being laboriously constructed by a
site survey process.
Zee works according to the following principle – a
user’s traveled distance and direction might be obtained
from accelerometer and compass/gyroscope data. Now,
with the help of a map, certain constraints can be applied to the user’s traversed path which may ultimately
result in revealing the user’s final location. For example,
according to the map, only one pathway inside the area
may accommodate the user’s traveled path at a certain
time. Consequently, we can also infer the user’s starting
location, and thereby can annotate the whole pathway
locations with measured Wi-Fi data. Zee has two main
UnLoc [66], an unsupervised indoor localization
scheme, utilizes urban dead-reckoning [80] to track a
user while recalibrating its estimate whenever it encounters a landmark (e.g., an entity with known location).
The landmarks are categorized into two: i) seed landmarks (SLMs) are the elevators, escalators, stairs, entrances etc., that may exhibit distinct signatures in the
inertial sensory dimension, and ii) organic landmarks
(OLM) are small confined geographical areas that exhibit distinct patterns from many sensed signals (Wi-Fi,
inertial sensors, etc.). UnLoc identifies both types of
landmarks in an automated manner during the crowdsourcing process. For example, a Finite State Machine
(FSM) is used on the accelerometer readings to separate
the various SLMs, while an OLM is identified based on
two different locations’ similarity metric, λ ∈ {0, 1}:
1 X min(s1k , s2k )
|m| ∀k∈m max(s1k , s2k )
where s1k and s2k denote the RSSs of AP k ∈ m at location l1 and l2 respectively. The rationale behind (2) is
to add proportionally larger weights to λ when an AP’s
signal is strong at both l1 and l2 , and vice versa. (2)
helps to identify Wi-Fi based OLMs in a way that all
locations within the Wi-Fi landmark exhibit lower similarity (e.g., λ < 0.4) with the rest. Moreover, all estimated locations within the Wi-Fi based OLM must be
confined inside a small region (4m2 ) to be considered as
a single landmark. Similar technique has been adopted
for identifying other inertial sensor based OLMs.
UnLoc works in the following manner – the mobile
user starts the itinerary whenever he/she encounters the
first landmark (e.g., entrance), and subsequently resets
the position whenever other landmarks are sensed. The
dead-reckoning approach is used to track him/her between landmarks. The locations of the landmarks are
made more accurate as more user traces become available which in turn improve dead-reckoning tracking of
subsequent handhelds. This recursive process continues
to improve UnLoc’s offered accuracy over time, while
the first few crowdsourcing users might suffer from inferior accuracy.
UnLoc does not require the floorplan to operate, but
it needs only one seed landmark’s location during bootstrapping phase (e.g., the entrance) to be treated as origin for the space of interest. UnLoc demonstrates a median error of 1.69m across three different indoor settings
including a shopping mall; however they conducted all
the experiments with a single mobile device.
• Placement Independent Motion Estimator (PIME)
identifies whether the user is walking or not; irrespective of where the handheld might be placed
(e.g., front/back pockets, bags, etc.), and also estimates the step count and heading offset (HO).
• Augmented Particle Filter (APF) maintains a four
dimensional joint probability distribution of the
user’s location (x, y), stride length, and HO along
his/her traversed path.
The APF then runs belief back-propagation (with map
constraints) to correct the user’s path history as described before. This yields a time-indexed sequence of
the user’s estimated locations that can be annotated with
the measured Wi-Fi RSSs. As the training database gets
filled with more Wi-Fi measurements, the APF may obtain a more confined belief of the user’s initial location
distribution (for the first user, it would be uniformly distributed over the whole space of interest) by comparing
the Wi-Fi scans with the existing database. The WiFi measurements obtained from subsequent walk are in
turn used to refine the existing database.
Zee is not a localization algorithm, rather an innovative means of constructing the training radio-map from
crowdsourcing which could be used by any location fingerprinting techniques. With Zee’s crowdsourced data,
Horus [10] and EZ [49] were seen to achieve a median
localization error of 3m, comparable to the performance
achieved by using the typical site surveyed training data.
Hossain et al. / Computer Communications 00 (2015) 1–18
Zee does not require any location fix (e.g., GPS lock),
and is also free from requirement of knowledge of the
APs’ locations. However, it requires the availability of
a map of the space of interest during the crowdsourcing
LiFS offers 89% room-level accuracy (average localization error 5.8m) inside a medium sized academic
building. It does not need to know the APs’ locations,
but requires map to transform it into a stress free floor
plan with associated fingerprints.
4.9. LiFS
4.10. Walkie-Markie
Locating in fingerprint space (LiFS) [74] is an indoor
localization system from the same group who invented
WILL discussed in Section 4.6. LiFS follows similar
principle as WILL [67], and also uses off-the-shelf WiFi infrastructure, and inertial sensors (accelerometer) of
mobile devices in order to locate a target. LiFS training
phase is divided into three main tasks: i) transforming
the map into a stress free floor plan, ii) creating fingerprint space, and iii) mapping between fingerprint and
real location. The client’s location queries are then answered with the closest location corresponding to the
observed fingerprint (nearest neighbor); even though
other searching algorithms can easily be adopted.
The stress free floor plan is created by first sampling
the real floor plan into grids, where two consecutive
grids’ separation is set at 2m. By calculating walking
distances between all sampled locations (via accelerometer readings of user traces), a distance matrix D = [di j ]
can be obtained between every two locations i and j. In
a stress free floor plan, the Euclidean distance between a
pair of points reflects the walking distance of their corresponding locations in a real floor plan. Fingerprints are
recorded during a normal user’s itinerary together with
the walking distances (via computing steps multiplied
by stride length) between two consecutive recorded fingerprints. These fingerprints are pre-processed to filter out the multiple similar fingerprints taken possibly at
the same positions. The fingerprint space is then constructed by connecting the fingerprints collected with
the distance information using Floyd-Warshall shortest
path algorithm [81]. Subsequently, the mapping between each fingerprint with a real location can be performed utilizing the spatial similarity between the constructed stress-free floor plan and the fingerprint space.
Similar to WILL [67], betweenness centrality and kMeans algorithm are used to identify the fingerprints
along the corridors and rooms, respectively. Thereafter,
fingerprints observed at doors (named reference points)
are utilized to match the different clusters (i.e., corridors and rooms), thereby establishing the correspondence between the stress free floor plan and the fingerprint space. Subsequently, the points in each cluster are
mapped to sample locations in its corresponding rooms
by choosing the nearest neighbor for each point.
Walkie-Markie [82] generates indoor pathway maps
from crowdsourcing user traces without any a priori
knowledge about the space of interest (e.g., map, propagation characteristics, etc.). They utilize RSS trend (increasing or decreasing) rather than its absolute values by
identifying Wi-Fi marks inside the building where RSS
trend tripping point (RTTP) (i.e., maximum) occur for
an AP. They argue that using RTTP’s location as WiFi marks is free from mobile device heterogeneity, and
other environment factors affecting absolute RSS. These
Wi-Fi marks are then placed inside the 2D plane by their
graph embedding algorithm utilizing user trajectories.
On one hand, the created pathway maps leverages the
localization of users when they pass a Wi-Fi mark, and
utilizes dead-reckoning to localize him/her in between
two Wi-Fi marks. On the other hand, the pathway maps
can also be utilized to create the complete radio map of
the space of interest, thereby opening the door for traditional fingerprinting techniques to be used on top of
Wi-Fi mark is identified by {BSSID, (D1 , D2 ), N}
where BSSID is the AP’s MAC, D1 and D2 are the
steady walking directions approaching and leaving the
RTTP, respectively. Furthermore, the neighbor information N of the AP is used to resolve ambiguity that
might occur in case of parallel corridors or similar turning styles. Walkie-Markie then uses the direction information together with distance information retrieved
from user trajectories in order to place the Wi-Fi marks
at real locations using their novel “Arturia” positioning
algorithm. Due to the possibility of Wi-Fi marks being
reported differently (i.e., at different locations) by multiple crowdsourced trajectories, a clustering algorithm
is used in order to cluster the Wi-Fi marks with slight
deviations into a single one. Utilizing the placed WiFi marks’ locations together with user trajectories, an
expansion and shrinking process is then applied to conceptualize the whole pathway map.
Walkie-Markie’s average localization error is 1.65m
outperforming RADAR (2.3m) in the same office floor
testbed of area 3600m2 . It neither requires the knowledge of the floor plan nor needs to know the locations
of the APs.
Hossain et al. / Computer Communications 00 (2015) 1–18
5. Discussion
ized without any a priori knowledge in the newer systems like UnLoc and Walkie-Markie. The recent systems (WILL, UnLoc, Zee, LiFS, Walkie-Markie) also
require additional sensors on the client devices besides
Wi-Fi to be operational which is argued to be quite commonplace for today’s handhelds. Very few systems like
EZ, WILL, Walkie-Markie address the client device heterogeneity issue which may have biased other systems’
accuracy or precision reporting (e.g., UnLoc and LiFS).
In this survey, we have only selected the class of
calibration-free techniques that utilizes the ubiquitous
Wi-Fi infrastructure as the main means for localization.
Moreover, it incorporates the off-the-shelf hardware integrated with various handhelds making the system to
be an easily deployable low-cost solution. We feel this
class of systems has huge potential since it follows the
traditional fingerprinting approach’s principle that already showed promise for indoor localization.
Table 3 and 4 can help the reader to quickly identify
the appropriate calibration-free solutions that can satisfy
a application’s need. If an application’s requirements is
specified in terms of the various performance properties, e.g., accuracy/precision, cost/complexity, addressing device heterogeneity, map requirement, usage of
inertial sensors, etc., the corresponding solutions that
satisfy these requirements fully or partially can easily be identified by going through the summarized results/findings presented inside the tables. For example,
if one is interested in the class of calibration-free positioning systems that do not require any inertial sensors
but addresses the device heterogeneity, one quick look
at the last two columns of Table 4 will reveal that EZ and
WiGEM will fall into that class of positioning systems.
Another class of calibration-free techniques, namely
Simultaneous Localization and Mapping (SLAM) [84]
utilizes only the inertial sensors to iteratively construct
the indoor environment map, and localizes the user
within this map. SLAM was initially targeted at robotics
and autonomous vehicles field in order to construct a
map within an unknown environment, while at the same
time keeping track of their locations. However, it is
slowly finding its way in the calibration-free indoor localization as well where a user can be located inside a
map that has been created by someone else. It mainly
uses various filtering (e.g., Kalman filter, Particle filter, etc. [85, 86]) and dead-reckoning [80] techniques
utilizing the inertial sensor measurements to build the
map without any a priori knowledge, and pinpoint the
location inside it. A lot of variants of SLAM exist
in literature. FootSLAM [87] and ActionSLAM [88]
use foot-mounted and body-mounted inertial measurement units (IMUs) respectively to track a user’s motion;
Some of the recent calibration-free fingerprinting
IPSs discussed in the previous section are evaluated
w.r.t. performance comparison criteria discussed in Section 3 in Table 3 and 4. From the tabulated evaluation
results, one can select the IPS that may best suit his/her
location-based application’s requirements. In the following, we provide an overall discussion based on the
None of the work seems to report accuracy and precision in terms of both localization performance numbers (e.g., median or average error) and room-level correctness with the exception of LiFS which reports average localization error and room-level error to be 5.8m
and 11%, respectively. The rest either provide numbers (e.g., SDM, EZ, etc.) or the room-level accuracy
(e.g., WILL). As discussed in Section 2, both types of
performance evaluation results are important for indoor
environment. The performance results also vary according to the testbed size as reported in EZ [49]. Both
scalability and robustness issues are addressed in newer
approaches (e.g., WILL, UnLoc, Zee, LiFS, WalkieMarkie, etc.) rather than the earlier ones (e.g., TIX,
SDM, OIL, etc.). The newer systems are also quite costeffective in terms of its deployment parameters, and operational characteristics except for LiFS [74]. However,
handling security and privacy gives a different picture,
i.e., most of the newer approaches like WILL, UnLoc,
Zee, LiFS, Walkie-Markie do not address it. Most of
the approaches’ responsiveness to location queries are
fast with the exception of TIX. All of them utilize existing Wi-Fi infrastructure as the wireless technology in
their localization algorithm while a few also use inertial
sensors on top of it (e.g., UnLoc, Walkie Markie, etc.).
The map requirement parameter produced mixed observations for the calibration-free systems as seen in Table 4. The ones that required map knowledge argue
that a user needs to be shown inside a map to consume
location-based services [73]; so it does not make sense
assuming the unavailability of the map. Furthermore,
various map construction tools are already available in
the literature [83]; so the map availability assumption
can be safely made. Most of the earlier works (TIX,
SDM, OIL, or even EZ) require location fix from known
locations while the more recent ones do not. User participation is implicit inside the most recent ones’ workings. On the contrary, either explicit user participation
was generally required for earlier research (e.g., OIL) or
the user participation was ignored altogether (e.g., TIX
or SDM). TIX and SDM use APs with known locations
as landmarks whereas organic landmark is conceptual12
Hossain et al. / Computer Communications 00 (2015) 1–18
deployment in an indoor environment. We selected 10
such calibration-free schemes to discuss in detail, and
point out their advantages and disadvantages w.r.t. our
identified performance comparison properties. From
this survey, the readers will have a comprehensive understanding of the existing calibration-free techniques
in the literature, especially of the 10 IPSs discussed
elaborately. As can be seen from the discussion, each
IPS has its own design characteristics, e.g., it uses a certain kind of technology like Wi-Fi or may include the
inertial sensor measurements as well. Moreover, one
may work well compared to the other under certain constraints (e.g., with the availability of map or the lack of
it). Section 5 provides a discussion on the suitability of
a particular calibration-free IPS under certain scenarios
or requirements.
We foresee a few future research work directions in
order to enhance the prospects of widespread deployability of such techniques. For example, none of the
calibration-free research touches upon the important security and privacy issue explicitly; although the prior
research of this family (e.g., TIX, SDM and OIL) incorporates it implicitly because of its client-based location computation nature. Starting from EZ [49], the
crowdsensing way of eliminating the calibration phase
warrants the tackling of security and privacy issue even
more which has not been addressed. Even though the
user participation is implicit, a malicious user can easily tamper with the system by providing incorrect measurements, thereby delaying the convergence of the system, and also degrading its offered accuracy. Energy
efficiency is another aspect for consideration for these
crowdsensing approaches. Most of them assume the
client devices to turn on the Wi-Fi, and other inertial
sensors for obtaining the measurements, which may
drain their batteries quickly. In practical scenarios, this
might be a road block for deploying such systems. Even
though one of the motivation of such calibration-free
techniques was its fast and easily deployable quality; in
practical sense, it may still require a bit more investigation. For example, in order to offer reasonable accuracy
and precision, the participating users may need to be
uniformly distributed over the whole space of interest
which might be difficult to ensure in practice.
PlaceSLAM [89] is an improvement over FootSLAM
by explicitly involving users to input feedback about
his/her known physical world. WiFi-SLAM [90] uses
Gaussian Process Latent Variable Model (GPLVM) together with motion dynamics model to label the unlabeled signal strength data in order to perform efficient
localization. GraphSLAM [91] is an improvement over
it. Unlike other SLAMs, SmartSLAM[83] uses smartphone’s accelerometer, compass, and Wi-Fi modules to
observe the device’s movement and environment, and
thereby construct the indoor map. Although this family
of SLAM techniques is quite promising to relieve the
exhaustive calibration of a typical fingerprinting technique; their working principle is quite different from the
ones discussed in Section 4. Therefore, none of them
has been included for thorough evaluation purpose in
this article. Moreover, they may require customized inertial sensors that need to be integrated inside a moving
object specially (e.g., foot-mounted, body-mounted or
having a specific orientation, etc.) to ensure accurate
measurements. This may not be in accordance with the
property that the low-cost easily deployable IPS must
adhere to.
In terms of commercial prospect, the calibrationfree indoor positioning start-ups are recently making its
presence felt. Most of these commercial products seem
to follow the same principles adopted by the calibrationfree research that we have discussed in this article, i.e.,
they use and process the off-the-shelf inertial sensor
measurements that are readily available within a modern smartphone. Navigine [92] uses Wi-Fi signals and
the inertial sensors found in the smartphones for navigation with claimed accuracy of 1 ∼ 2m. On the contrary,
Navisens [93] only uses the inertial sensor measurements. Both of these products use filtering algorithm to
pinpoint the final location estimate. [94] relies
on a combination of dead-reckoning, sensor fusion and
partial WiFi fingerprinting for indoor navigation. All
of them try to provide SaaS (software as a service) by
giving out an SDK for the developers to program various applications on top of their location service to run
in their smartphones.
6. Conclusion and Future Work
In this article, we reviewed the calibration-free fingerprinting techniques w.r.t. both traditional and some
newer performance comparison criteria that we identify
based on the systems’ operating principles and characteristics. This newer breed of fingerprinting solutions
try to get rid of the training phase of a typical fingerprinting system that generally hinders its widespread
Yes; Even
though their
fingerprints are
collected at
Median error ∼ 3m inside
medium sized building (2275
m2 ) when used together with
EZ or Horus
89% room level accuracy
inside medium sized academic
building (1600 m2 )
Average error ∼ 1.65m inside a
medium sized office floor
(3600 m2 ); also conducted
some experiments inside a
shopping mall
UnLoc [66]
Zee [73]
LiFS [74]
WalkieMarkie [82]
Requires modified APs’
operations or additional
sniffer elements; simple
Uses off the shelf h/w;
Complex training phase
and also requires little
calibration effort to
construct the stress free
floor plan
Uses off the shelf h/w;
novel algorithm to
identify Wi-Fi marks,
their real locations and the
overall conceptualization
of the whole pathway map
Wi-Fi (also
reliant on
inertial sensors
for localizing)
Wi-Fi (also
reliant on
inertial sensors
for localizing)
Uses off the shelf h/w;
Simple techniques to
identify SLM and OLM
together with
dead-reckoning scheme
Uses off the shelf h/w;
Constrained based simple
filtering techniques
Requires sniffers or
modified APs’ operations;
complex GMM creation
and EM algorithm
Uses off the shelf h/w;
Complex mapping of
virtual floor into physical
layout with associated
Uses off the shelf h/w;
complex algorithm
Requires modified APs’
operations; simple
computationally light
Uses off-the-shelf h/w;
simple MLE algorithm
Cost & Complexity
Convergence to training database is fast;
so also the answer to location queries
using simple matching and
Construction of training database is
incremental; response to client location
queries is fast using simple pattern
matching technique
Construction of training database is fast;
training database is used with other
algorithms to answer location queries
(e.g., EZ, RADAR, etc.)
Convergence to training database is fast;
so also the answer to location queries
using simple matching and
Construction of training database is
incremental; response to client location
queries is fast using simple pattern
matching technique
WiGEM model creation can be offloaded
to cloud server machines since
infrastructure-based approach; increases
monotonically with learning samples.
Response to location queries in real-time,
but responsiveness depends on the
granularity of the location area used.
No database creation; slow responsiveness
to client queries that involves first
retrieving the RSSs from APs (inter-AP
measurements) and client, followed by
mapping function, and then location
No database creation, only linear
coefficients estimations by the APs; client
computes its location quickly using
lateration technique
Server updates client’s local cache, and
also filters out erroneous feedback; fast
response time using simple MLE by the
client itself
Because of GA, construction of EZ
database is time intensive; answer to
location queries of EZ clients in real time
Table 3: Performance of state-of-the-art site survey free research w.r.t. traditional performance comparison properties.
Median error ∼ 1.69m across
three different indoor setups
(largest being 4000 m2 )
Yes; Performs major
(occasionally) and
minor (frequently)
updates that keeps
the training database
up to date
WILL [67]
Partially; model
86% room level accuracy
inside medium sized academic
building (1600 m2 )
WiGEM [78]
No; finer
RSS reading is
Yes; client
computes its
own location
Partially; model
No; client has
to maintain
local cache of
Accuracy and precision
numbers are not reported
OIL [77]
Yes; client
computes its
own location
Yes; client
computes its
own location
Security &
Partially; linear
mapping of RSS and
log (distance)
because of AP
Average error ∼ 3m inside
small building (598 m2 )
SDM [76]
Median error ∼ 2m inside
small building (486 m2 ) and
7m inside big building (12600
m2 )
Better than other model based
techniques but poorer than
RADAR [8] and
probabilistic [10] approaches
with finer granularity inside
two testbeds - one small (600
m2 ) and another medium sized
(3250 m2 ) academic buildings
Partially; linear
mapping of RSS and
log (distance)
because of AP
Average error ∼ 5.4m inside an
office building (1020 m2 )
TIX [75]
EZ [49]
Accuracy & Precision
Hossain et al. / Computer Communications 00 (2015) 1–18
OIL [77]
EZ [49]
WiGEM [78]
WILL [67]
UnLoc [66]
Zee [73]
LiFS [74]
Yes; from APs with known
No; only one landmark’s location
with absolute location during
Yes; Wi-Fi
marks retrieved
from user traces
Mostly implicit; only
stress free floor plan
construction requires
explicit participation
Yes; accelerometer,
gyroscope & compass
Yes; accelerometer
Yes; usage of RTTP to identify
Wi-Fi marks, and also RSS
differences during clustering of
Wi-Fi marks
No; Same device used for the
Yes; accelerometer,
gyroscope & compass
N/A; produces training radio-map
via crowdsourcing to be used with
other fingerprinting techniques
Yes; accelerometer,
gyroscope & compass
No; all experiments are performed
with one type of device
Yes; accelerometer
Yes; supported with experimental
results, however explanation is
missing as to why their approach
can handle device heterogeneity
Yes; using RSS stacking difference
similar to [42, 44]
No; presented results do not verify
its claim of addressing device
Address Device Heterogeneity
Need Additional
Yes; both SLM
and OLM
retrieved from
user traces
Yes; APs treated
as landmarks
Usage of
Yes; APs treated
as landmarks
No; only one SLM’s location
needed during bootstrapping
Yes; from APs with known
Yes, from users
Yes; needs occasional fix from
known locations (e.g., GPS lock)
Seamless User
Acquiring Location Fix
Table 4: Performance of state-of-the-art site survey free research w.r.t. calibration-free performance comparison properties.
SDM [76]
Walkie-Markie [82]
TIX [75]
Hossain et al. / Computer Communications 00 (2015) 1–18
Hossain et al. / Computer Communications 00 (2015) 1–18
[1] J. Hightower, G. Borriello, Location systems for ubiquitous
computing, Computer 34 (8) (2001) 57–66.
[2] K. Pahlavan, X. Li, J. Maleka, Indoor geolocation science and
technology, IEEE Communications Mag. 40 (2) (2002) 112–
[3] K. Muthukrishnan, M. Lijding, P. Havinga, Towards smart surroundings: Enabling techniques and technologies for localization, in: T. Strang, C. Linnhoff-Popien (Eds.), Location and
Context-Awareness, Vol. 3479 of Lecture Notes in Computer
Science, Springer Berlin Heidelberg, 2005, pp. 350–362.
[4] R. Want, A. Hopper, V. Falcão, J. Gibbons, The active badge
location system, ACM Trans. on Information Systems 10 (1)
(1992) 91–102.
[5] A. Ward, A. Jones, A. Hopper, A new location technique for
the active office, IEEE Personal Communications 4 (5) (1997)
[6] N. Priyantha, A. Chakraborty, H. Balakrishnan, The cricket
location-support system, in: Proc. ACM MobiCom’00, Boston,
MA, 2000, pp. 32–43.
[7] L. Ni, Y. Liu, Y. C. Lau, A. Patil, LANDMARC: indoor location
sensing using active RFID, in: Proc. of IEEE PerCom, 2003, pp.
[8] P. Bahl, V. N. Padmanabhan, RADAR: An in-building RF-based
user location and tracking system, in: Proc. IEEE INFOCOM,
Tel Aviv, Israel, 2000, pp. 775–784.
[9] P. Castro, P. Chiu, T. Kremenek, R. R. Muntz, A probabilistic
room location service for wireless networked environments, in:
Proc. UbiComp’01, Atlanta, GA, 2001, pp. 18–34.
[10] M. Youssef, A. Agrawala, The horus wlan location determination system, in: Proc. of ACM MobiSys, 2005, pp. 205–218.
[11] T. Roos, P. Myllymki, H. Tirri, P. Misikangas, J. Sievnen, A
probabilistic approach to wlan user location estimation, International Journal of Wireless Information Networks 9 (2002) 155–
[12] Y. Gwon, R. Jain, T. Kawahara, Robust indoor location estimation of stationary and mobile users, in: Proc. IEEE INFOCOM’04, 2004, pp. 1032–1043.
[13] T. King, S. Kopf, T. Haenselmann, C. Lubberger, W. Effelsberg,
COMPASS: a probabilistic indoor positioning system based on
802.11 and digital compasses, in: In Proc. of the 1st International Workshop on Wireless Network Testbeds, Experimental
Evaluation & Characterization, 2006, pp. 34–40.
[14] N. Bulusu, J. Heidemann, D. Estrin, GPS-less low-cost outdoor
localization for very small devices, IEEE Personal Communications Magazine 7 (5) (2000) 28–34.
[15] T. He, C. Huang, B. M. Blum, J. A. Stankovic, T. Abdelzaher,
Range-free localization schemes in large scale sensor networks,
in: Proc. ACM/IEEE Mobicom’03, 2003, pp. 81–95.
[16] A. Savvides, C.-C. Han, M. B. Strivastava, Dynamic finegrained localization in ad-hoc networks of sensors, in: In Proc.
of MobiCom, New York, NY, USA, 2001, pp. 166–179.
[17] H. Wu, C. Wang, N.-F. Tzeng, Novel self-configurable positioning technique for multi-hop wireless networks, IEEE/ACM
Transactions on Networking 13 (3) (2005) 609–621.
[18] N. B. Priyantha, H. Balakrishnan, E. Demaine, S. Teller,
Anchor-free distributed localization in sensor networks, in:
Proc. of SenSys, 2003, pp. 340–341.
[19] K. Yedavalli, B. Krishnamachari, S. Ravula, B. Srinivasan,
Ecolocation: a sequence based technique for rf localization in
wireless sensor networks, in: Proc. ISPN’05, 2005, pp. 285–
[20] Cybernet Interactive, Firefly Motion Capture System.
Northern Digital Inc., OPTOTRAK.
Sonitor, Sonitor Technologies.
UbiSense, Ubisense Real-Time Location Intelligence –
Tadlys Last Yard Solutions., Topaz.
iBeacon, ios: Understanding ibeacon.
MotionStar, Ascension Technology Coporation.
B. Brumitt, B. Meyers, J. Krumm, A. Kern, S. A. Shafer, Easyliving: Technologies for intelligent environments, in: Proceedings of the 2nd International Symposium on Handheld and
Ubiquitous Computing, 2000, pp. 12–29.
Ekahau, Real-Time Location System.
GiPStech, Indoor Localization.
R. Battiti, M. Brunato, A. Villani, Statistical learning theory for
location fingerprinting in wireless LANs, Tech. Rep. DIT-020086, Universita di Trento, Dipartimento di Informatica e Telecomunicazioni (Oct. 2002).
K. Kaemarungsi, P. Krishnamurthy, Modeling of indoor positioning systems based on location fingerprinting, in: Proc. of
IEEE INFOCOM’04, 2004, pp. 1012–1022.
A. Haeberlen, E. Flannery, A. M. Ladd, A. Rudys, D. S. Wallach, L. E. Kavraki, Practical robust localization over largescale 802.11 wireless networks, in: Proc. ACM MobiCom’04,
Philadelphia, PA, 2004, pp. 70–84.
C. tseng Huang, C. hsuan Wu, Y.-N. Yao-Nan Lee, J.-T. Chen, A
novel indoor rss-based position location algorithm using factor
graphs, IEEE Transactions on Wireless Communications 8 (6)
(2009) 3050–3058.
J. geun Park, D. Curtis, S. Teller, J. Ledlie, Implications of device diversity for organic localization, in: In Proc. of IEEE INFOCOM, 2011, pp. 3182–3190.
A. Kushki, K. Plataniotis, A. Venetsanopoulos, Kernel-based
positioning in wireless local area networks, IEEE Transactions
on Mobile Computing 6 (6) (2007) 689–705.
R. Battiti, T. L. Nhat, A. Villani, Location-aware computing:
a neural network model for determining location in wireless
LANs, Tech. Rep. DIT-02-0083, Universita di Trento, Dipartimento di Informatica e Telecomunicazioni (2002).
S. Saha, K. Chaudhuri, D. Sanghi, P. Bhagwat, Location determination of a mobile device using IEEE 802.11b access point
signals, in: In Proc. of IEEE WCNC, Vol. 3, 2003, pp. 1987–
Z.-L. Wu, C.-H. Li, J.-Y. Ng, K. Leung, Location estimation via
support vector regression, IEEE Transactions on Mobile Computing 6 (3) (2007) 311–321.
H. Zou, H. Wang, L. Xie, Q.-S. Jia, An rfid indoor positioning
system by using weighted path loss and extreme learning machine, in: 1st IEEE International Conference on Cyber-Physical
Systems, Networks, and Applications (CPSNA), 2013, pp. 66–
K. Kaemarungsi, P. Krishnamurthy, Properties of indoor received signal strength for WLAN location fingerprinting, in:
Proc. MobiQuitous’04, San Diego, CA, 2004, pp. 14–23.
Hossain et al. / Computer Communications 00 (2015) 1–18
[61] A. K. M. Mahtab Hossain, W. S. Soh, A comprehensive study
of bluetooth signal parameters for localization, in: Proc. IEEE
PIMRC, Athens, Greece, 2007, pp. 1–5.
[62] A. Franssens, Impact of multiple inquirers on the bluetooth discovery process – and its application to localization, Master’s thesis, University of Twente, Netherlands (2010).
[63] C. Smowton, J. R. Lorch, D. Molnar, S. Saroiu, A. Wolman,
Seamless customer identification, Tech. Rep. MSR-TR-201331, Microsoft Research (2013).
[64] U. Ahmad, A. Gavrilov, U. Nasir, M. Iqbal, S. J. Cho, S. Lee, Inbuilding localization using neural networks, in: IEEE International Conference on Engineering of Intelligent Systems, 2006,
pp. 1–6.
[65] C. Laoudias, D. Eliades, P. Kemppi, C. Panayiotou, M. Polycarpou, Indoor localization using neural networks with location fingerprints, in: C. Alippi, M. Polycarpou, C. Panayiotou, G. Ellinas (Eds.), Artificial Neural Networks ICANN 2009, Vol. 5769
of Lecture Notes in Computer Science, Springer Berlin Heidelberg, 2009, pp. 954–963.
[66] H. Wang, S. Sen, A. Elgohary, M. Farid, M. Youssef, R. R.
Choudhury, No need to war-drive: Unsupervised indoor localization, in: Proc. of ACM MobiSys’12, 2012, pp. 197–210.
[67] C. Wu, Z. Yang, Y. Liu, W. Xi, WILL: Wireless indoor localization without site survey, in: Proc. of IEEE INFOCOM, 2012,
pp. 64–72.
[68] E. Bhasker, S. Brown, W. G. Griswold, Employing user feedback for fast, accurate, low-maintenance geolocationing, in: In
Proc. of IEEE PerCom, 2004, pp. 111–120.
[69] A. K. M. Mahtab Hossain, H. Nguyen Van, W.-S. Soh, Utilization of user feedback in indoor positioning system, Pervasive
and Mobile Computing (PMC) 6 (4) (2010) 467–481.
[70] T. Gallagher, B. Li, A. Dempster, C. Rizos, Database updating
through user feedback in fingerprint-based wi-fi location systems, in: Ubiquitous Positioning Indoor Navigation and Location Based Service (UPINLBS), 2010, pp. 1–8.
[71] Y. Luo, O. Hoeber, Y. Chen, Enhancing wi-fi fingerprinting
for indoor positioning using human-centric collaborative feedback, Human-centric Computing and Information Sciences 3 (1)
(2013) 1–23.
[72] P. Krishnan, A. S. Krishnakumar, W. Ju, C. Mallows, and S.
Ganu, A system for lease: location estimation assisted by stationary emitters for indoor rf wireless networks, in: IEEE INFOCOM 2004, Hong Kong, 2004, pp. 1001–1011.
[73] A. Rai, K. K. Chintalapudi, V. N. Padmanabhan, R. Sen, Zee:
Zero-effort crowdsourcing for indoor localization, in: Proc. of
ACM MobiCom’12, 2012, pp. 293–304.
[74] Z. Yang, C. Wu, Y. Liu, Locating in fingerprint space: Wireless
indoor localization with little human intervention, in: Proc. of
ACM MobiCom’12, 2012, pp. 269–280.
[75] Y. Gwon, R. Jain, Error characteristics and calibration-free techniques for wireless lan-based location estimation, in: Proc. of
the Second ACM MobiWac, New York, NY, USA, 2004, pp.
[76] H. Lim, L.-C. Kung, J. C. Hou, H. Luo, Zero-configuration indoor localization over ieee 802.11 wireless infrastructure, Wireless Network 16 (2) (2010) 405–420.
[77] J.-g. Park, B. Charrow, D. Curtis, J. Battat, E. Minkov, J. Hicks,
S. Teller, J. Ledlie, Growing an organic indoor location system,
in: Proc. of ACM MobiSys, MobiSys ’10, 2010, pp. 271–284.
[78] A. Goswami, L. E. Ortiz, S. R. Das, WiGEM: A Learning-based
Approach for Indoor Localization, in: Proc. of ACM CoNEXT,
2011, pp. 1–12.
[79] C. Wu, Z. Yang, Y. Liu, W. Xi, Site-survey-free wireless localization using mobile phones, Tech. rep., Hong Kong University
of Science and Technology (2011).
[41] A. Ladd, K. Bekris, G. Marceau, A. Rudys, L. Kavraki, D. Wallach, Robotics-based location sensing using wireless ethernet,
Tech. Rep. TR02-393, Department of Computer Science, Rice
University (2002).
[42] A. K. M. Mahtab Hossain, Y. Jin, W.-S. Soh, H. N. Van, SSD:
A robust rf location fingerprint addressing mobile devices’ heterogeneity, IEEE TMC 12 (1) (2013) 65–77.
[43] M. B. Kjærgaard, Indoor location ffingerprinting with heterogeneous clients, Pervasive and Mobile Computing 7 (1) (2011)
[44] F. Dong, Y. Chen, J. Liu, Q. Ning, S. Piao, A calibrationfree localization solution for handling signal strength variance,
in: Mobile Entity Localization and Tracking in GPS-less Environnments, Vol. 5801 of Lecture Notes in Computer Science,
Springer Berlin Heidelberg, 2009, pp. 79–90.
[45] P. Tao, A. Rudys, A. M. Ladd, D. S. Wallach, Wireless
LAN location-sensing for security applications, in: Proc. ACM
WiSe’03, San Diego, CA, 2003, pp. 11–20.
[46] H. Liu, H. Darabi, P. Banerjee, J. Liu, Survey of wireless indoor
positioning techniques and systems, IEEE Transactions on Systems, Man, and Cybernetics, Part C: Applications and Reviews
37 (6) (2007) 1067–1080.
[47] Y. Gu, A. Lo, I. Niemegeers, A survey of indoor positioning
systems for wireless personal networks, IEEE Communications
Surveys Tutorials 11 (1) (2009) 13–32.
[48] E. C. L. Chan, G. Baciu, Introduction to Wireless Localization:
With iPhone SDK Examples, Wiley-IEEE Press, 2012.
[49] K. Chintalapudi, A. Padmanabha Iyer, V. N. Padmanabhan, Indoor localization without the pain, in: Proc. of ACM MobiCom’10, 2010, pp. 173–184.
[50] M. Brunato, C. K. Kall, Transparent location fingerprinting for
wireless services, in: Proc. Med-Hoc-Net, 2002.
[51] A. Beresford, F. Stajano, Location privacy in pervasive computing, IEEE Pervasive Computing 2 (1) (2003) 46–55.
[52] J. Hallberg, M. Nilsson, K. Synnes, Positioning with Bluetooth,
in: Proc. ICT, 2003, pp. 954–958.
[53] Zebra Location Solutions, Dart Ultra-Wide Band (UWB) Real
Time Location System RTLS- RFID tracking.
[54] C. V. Lopes, A. Haghighat, A. Mandal, T. Givargis, P. Baldi,
Localization of Off-the-shelf Mobile Devices Using Audible
Sound: Architectures, Protocols and Performance Assessment,
SIGMOBILE Mobile Computer Communication Review 10 (2)
(2006) 38–50.
[55] Radianse, Radianse — RTLS Solutions for Healthcare.
[56] C. Brignone, T. Connors, G. Lyon, S. Pradhan, An autonomous,
self-assembling sensor network for indoor asset and systems
management, Tech. Rep. HPL-2003-41, HP Labs (2003).
[57] A. LaMarca, Y. Chawathe, S. Consolvo, J. Hightower, I. Smith,
J. Scott, T. Sohn, J. Howard, J. Hughes, F. Potter, J. Tabert,
P. Powledge, G. Borriello, B. Schilit, Place lab: Device positioning using radio beacons in the wild, in: Proc. of Pervasive
Computing, 2005, pp. 116–133.
[58] Skyhook, Skyhook Wireless.
[59] M. Azizyan, I. Constandache, R. Roy Choudhury, Surroundsense: Mobile phone localization via ambience fingerprinting,
in: Proc. of ACM MobiCom’09, MobiCom ’09, 2009, pp. 261–
[60] M. Papandrea, S. Giordano, Location prediction and mobility
modelling for enhanced localization solution, Journal of Ambient Intelligence and Humanized Computing 5 (3) (2014) 279–
Hossain et al. / Computer Communications 00 (2015) 1–18
[80] I. Constandache, R. Choudhury, I. Rhee, Towards mobile phone
localization without war-driving, in: Proc. of IEEE INFOCOM,
2010, pp. 1–9.
[81] T. H. Cormen, C. Stein, R. L. Rivest, C. E. Leiserson, Introduction to Algorithms, 2nd Edition, McGraw-Hill Higher Education, 2001.
[82] G. Shen, Z. Chen, P. Zhang, T. Moscibroda, Y. Zhang, Walkiemarkie: Indoor pathway mapping made easy, in: Proceedings
of the 10th USENIX Conference on Networked Systems Design
and Implementation, Berkeley, CA, USA, 2013, pp. 85–98.
[83] H. Shin, Y. Chon, H. Cha, Unsupervised construction of an indoor floor plan using a smartphone, IEEE Transactions on Systems, Man, and Cybernetics, Part C: Applications and Reviews
42 (6) (2012) 889–898.
[84] H. Durrant-Whyte, T. Bailey, Simultaneous localization and
mapping: part I, IEEE Robotics Automation Magazine 13 (2)
(2006) 99–110.
[85] G. Welch, G. Bishop, An introduction to the kalman filter, Tech.
rep., Chapel Hill, NC, USA (1995).
[86] M. Montemerlo, S. Thrun, D. Koller, B. Wegbreit, Fastslam: A
factored solution to the simultaneous localization and mapping
problem, in: In Proceedings of the AAAI National Conference
on Artificial Intelligence, 2002, pp. 593–598.
[87] P. Robertson, M. Angermann, B. Krach, Simultaneous localization and mapping for pedestrians using only foot-mounted inertial sensors, in: In Proc. of ACM Ubicomp, 2009, pp. 93–96.
[88] M. Hardegger, D. Roggen, S. Mazilu, G. Troster, Actionslam:
Using location-related actions as landmarks in pedestrian slam,
in: In Proc. of Indoor Positioning and Indoor Navigation (IPIN),
2012, pp. 1–10.
[89] P. Robertson, M. Angermann, M. Khider, Improving simultaneous localization and mapping for pedestrian navigation and
automatic mapping of buildings by using online human-based
feature labeling, in: IEEE/ION Position Location and Navigation Symposium (PLANS), 2010, pp. 365–374.
[90] B. Ferris, D. Fox, N. Lawrence, WiFi-SLAM Using Gaussian
Process Latent Variable Models, in: In Proc. of the 20th International Joint Conference on Artifical Intelligence, 2007, pp.
[91] J. Huang, D. Millman, M. Quigley, D. Stavens, S. Thrun, A. Aggarwal, Efficient, generalized indoor wifi graphslam, in: IEEE
International Conference on Robotics and Automation (ICRA),
2011, pp. 1038–1043.
[92] Navigine, Navigine | Universal Platform for Navigation.
[93] Navisens, Navisens | Location Aware.
[94], | Accurate real-time Indoor Localization .
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF