empirical)performance)of)) self0calibrang) wifi)locaon...
TRANSCRIPT
Empirical Performance of Self-‐calibra2ng WiFi Loca2on Systems
Daniel Turner Stefan Savage Alex C. Snoeren
UCSD 10/05/2011
Localiza2on Outdoors
• GPS tells us where we are outdoors • GPS + smart phone
Localiza2on Outdoors
• GPS tells us where we are outdoors • GPS + smart phone
– Turn-‐by-‐turn driving direc2ons
Localiza2on Outdoors
• GPS tells us where we are outdoors • GPS + smart phone
– Turn-‐by-‐turn driving direc2ons – Yelp
Localiza2on Indoors
• Sadly GPS works poorly indoors • Indoor localiza2on services
– Tour guides in museums – Adver2sing in Malls – The next BIG thing
• So how do we get indoor ‘GPS’ – For next genera2on of applica2ons
• How much accuracy do we need? – 3-‐5 meters depending on applica2on
Custom Hardware
• Deploy new localiza2on infrastructure – Bluetooth, RFID, Sonar, ..
• AoA, TDoA, and Signal Decay – Requires special purpose infrastructure
Leveraging WiFi Infrastructure
• 802.11 (WiFi) loca2on systems – Median accuracy around 3 Meters – Requires significant manual calibra2on
RSSI
• WiFi exposes Received Signal Strength Indica2on (RSSI) – measurement of the power present in a received packet
– Higher value indicated stronger signal – RSSI is a unit-‐less quan2ty
Fingerprint Based Approaches
AP 3 AP 2
AP 1
AP 4
Fingerprint Based Approaches
AP 3 AP 2
AP 1
AP 4
Fingerprint Based Approaches
AP 3
AP 1: 44.1 RSSI AP 2: 22.7 AP 3: 19.8 AP 4: 35
AP 2
AP 1
AP 4
Fingerprint Based Approaches
AP 3
AP 1: 44.1 RSSI AP 2: 22.7 AP 3: 19.8 AP 4: 35
AP 1: 42.1 RSSI AP 2: 27.7 AP 3: 24.8 AP 4: 31
AP 2
AP 1
AP 4
Fingerprint Based Approaches
AP 3 AP 2
AP 1
AP 4
Fingerprint Based Approaches
AP 3 AP 2
AP 1
AP 4
AP 1: 42.1 RSSI AP 2: 23.4 AP 3: 22.6 AP 4: 39
The Next Challenge
• Indoor localiza2on is possible • Can be done using only WiFi
• Can it be done w/o human calibra2on? – Only Calibra2on data from AP to AP transmissions
Self-‐Calibra2on
• We can leverage intra-‐AP transmissions to build models
AP 3
AP 1
AP 2
AP 4
Self-‐Calibra2on
• We can leverage intra-‐AP transmissions to build models
• Signal-‐Distance pairs
AP 3
AP 1
AP 2
AP 4
Sender Distance (M)
Mean RSSI
Varia2on
AP 2 4.6 48.9 1.2
AP 1 10.0 42.3 0.8
AP 4 15.4 12.1 0.7
Plug-‐n-‐Play for Indoor Localiza2on
• Self-‐Calibra2ng WiFi loca2on systems – Recognized that WiFi decays exponen2ally – Use models to map RSSI to distance
– Leverage WiFi infrastructure to calibrate models – Published 2-‐5 Meters of accuracy
Published Self-‐Calibra2ng Systems
• Algorithms uses varying amounts of data to train models – In theory more training data should lead to beger accuracy
• Triangular Interpola5on and eXtrapola5on (TiX) Gwon and Jain 2004
• Lease Krishnakumar et al 2004 • Bayesian indoor posi2oning system D. Madigan et al 2005
• Ariadne y. Ji et al 2006
• Probabilis5c Method Moares and Nunes 2006
• Zero-‐configura5on and robust indoor localiza5on (ZCFG) Lim et al. 2006
• EZ Localiza2on Chintalapudi et atl. 2010
Picking a System to Deploy
• There are mul2ple Self-‐Calibra2ng systems
• Published performance all under 5 Meters
• Not clear which we should(n’t) use
Study # of APs Published Accuracy
Probabilis2c (de Moraes & Nunes) 4 1.84
TiX (Gwon & Jain) 4 5
ZCFG (Lim et al.) 6 2.5
Our Environment
• Test Environment – 30k Sq. Ft. per floor – 10 APs per Floor
• Channels 1,6, and 11
Our Environment
• Test Environment – 30k Sq. Ft. per floor – 10 APs per Floor
• Channels 1,6, and 11
Study Square Feet # of Aps Aps/Sq. Ft.
Probabilis2c (de Moraes & Nunes) 1,722 4 0.0023
TiX (Gwon & Jain) 6,717 4 0.0006
ZCFG (Lim et al.) 598 6 0.01
Our Environment
• Test Environment – 30k Sq. Ft. per floor – 10 APs per Floor
• Channels 1,6, and 11
Study Square Feet # of Aps Aps/Sq. Ft.
Probabilis2c (de Moraes & Nunes) 1,722 4 0.0023
TiX (Gwon & Jain) 6,717 4 0.0006
ZCFG (Lim et al.) 598 6 0.01
Cisco Recommends 3,000-‐5,000 1 0.0002-‐0.0003
Study Square Feet # of Aps Aps/Sq. Ft.
Probabilis2c (de Moraes & Nunes) 1,722 4 0.0023
TiX (Gwon & Jain) 6,717 4 0.0006
ZCFG (Lim et al.) 598 6 0.01
Our Environment
• Test Environment – 30k Sq. Ft. per floor – 10 APs per Floor
• Channels 1,6, and 11
Study Square Feet # of Aps Aps/Sq. Ft.
Probabilis2c (de Moraes & Nunes) 1,722 4 0.0023
TiX (Gwon & Jain) 6,717 4 0.0006
ZCFG (Lim et al.) 598 6 0.01
Cisco Recommends 3,000-‐5,000 1 0.0002-‐0.0003
UCSD CSE Building 30,000 10 0.0003
Study Square Feet # of Aps Aps/Sq. Ft.
Probabilis2c (de Moraes & Nunes) 1,722 4 0.0023
TiX (Gwon & Jain) 6,717 4 0.0006
ZCFG (Lim et al.) 598 6 0.01
Cisco Recommends 3,000-‐5,000 1 0.0002-‐0.0003
Study Square Feet # of Aps Aps/Sq. Ft.
Probabilis2c (de Moraes & Nunes) 1,722 4 0.0023
TiX (Gwon & Jain) 6,717 4 0.0006
ZCFG (Lim et al.) 598 6 0.01
Contribu2ons
• Perform head-‐to-‐head comparisons – 4 different algorithms
• Algorithms don’t work as well as published – As much as 4x worse than previously published
• We provide insight into why not – Performance loss due to environmental issues – Performance loss due to systemic issues
Outline
• Mo2va2on • Algorithms
• Experimental set-‐up / results
• Causes of Error • Conclusion
Nearest AP
AP 3
AP 1
AP 2
AP 4
Nearest AP technologies including Bluetooth signals [12], RFID proximity[19], and ultrasonic acoustic waves [3], among others.
Systems that use signal strength measurements, or, morespecifically in the case of WiFi, the received signal strengthindication (RSSI)—a unit-less measure of signal power asso-ciated with a packet as measured by a receiving device—canbe subdivided into two categories: those that require manualcalibration and those that do not. While the goal of this paperis to evaluate self-calibrating 802.11-based location systems,we briefly summarize the approaches based upon manualcalibration for context.
Many of the early (and still most accurate) 802.11-basedlocation systems require manual calibration. During this initialsite survey an installer collects client fingerprints (RSSI valuesfor packets broadcast by APs) at a large number of knownlocations to build up a fine-grained fingerprint database thatcan be used for future client localization. In general thesesystems have three major manifestations: point matching,probabilistic matching, and Bayesian networks.
Point matching, also known as nearest neighbor in signal
strength [2], starts by capturing a client fingerprint and com-paring it against the initial fingerprint database scanning for thefingerprint that minimizes a similarity metric, often Euclideandistance. The client’s location is determined to be the locationof the offline fingerprint that minimizes the metric.
The second set of methods once again starts with the usertaking a client fingerprint. Then the user computes the proba-bility of receiving the fingerprint at each location, l = (x, y),in the building. Assuming that the distribution of RSSIs in thefingerprint is Gaussian, it is straightforward to determine an lsuch that P (s|l) is maximized [1], [25].
Algorithms based on Bayesian networks are the most com-plicated. They depend on building a Hidden Markov Model(HMM) in which the RSSIs represent the observed state, andthe distance from the APs (and hence the x, y coordinatesof the client) are the hidden state. The relationship betweenvisible and hidden states is encoded via a signal decay modeland then the fingerprint database is used to train the parametersof the model [16], [21].
Chintalapudi et al. recently proposed EZ Localization as amethodology for calibration-free wireless localization in anindoor environment [6]. EZ uses WiFi-enabled smart phonesto take fingerprints at unknown locations in a building. Ratherthan using the location of APs to ground their model, however,they leverage the GPS units available on the client devices.When a phone is able to get a GPS lock they opportunisticallyuse the measurement as an anchor point. They then treat thefingerprints as constraints to be solved by a genetic algorithm.The authors report that EZ has low error in their environment,but remains unable to surpass the accuracy of systems thatrequire manual calibration.
III. SELF-CALIBRATING ALGORITHMS
Because manually calibrated systems may require takinghundreds of fingerprints per floor, a second class of systemsthat requires no manual calibration has been proposed. Instead,
1
3 4
2
(a) Simplified network diagram
Device Mean RSSIAP1 24.4AP2 47.8AP3 45.5AP4 40.6
(b) Client fingerprint
Fig. 1. An example network and fingerprint corresponding to signal A
these self-calibrating systems leverage information regardingthe location of WiFi access points to conduct a trainingphase in which they collect system fingerprints, or sets ofRSSI measurements taken of (and by) the stationary APs inthe network. After the self-calibration stage, the fingerprintdatabase is combined with pre-configured information aboutthe physical location of the APs to compute a model of therelationship between RSSI and physical location. This modelis then used to answer on-line queries about client location.
A. Overview
It is convenient to think of self-calibrating algorithms asfitting into three categories based on the amount of data usedto perform calibration. To assist the reader in understandingeach algorithm we give examples of how they work in a simplewireless network. Our explanatory network, diagrammed inFigure 1(a), has four APs and one client. Each AP knowsthe distance between it and every other AP, for example thedistance between AP3 and AP4 is ten meters and labeled D.The 802.11 signal emitted by the client when sending a packetis labeled A. The RSSI values for one transmission as seen bythe four APs, i.e. a client fingerprint, are documented in thetable in Figure 1(b). Conversely, the 802.11 signal generatedwhen an AP sends a packet into the network is labeled C.When an AP sends a packet it is heard by everyone withinrange, though all but the intended destination will usuallyignore it when the network is operational. In the systemtraining phase, however, APs do not ignore packets sent fromother APs, labeled B, but instead record their RSSI values.Table I shows an example of training data that might observedby AP4 in this network.
Baseline: In order to provide some context in which tointerpret the performance of the localization system classes, weconsider a simple, calibration-free baseline that leverages onlya single RSSI from a client transmission, labeled A in Figure1(a). This nearest-neighbor approach requires no calibration:to locate a client it simply determines which AP heard theclient’s transmission with the strongest RSSI and declares theclient to be co-located with that AP. For example if the systemobtained the fingerprint in Figure 1(b) it would estimate theclient to be at the same location as AP2.
Class I: Our baseline algorithm, while simple, has twomajor limitations. First, it can choose only locations occupiedby an AP. Second, it ignores all information from the (n− 1)
AP 3
AP 1
AP 2
AP 4
Nearest AP technologies including Bluetooth signals [12], RFID proximity[19], and ultrasonic acoustic waves [3], among others.
Systems that use signal strength measurements, or, morespecifically in the case of WiFi, the received signal strengthindication (RSSI)—a unit-less measure of signal power asso-ciated with a packet as measured by a receiving device—canbe subdivided into two categories: those that require manualcalibration and those that do not. While the goal of this paperis to evaluate self-calibrating 802.11-based location systems,we briefly summarize the approaches based upon manualcalibration for context.
Many of the early (and still most accurate) 802.11-basedlocation systems require manual calibration. During this initialsite survey an installer collects client fingerprints (RSSI valuesfor packets broadcast by APs) at a large number of knownlocations to build up a fine-grained fingerprint database thatcan be used for future client localization. In general thesesystems have three major manifestations: point matching,probabilistic matching, and Bayesian networks.
Point matching, also known as nearest neighbor in signal
strength [2], starts by capturing a client fingerprint and com-paring it against the initial fingerprint database scanning for thefingerprint that minimizes a similarity metric, often Euclideandistance. The client’s location is determined to be the locationof the offline fingerprint that minimizes the metric.
The second set of methods once again starts with the usertaking a client fingerprint. Then the user computes the proba-bility of receiving the fingerprint at each location, l = (x, y),in the building. Assuming that the distribution of RSSIs in thefingerprint is Gaussian, it is straightforward to determine an lsuch that P (s|l) is maximized [1], [25].
Algorithms based on Bayesian networks are the most com-plicated. They depend on building a Hidden Markov Model(HMM) in which the RSSIs represent the observed state, andthe distance from the APs (and hence the x, y coordinatesof the client) are the hidden state. The relationship betweenvisible and hidden states is encoded via a signal decay modeland then the fingerprint database is used to train the parametersof the model [16], [21].
Chintalapudi et al. recently proposed EZ Localization as amethodology for calibration-free wireless localization in anindoor environment [6]. EZ uses WiFi-enabled smart phonesto take fingerprints at unknown locations in a building. Ratherthan using the location of APs to ground their model, however,they leverage the GPS units available on the client devices.When a phone is able to get a GPS lock they opportunisticallyuse the measurement as an anchor point. They then treat thefingerprints as constraints to be solved by a genetic algorithm.The authors report that EZ has low error in their environment,but remains unable to surpass the accuracy of systems thatrequire manual calibration.
III. SELF-CALIBRATING ALGORITHMS
Because manually calibrated systems may require takinghundreds of fingerprints per floor, a second class of systemsthat requires no manual calibration has been proposed. Instead,
1
3 4
2
(a) Simplified network diagram
Device Mean RSSIAP1 24.4AP2 47.8AP3 45.5AP4 40.6
(b) Client fingerprint
Fig. 1. An example network and fingerprint corresponding to signal A
these self-calibrating systems leverage information regardingthe location of WiFi access points to conduct a trainingphase in which they collect system fingerprints, or sets ofRSSI measurements taken of (and by) the stationary APs inthe network. After the self-calibration stage, the fingerprintdatabase is combined with pre-configured information aboutthe physical location of the APs to compute a model of therelationship between RSSI and physical location. This modelis then used to answer on-line queries about client location.
A. Overview
It is convenient to think of self-calibrating algorithms asfitting into three categories based on the amount of data usedto perform calibration. To assist the reader in understandingeach algorithm we give examples of how they work in a simplewireless network. Our explanatory network, diagrammed inFigure 1(a), has four APs and one client. Each AP knowsthe distance between it and every other AP, for example thedistance between AP3 and AP4 is ten meters and labeled D.The 802.11 signal emitted by the client when sending a packetis labeled A. The RSSI values for one transmission as seen bythe four APs, i.e. a client fingerprint, are documented in thetable in Figure 1(b). Conversely, the 802.11 signal generatedwhen an AP sends a packet into the network is labeled C.When an AP sends a packet it is heard by everyone withinrange, though all but the intended destination will usuallyignore it when the network is operational. In the systemtraining phase, however, APs do not ignore packets sent fromother APs, labeled B, but instead record their RSSI values.Table I shows an example of training data that might observedby AP4 in this network.
Baseline: In order to provide some context in which tointerpret the performance of the localization system classes, weconsider a simple, calibration-free baseline that leverages onlya single RSSI from a client transmission, labeled A in Figure1(a). This nearest-neighbor approach requires no calibration:to locate a client it simply determines which AP heard theclient’s transmission with the strongest RSSI and declares theclient to be co-located with that AP. For example if the systemobtained the fingerprint in Figure 1(b) it would estimate theclient to be at the same location as AP2.
Class I: Our baseline algorithm, while simple, has twomajor limitations. First, it can choose only locations occupiedby an AP. Second, it ignores all information from the (n− 1)
• Device is located at the AP that with the greatest RSSI
AP 3
AP 1
AP 2
AP 4
Nearest AP technologies including Bluetooth signals [12], RFID proximity[19], and ultrasonic acoustic waves [3], among others.
Systems that use signal strength measurements, or, morespecifically in the case of WiFi, the received signal strengthindication (RSSI)—a unit-less measure of signal power asso-ciated with a packet as measured by a receiving device—canbe subdivided into two categories: those that require manualcalibration and those that do not. While the goal of this paperis to evaluate self-calibrating 802.11-based location systems,we briefly summarize the approaches based upon manualcalibration for context.
Many of the early (and still most accurate) 802.11-basedlocation systems require manual calibration. During this initialsite survey an installer collects client fingerprints (RSSI valuesfor packets broadcast by APs) at a large number of knownlocations to build up a fine-grained fingerprint database thatcan be used for future client localization. In general thesesystems have three major manifestations: point matching,probabilistic matching, and Bayesian networks.
Point matching, also known as nearest neighbor in signal
strength [2], starts by capturing a client fingerprint and com-paring it against the initial fingerprint database scanning for thefingerprint that minimizes a similarity metric, often Euclideandistance. The client’s location is determined to be the locationof the offline fingerprint that minimizes the metric.
The second set of methods once again starts with the usertaking a client fingerprint. Then the user computes the proba-bility of receiving the fingerprint at each location, l = (x, y),in the building. Assuming that the distribution of RSSIs in thefingerprint is Gaussian, it is straightforward to determine an lsuch that P (s|l) is maximized [1], [25].
Algorithms based on Bayesian networks are the most com-plicated. They depend on building a Hidden Markov Model(HMM) in which the RSSIs represent the observed state, andthe distance from the APs (and hence the x, y coordinatesof the client) are the hidden state. The relationship betweenvisible and hidden states is encoded via a signal decay modeland then the fingerprint database is used to train the parametersof the model [16], [21].
Chintalapudi et al. recently proposed EZ Localization as amethodology for calibration-free wireless localization in anindoor environment [6]. EZ uses WiFi-enabled smart phonesto take fingerprints at unknown locations in a building. Ratherthan using the location of APs to ground their model, however,they leverage the GPS units available on the client devices.When a phone is able to get a GPS lock they opportunisticallyuse the measurement as an anchor point. They then treat thefingerprints as constraints to be solved by a genetic algorithm.The authors report that EZ has low error in their environment,but remains unable to surpass the accuracy of systems thatrequire manual calibration.
III. SELF-CALIBRATING ALGORITHMS
Because manually calibrated systems may require takinghundreds of fingerprints per floor, a second class of systemsthat requires no manual calibration has been proposed. Instead,
1
3 4
2
(a) Simplified network diagram
Device Mean RSSIAP1 24.4AP2 47.8AP3 45.5AP4 40.6
(b) Client fingerprint
Fig. 1. An example network and fingerprint corresponding to signal A
these self-calibrating systems leverage information regardingthe location of WiFi access points to conduct a trainingphase in which they collect system fingerprints, or sets ofRSSI measurements taken of (and by) the stationary APs inthe network. After the self-calibration stage, the fingerprintdatabase is combined with pre-configured information aboutthe physical location of the APs to compute a model of therelationship between RSSI and physical location. This modelis then used to answer on-line queries about client location.
A. Overview
It is convenient to think of self-calibrating algorithms asfitting into three categories based on the amount of data usedto perform calibration. To assist the reader in understandingeach algorithm we give examples of how they work in a simplewireless network. Our explanatory network, diagrammed inFigure 1(a), has four APs and one client. Each AP knowsthe distance between it and every other AP, for example thedistance between AP3 and AP4 is ten meters and labeled D.The 802.11 signal emitted by the client when sending a packetis labeled A. The RSSI values for one transmission as seen bythe four APs, i.e. a client fingerprint, are documented in thetable in Figure 1(b). Conversely, the 802.11 signal generatedwhen an AP sends a packet into the network is labeled C.When an AP sends a packet it is heard by everyone withinrange, though all but the intended destination will usuallyignore it when the network is operational. In the systemtraining phase, however, APs do not ignore packets sent fromother APs, labeled B, but instead record their RSSI values.Table I shows an example of training data that might observedby AP4 in this network.
Baseline: In order to provide some context in which tointerpret the performance of the localization system classes, weconsider a simple, calibration-free baseline that leverages onlya single RSSI from a client transmission, labeled A in Figure1(a). This nearest-neighbor approach requires no calibration:to locate a client it simply determines which AP heard theclient’s transmission with the strongest RSSI and declares theclient to be co-located with that AP. For example if the systemobtained the fingerprint in Figure 1(b) it would estimate theclient to be at the same location as AP2.
Class I: Our baseline algorithm, while simple, has twomajor limitations. First, it can choose only locations occupiedby an AP. Second, it ignores all information from the (n− 1)
• Device is located at the AP that with the greatest RSSI
AP 3
AP 1
AP 2
AP 4
Self-‐Calibra2ng System Phases
• Algorithms have two phases – 1) Determine distance from device and each AP – 2) Combine these distance es2ma2ons
• Phase 1 uses models to map distance between each AP and the device to be located
Class III: The final class of algorithms we study makes onesignificant conceptual addition to the third class of algorithms:It recognizes that each AP is able to map the distance fromitself to the client via RSSI, but, also, that the other (n − 1)APs can map the distance between the client and AP fromtheir own perspective [11], [17]. For example, AP4 can notonly estimate the distance between itself and the client, butalso the distance between the client and the other APs. Thusfinal estimation of the distance between AP4 and the clientis a (linear) combination of each AP’s distance estimationbetween the client and AP4. We chose ZCFG [17] as ourrepresentative algorithm for Class III because of its highlevel of sophistication and built-in resistance to error fromenvironmental noise.
ZCFG, like the previous algorithms, requires a trainingphase during which each AP records the average RSSI forall APs in range. Unlike the TiX algorithm, the calibration ofZCFG is done in real time. ZCFG uses a signal distance map(SDM) to “describe the relation between the RSS(I)s and thegeographic distance” [17]. Every time ZCFG wants to locatea device, it identifies each of the n APs that observed a packetfrom the device. It then builds the n × n dimensional SDMaccording to the following equation, T = log(D)ST (SST )−1
where S,D are n×n dimensional matrices such that Sij , Dij
are the RSSI as observed by AP i from AP j and the physicaldistance between AP i and j respectively. ZCFG then finds thedistance between the device and each of the n APs accordingto the equation: dn = eTsn where sn is the column vectorof the observed RSSIs. Finally, ZCFG uses a multilaterationalgorithm to determine a location estimation.
IV. EXPERIMENTAL EVALUATION
Our goal is to perform a fair and accurate head-to-headcomparison of the three exemplar algorithms presented in theprevious section. We start by quantifying the performance ofthe algorithms in several locations in our office building. Moreimportantly, we attempt to understand limiting factors that willeffect most real-world deployments. We also identify severalpractical issues that were not discussed by the algorithms’authors in their original publications and, where possible, offerguidance in addressing them.
A. Experimental setupOur test building has approximately 145,000 square feet
over four stories and a basement. All experiments were con-ducted on the third floor which has ten access points dividedbetween channels 1, 6, and 11.
The APs in our building are not under our control. Thismeans that we are unable to use them to observe traffic.Instead, we monitor traffic using passive wireless monitors [4].There are 48 pairs of such monitors installed throughout ourbuilding, with approximately 12 pairs (24 distinct monitoringdevices with two radios each) located on each floor. Likethe initial AP placement, the monitor layout was specificallydesigned to provide a maximal coverage of each floor and,like the APs, should neither favor nor disfavor any specific
Fig. 3. Locations on the 3rd floor
location. The locations of the APs and monitors are indicatedin Figure 3.
For training the algorithms, we opportunistically use theexisting AP traffic (i.e., as though each AP was a client) as weknow the precise location of each AP. This is sufficient for con-structing the fingerprint database for the trivial, probabilistic,and TiX algorithms. Since the online training process of ZCFGrequires an architecture in which packets are transmitted aswell as received, we implement this by forcing each monitorto send pings to each of the others. Since there are 11 pairs ofmonitors but only 10 APs on the floor we consider, ZCFG will,in principle, have more training data to work with. This shouldgive ZCFG an advantage over the other three algorithms, butSection IV-D shows that this advantage does not materialize.
To build the fingerprint database we allow each algorithmto monitor the broadcast packets sent from the APs in thebuilding for approximately ten seconds. We find that thereis sufficient traffic in ten seconds to allow the mean RSSIvalue to be minimally affected by abnormally large or smallinstantaneous values. It is possible to locate a client from onlya single packet. However, we averaged two seconds worth ofRSSI values from the client. This time frame is a balancebetween reducing the effect of an outlier in the RSSI valuesand effectively locating a user even if they are in motion.Throughout this paper we report RSSI values instead of dB;RSSI values can be converted to dB by subtracting 95.
Probabilis2c Method (Moares and Nunes)
• Phase 1: Building a Radio Propaga2on Map – Overlay grid w/ probability that device is present
Class III: The final class of algorithms we study makes onesignificant conceptual addition to the third class of algorithms:It recognizes that each AP is able to map the distance fromitself to the client via RSSI, but, also, that the other (n − 1)APs can map the distance between the client and AP fromtheir own perspective [11], [17]. For example, AP4 can notonly estimate the distance between itself and the client, butalso the distance between the client and the other APs. Thusfinal estimation of the distance between AP4 and the clientis a (linear) combination of each AP’s distance estimationbetween the client and AP4. We chose ZCFG [17] as ourrepresentative algorithm for Class III because of its highlevel of sophistication and built-in resistance to error fromenvironmental noise.
ZCFG, like the previous algorithms, requires a trainingphase during which each AP records the average RSSI forall APs in range. Unlike the TiX algorithm, the calibration ofZCFG is done in real time. ZCFG uses a signal distance map(SDM) to “describe the relation between the RSS(I)s and thegeographic distance” [17]. Every time ZCFG wants to locatea device, it identifies each of the n APs that observed a packetfrom the device. It then builds the n × n dimensional SDMaccording to the following equation, T = log(D)ST (SST )−1
where S,D are n×n dimensional matrices such that Sij , Dij
are the RSSI as observed by AP i from AP j and the physicaldistance between AP i and j respectively. ZCFG then finds thedistance between the device and each of the n APs accordingto the equation: dn = eTsn where sn is the column vectorof the observed RSSIs. Finally, ZCFG uses a multilaterationalgorithm to determine a location estimation.
IV. EXPERIMENTAL EVALUATION
Our goal is to perform a fair and accurate head-to-headcomparison of the three exemplar algorithms presented in theprevious section. We start by quantifying the performance ofthe algorithms in several locations in our office building. Moreimportantly, we attempt to understand limiting factors that willeffect most real-world deployments. We also identify severalpractical issues that were not discussed by the algorithms’authors in their original publications and, where possible, offerguidance in addressing them.
A. Experimental setupOur test building has approximately 145,000 square feet
over four stories and a basement. All experiments were con-ducted on the third floor which has ten access points dividedbetween channels 1, 6, and 11.
The APs in our building are not under our control. Thismeans that we are unable to use them to observe traffic.Instead, we monitor traffic using passive wireless monitors [4].There are 48 pairs of such monitors installed throughout ourbuilding, with approximately 12 pairs (24 distinct monitoringdevices with two radios each) located on each floor. Likethe initial AP placement, the monitor layout was specificallydesigned to provide a maximal coverage of each floor and,like the APs, should neither favor nor disfavor any specific
Fig. 3. Locations on the 3rd floor
location. The locations of the APs and monitors are indicatedin Figure 3.
For training the algorithms, we opportunistically use theexisting AP traffic (i.e., as though each AP was a client) as weknow the precise location of each AP. This is sufficient for con-structing the fingerprint database for the trivial, probabilistic,and TiX algorithms. Since the online training process of ZCFGrequires an architecture in which packets are transmitted aswell as received, we implement this by forcing each monitorto send pings to each of the others. Since there are 11 pairs ofmonitors but only 10 APs on the floor we consider, ZCFG will,in principle, have more training data to work with. This shouldgive ZCFG an advantage over the other three algorithms, butSection IV-D shows that this advantage does not materialize.
To build the fingerprint database we allow each algorithmto monitor the broadcast packets sent from the APs in thebuilding for approximately ten seconds. We find that thereis sufficient traffic in ten seconds to allow the mean RSSIvalue to be minimally affected by abnormally large or smallinstantaneous values. It is possible to locate a client from onlya single packet. However, we averaged two seconds worth ofRSSI values from the client. This time frame is a balancebetween reducing the effect of an outlier in the RSSI valuesand effectively locating a user even if they are in motion.Throughout this paper we report RSSI values instead of dB;RSSI values can be converted to dB by subtracting 95.
Probabilis2c Method (Moares and Nunes)
• Phase 1: Building a Radio Propaga2on Map – Overlay grid w/ probability that device is present
Class III: The final class of algorithms we study makes onesignificant conceptual addition to the third class of algorithms:It recognizes that each AP is able to map the distance fromitself to the client via RSSI, but, also, that the other (n − 1)APs can map the distance between the client and AP fromtheir own perspective [11], [17]. For example, AP4 can notonly estimate the distance between itself and the client, butalso the distance between the client and the other APs. Thusfinal estimation of the distance between AP4 and the clientis a (linear) combination of each AP’s distance estimationbetween the client and AP4. We chose ZCFG [17] as ourrepresentative algorithm for Class III because of its highlevel of sophistication and built-in resistance to error fromenvironmental noise.
ZCFG, like the previous algorithms, requires a trainingphase during which each AP records the average RSSI forall APs in range. Unlike the TiX algorithm, the calibration ofZCFG is done in real time. ZCFG uses a signal distance map(SDM) to “describe the relation between the RSS(I)s and thegeographic distance” [17]. Every time ZCFG wants to locatea device, it identifies each of the n APs that observed a packetfrom the device. It then builds the n × n dimensional SDMaccording to the following equation, T = log(D)ST (SST )−1
where S,D are n×n dimensional matrices such that Sij , Dij
are the RSSI as observed by AP i from AP j and the physicaldistance between AP i and j respectively. ZCFG then finds thedistance between the device and each of the n APs accordingto the equation: dn = eTsn where sn is the column vectorof the observed RSSIs. Finally, ZCFG uses a multilaterationalgorithm to determine a location estimation.
IV. EXPERIMENTAL EVALUATION
Our goal is to perform a fair and accurate head-to-headcomparison of the three exemplar algorithms presented in theprevious section. We start by quantifying the performance ofthe algorithms in several locations in our office building. Moreimportantly, we attempt to understand limiting factors that willeffect most real-world deployments. We also identify severalpractical issues that were not discussed by the algorithms’authors in their original publications and, where possible, offerguidance in addressing them.
A. Experimental setupOur test building has approximately 145,000 square feet
over four stories and a basement. All experiments were con-ducted on the third floor which has ten access points dividedbetween channels 1, 6, and 11.
The APs in our building are not under our control. Thismeans that we are unable to use them to observe traffic.Instead, we monitor traffic using passive wireless monitors [4].There are 48 pairs of such monitors installed throughout ourbuilding, with approximately 12 pairs (24 distinct monitoringdevices with two radios each) located on each floor. Likethe initial AP placement, the monitor layout was specificallydesigned to provide a maximal coverage of each floor and,like the APs, should neither favor nor disfavor any specific
Fig. 3. Locations on the 3rd floor
location. The locations of the APs and monitors are indicatedin Figure 3.
For training the algorithms, we opportunistically use theexisting AP traffic (i.e., as though each AP was a client) as weknow the precise location of each AP. This is sufficient for con-structing the fingerprint database for the trivial, probabilistic,and TiX algorithms. Since the online training process of ZCFGrequires an architecture in which packets are transmitted aswell as received, we implement this by forcing each monitorto send pings to each of the others. Since there are 11 pairs ofmonitors but only 10 APs on the floor we consider, ZCFG will,in principle, have more training data to work with. This shouldgive ZCFG an advantage over the other three algorithms, butSection IV-D shows that this advantage does not materialize.
To build the fingerprint database we allow each algorithmto monitor the broadcast packets sent from the APs in thebuilding for approximately ten seconds. We find that thereis sufficient traffic in ten seconds to allow the mean RSSIvalue to be minimally affected by abnormally large or smallinstantaneous values. It is possible to locate a client from onlya single packet. However, we averaged two seconds worth ofRSSI values from the client. This time frame is a balancebetween reducing the effect of an outlier in the RSSI valuesand effectively locating a user even if they are in motion.Throughout this paper we report RSSI values instead of dB;RSSI values can be converted to dB by subtracting 95.
Probabilis2c Method (Moares and Nunes)
• Phase 1: Building a Radio Propaga2on Map – Overlay grid w/ probability that device is present – Probability determined by large-‐scale decay model
Class III: The final class of algorithms we study makes onesignificant conceptual addition to the third class of algorithms:It recognizes that each AP is able to map the distance fromitself to the client via RSSI, but, also, that the other (n − 1)APs can map the distance between the client and AP fromtheir own perspective [11], [17]. For example, AP4 can notonly estimate the distance between itself and the client, butalso the distance between the client and the other APs. Thusfinal estimation of the distance between AP4 and the clientis a (linear) combination of each AP’s distance estimationbetween the client and AP4. We chose ZCFG [17] as ourrepresentative algorithm for Class III because of its highlevel of sophistication and built-in resistance to error fromenvironmental noise.
ZCFG, like the previous algorithms, requires a trainingphase during which each AP records the average RSSI forall APs in range. Unlike the TiX algorithm, the calibration ofZCFG is done in real time. ZCFG uses a signal distance map(SDM) to “describe the relation between the RSS(I)s and thegeographic distance” [17]. Every time ZCFG wants to locatea device, it identifies each of the n APs that observed a packetfrom the device. It then builds the n × n dimensional SDMaccording to the following equation, T = log(D)ST (SST )−1
where S,D are n×n dimensional matrices such that Sij , Dij
are the RSSI as observed by AP i from AP j and the physicaldistance between AP i and j respectively. ZCFG then finds thedistance between the device and each of the n APs accordingto the equation: dn = eTsn where sn is the column vectorof the observed RSSIs. Finally, ZCFG uses a multilaterationalgorithm to determine a location estimation.
IV. EXPERIMENTAL EVALUATION
Our goal is to perform a fair and accurate head-to-headcomparison of the three exemplar algorithms presented in theprevious section. We start by quantifying the performance ofthe algorithms in several locations in our office building. Moreimportantly, we attempt to understand limiting factors that willeffect most real-world deployments. We also identify severalpractical issues that were not discussed by the algorithms’authors in their original publications and, where possible, offerguidance in addressing them.
A. Experimental setupOur test building has approximately 145,000 square feet
over four stories and a basement. All experiments were con-ducted on the third floor which has ten access points dividedbetween channels 1, 6, and 11.
The APs in our building are not under our control. Thismeans that we are unable to use them to observe traffic.Instead, we monitor traffic using passive wireless monitors [4].There are 48 pairs of such monitors installed throughout ourbuilding, with approximately 12 pairs (24 distinct monitoringdevices with two radios each) located on each floor. Likethe initial AP placement, the monitor layout was specificallydesigned to provide a maximal coverage of each floor and,like the APs, should neither favor nor disfavor any specific
Fig. 3. Locations on the 3rd floor
location. The locations of the APs and monitors are indicatedin Figure 3.
For training the algorithms, we opportunistically use theexisting AP traffic (i.e., as though each AP was a client) as weknow the precise location of each AP. This is sufficient for con-structing the fingerprint database for the trivial, probabilistic,and TiX algorithms. Since the online training process of ZCFGrequires an architecture in which packets are transmitted aswell as received, we implement this by forcing each monitorto send pings to each of the others. Since there are 11 pairs ofmonitors but only 10 APs on the floor we consider, ZCFG will,in principle, have more training data to work with. This shouldgive ZCFG an advantage over the other three algorithms, butSection IV-D shows that this advantage does not materialize.
To build the fingerprint database we allow each algorithmto monitor the broadcast packets sent from the APs in thebuilding for approximately ten seconds. We find that thereis sufficient traffic in ten seconds to allow the mean RSSIvalue to be minimally affected by abnormally large or smallinstantaneous values. It is possible to locate a client from onlya single packet. However, we averaged two seconds worth ofRSSI values from the client. This time frame is a balancebetween reducing the effect of an outlier in the RSSI valuesand effectively locating a user even if they are in motion.Throughout this paper we report RSSI values instead of dB;RSSI values can be converted to dB by subtracting 95.
Probabilis2c Method (Moares and Nunes)
• Phase 1: Building a Radio Propaga2on Map – Overlay grid w/ probability that device is present – Probability determined by large-‐scale decay model
• Trained on 1 signal-‐distance pair per RPM
Class III: The final class of algorithms we study makes onesignificant conceptual addition to the third class of algorithms:It recognizes that each AP is able to map the distance fromitself to the client via RSSI, but, also, that the other (n − 1)APs can map the distance between the client and AP fromtheir own perspective [11], [17]. For example, AP4 can notonly estimate the distance between itself and the client, butalso the distance between the client and the other APs. Thusfinal estimation of the distance between AP4 and the clientis a (linear) combination of each AP’s distance estimationbetween the client and AP4. We chose ZCFG [17] as ourrepresentative algorithm for Class III because of its highlevel of sophistication and built-in resistance to error fromenvironmental noise.
ZCFG, like the previous algorithms, requires a trainingphase during which each AP records the average RSSI forall APs in range. Unlike the TiX algorithm, the calibration ofZCFG is done in real time. ZCFG uses a signal distance map(SDM) to “describe the relation between the RSS(I)s and thegeographic distance” [17]. Every time ZCFG wants to locatea device, it identifies each of the n APs that observed a packetfrom the device. It then builds the n × n dimensional SDMaccording to the following equation, T = log(D)ST (SST )−1
where S,D are n×n dimensional matrices such that Sij , Dij
are the RSSI as observed by AP i from AP j and the physicaldistance between AP i and j respectively. ZCFG then finds thedistance between the device and each of the n APs accordingto the equation: dn = eTsn where sn is the column vectorof the observed RSSIs. Finally, ZCFG uses a multilaterationalgorithm to determine a location estimation.
IV. EXPERIMENTAL EVALUATION
Our goal is to perform a fair and accurate head-to-headcomparison of the three exemplar algorithms presented in theprevious section. We start by quantifying the performance ofthe algorithms in several locations in our office building. Moreimportantly, we attempt to understand limiting factors that willeffect most real-world deployments. We also identify severalpractical issues that were not discussed by the algorithms’authors in their original publications and, where possible, offerguidance in addressing them.
A. Experimental setupOur test building has approximately 145,000 square feet
over four stories and a basement. All experiments were con-ducted on the third floor which has ten access points dividedbetween channels 1, 6, and 11.
The APs in our building are not under our control. Thismeans that we are unable to use them to observe traffic.Instead, we monitor traffic using passive wireless monitors [4].There are 48 pairs of such monitors installed throughout ourbuilding, with approximately 12 pairs (24 distinct monitoringdevices with two radios each) located on each floor. Likethe initial AP placement, the monitor layout was specificallydesigned to provide a maximal coverage of each floor and,like the APs, should neither favor nor disfavor any specific
Fig. 3. Locations on the 3rd floor
location. The locations of the APs and monitors are indicatedin Figure 3.
For training the algorithms, we opportunistically use theexisting AP traffic (i.e., as though each AP was a client) as weknow the precise location of each AP. This is sufficient for con-structing the fingerprint database for the trivial, probabilistic,and TiX algorithms. Since the online training process of ZCFGrequires an architecture in which packets are transmitted aswell as received, we implement this by forcing each monitorto send pings to each of the others. Since there are 11 pairs ofmonitors but only 10 APs on the floor we consider, ZCFG will,in principle, have more training data to work with. This shouldgive ZCFG an advantage over the other three algorithms, butSection IV-D shows that this advantage does not materialize.
To build the fingerprint database we allow each algorithmto monitor the broadcast packets sent from the APs in thebuilding for approximately ten seconds. We find that thereis sufficient traffic in ten seconds to allow the mean RSSIvalue to be minimally affected by abnormally large or smallinstantaneous values. It is possible to locate a client from onlya single packet. However, we averaged two seconds worth ofRSSI values from the client. This time frame is a balancebetween reducing the effect of an outlier in the RSSI valuesand effectively locating a user even if they are in motion.Throughout this paper we report RSSI values instead of dB;RSSI values can be converted to dB by subtracting 95.
Probabilis2c Method
• Phase 1: Building a Radio Propaga2on Map
Class III: The final class of algorithms we study makes onesignificant conceptual addition to the third class of algorithms:It recognizes that each AP is able to map the distance fromitself to the client via RSSI, but, also, that the other (n − 1)APs can map the distance between the client and AP fromtheir own perspective [11], [17]. For example, AP4 can notonly estimate the distance between itself and the client, butalso the distance between the client and the other APs. Thusfinal estimation of the distance between AP4 and the clientis a (linear) combination of each AP’s distance estimationbetween the client and AP4. We chose ZCFG [17] as ourrepresentative algorithm for Class III because of its highlevel of sophistication and built-in resistance to error fromenvironmental noise.
ZCFG, like the previous algorithms, requires a trainingphase during which each AP records the average RSSI forall APs in range. Unlike the TiX algorithm, the calibration ofZCFG is done in real time. ZCFG uses a signal distance map(SDM) to “describe the relation between the RSS(I)s and thegeographic distance” [17]. Every time ZCFG wants to locatea device, it identifies each of the n APs that observed a packetfrom the device. It then builds the n × n dimensional SDMaccording to the following equation, T = log(D)ST (SST )−1
where S,D are n×n dimensional matrices such that Sij , Dij
are the RSSI as observed by AP i from AP j and the physicaldistance between AP i and j respectively. ZCFG then finds thedistance between the device and each of the n APs accordingto the equation: dn = eTsn where sn is the column vectorof the observed RSSIs. Finally, ZCFG uses a multilaterationalgorithm to determine a location estimation.
IV. EXPERIMENTAL EVALUATION
Our goal is to perform a fair and accurate head-to-headcomparison of the three exemplar algorithms presented in theprevious section. We start by quantifying the performance ofthe algorithms in several locations in our office building. Moreimportantly, we attempt to understand limiting factors that willeffect most real-world deployments. We also identify severalpractical issues that were not discussed by the algorithms’authors in their original publications and, where possible, offerguidance in addressing them.
A. Experimental setupOur test building has approximately 145,000 square feet
over four stories and a basement. All experiments were con-ducted on the third floor which has ten access points dividedbetween channels 1, 6, and 11.
The APs in our building are not under our control. Thismeans that we are unable to use them to observe traffic.Instead, we monitor traffic using passive wireless monitors [4].There are 48 pairs of such monitors installed throughout ourbuilding, with approximately 12 pairs (24 distinct monitoringdevices with two radios each) located on each floor. Likethe initial AP placement, the monitor layout was specificallydesigned to provide a maximal coverage of each floor and,like the APs, should neither favor nor disfavor any specific
Fig. 3. Locations on the 3rd floor
location. The locations of the APs and monitors are indicatedin Figure 3.
For training the algorithms, we opportunistically use theexisting AP traffic (i.e., as though each AP was a client) as weknow the precise location of each AP. This is sufficient for con-structing the fingerprint database for the trivial, probabilistic,and TiX algorithms. Since the online training process of ZCFGrequires an architecture in which packets are transmitted aswell as received, we implement this by forcing each monitorto send pings to each of the others. Since there are 11 pairs ofmonitors but only 10 APs on the floor we consider, ZCFG will,in principle, have more training data to work with. This shouldgive ZCFG an advantage over the other three algorithms, butSection IV-D shows that this advantage does not materialize.
To build the fingerprint database we allow each algorithmto monitor the broadcast packets sent from the APs in thebuilding for approximately ten seconds. We find that thereis sufficient traffic in ten seconds to allow the mean RSSIvalue to be minimally affected by abnormally large or smallinstantaneous values. It is possible to locate a client from onlya single packet. However, we averaged two seconds worth ofRSSI values from the client. This time frame is a balancebetween reducing the effect of an outlier in the RSSI valuesand effectively locating a user even if they are in motion.Throughout this paper we report RSSI values instead of dB;RSSI values can be converted to dB by subtracting 95.
Probabilis2c Method
• Phase 1: Building a Radio Propaga2on Map
Class III: The final class of algorithms we study makes onesignificant conceptual addition to the third class of algorithms:It recognizes that each AP is able to map the distance fromitself to the client via RSSI, but, also, that the other (n − 1)APs can map the distance between the client and AP fromtheir own perspective [11], [17]. For example, AP4 can notonly estimate the distance between itself and the client, butalso the distance between the client and the other APs. Thusfinal estimation of the distance between AP4 and the clientis a (linear) combination of each AP’s distance estimationbetween the client and AP4. We chose ZCFG [17] as ourrepresentative algorithm for Class III because of its highlevel of sophistication and built-in resistance to error fromenvironmental noise.
ZCFG, like the previous algorithms, requires a trainingphase during which each AP records the average RSSI forall APs in range. Unlike the TiX algorithm, the calibration ofZCFG is done in real time. ZCFG uses a signal distance map(SDM) to “describe the relation between the RSS(I)s and thegeographic distance” [17]. Every time ZCFG wants to locatea device, it identifies each of the n APs that observed a packetfrom the device. It then builds the n × n dimensional SDMaccording to the following equation, T = log(D)ST (SST )−1
where S,D are n×n dimensional matrices such that Sij , Dij
are the RSSI as observed by AP i from AP j and the physicaldistance between AP i and j respectively. ZCFG then finds thedistance between the device and each of the n APs accordingto the equation: dn = eTsn where sn is the column vectorof the observed RSSIs. Finally, ZCFG uses a multilaterationalgorithm to determine a location estimation.
IV. EXPERIMENTAL EVALUATION
Our goal is to perform a fair and accurate head-to-headcomparison of the three exemplar algorithms presented in theprevious section. We start by quantifying the performance ofthe algorithms in several locations in our office building. Moreimportantly, we attempt to understand limiting factors that willeffect most real-world deployments. We also identify severalpractical issues that were not discussed by the algorithms’authors in their original publications and, where possible, offerguidance in addressing them.
A. Experimental setupOur test building has approximately 145,000 square feet
over four stories and a basement. All experiments were con-ducted on the third floor which has ten access points dividedbetween channels 1, 6, and 11.
The APs in our building are not under our control. Thismeans that we are unable to use them to observe traffic.Instead, we monitor traffic using passive wireless monitors [4].There are 48 pairs of such monitors installed throughout ourbuilding, with approximately 12 pairs (24 distinct monitoringdevices with two radios each) located on each floor. Likethe initial AP placement, the monitor layout was specificallydesigned to provide a maximal coverage of each floor and,like the APs, should neither favor nor disfavor any specific
Fig. 3. Locations on the 3rd floor
location. The locations of the APs and monitors are indicatedin Figure 3.
For training the algorithms, we opportunistically use theexisting AP traffic (i.e., as though each AP was a client) as weknow the precise location of each AP. This is sufficient for con-structing the fingerprint database for the trivial, probabilistic,and TiX algorithms. Since the online training process of ZCFGrequires an architecture in which packets are transmitted aswell as received, we implement this by forcing each monitorto send pings to each of the others. Since there are 11 pairs ofmonitors but only 10 APs on the floor we consider, ZCFG will,in principle, have more training data to work with. This shouldgive ZCFG an advantage over the other three algorithms, butSection IV-D shows that this advantage does not materialize.
To build the fingerprint database we allow each algorithmto monitor the broadcast packets sent from the APs in thebuilding for approximately ten seconds. We find that thereis sufficient traffic in ten seconds to allow the mean RSSIvalue to be minimally affected by abnormally large or smallinstantaneous values. It is possible to locate a client from onlya single packet. However, we averaged two seconds worth ofRSSI values from the client. This time frame is a balancebetween reducing the effect of an outlier in the RSSI valuesand effectively locating a user even if they are in motion.Throughout this paper we report RSSI values instead of dB;RSSI values can be converted to dB by subtracting 95.
Probabilis2c Method
• Phase 1: Building a Radio Propaga2on Map
Probabilis2c Method
• Phase 1: Building a Radio Propaga2on Map
(a) AP4’s RPM (b) Final RPM derived from individual AP RPMs
Fig. 2. Example RPMs. Darker shading indicates higher probability.
APs that do not observe the strongest RSSI. To address thefirst issue probabilistic methods use data from the fingerprintdatabase, gathered during the training phase, to train a modelof decay that allows them to map an RSSI to locations otherthan the location of an AP. The probabilistic methods also useall n observed RSSIs when computing a final location.
Here, we consider the probabilistic algorithm proposed byde Moraes and Nunes [8] due to its reported mean errorof less than two meters and its foundation on a well-citedand understood RF propagation model [20]. The first step inthe client location phase of this algorithm is to estimate thedistance between each AP and the client. To do this each APthat hears the client builds a radio propagation map (RPM).An RPM is best thought of as a grid overlaid on the building’sfloor plan where each grid point, l = (x, y), has an associatedprobability, P [l|si], that the client is that location when signalstrength si is observed by AP i. P [l|si] is assumed to be aGaussian probability distribution, based on a large scale decaymodel where the model is trained on one entry in the offlinetraining database, Table I. An overly simple way to view theRPM is that the more similar an observed RSSI at location lis to the value the large scale decay model predicts, the morelikely P [l|si] will be large.
Once the probabilistic method has an RPM for each of then APs it combines them in the obvious manner, P [l|s] =Πn
i=1P [l|si] Finally a quick scan of the joint RPM is per-formed to find the point with the highest probability and thatpoint is declared as the client’s location. For example, if thealgorithm’s training phase produced Table I and subsequentlyobserved the client fingerprint in Figure 1(b), Figure 2(a)shows the RPM AP4 generates overlaid on our building’s floorplan while Figure 2(b) depicts the final RPM produced whenall of the four RPMs are combined into a final RPM.
Class II: The probabilistic algorithms leverages more infor-mation than our nearest-neighbor baseline: During the trainingphase, each AP observes (n−1) different RSSI-distance pairs,i.e., AP4 observes the RSSI values from AP1, AP2, andAP3. But, when an AP creates an RPM it uses only one
Device Distance (m) from sender Mean RSSI VarianceAP1 4.6 48.9 1.2AP2 10.0 42.3 0.8AP3 15.4 22.6 0.2
TABLE IAN EXAMPLE OF AP4’S OFFLINE TRAINING DATABASE.
RSSI/distance. Algorithms in our second class overcome thislimitation of probabilistic algorithms by using the reportedsignal strength at multiple APs simultaneously when attempt-ing to locate the client [10], [14], [18]. We have chosenTriangular Interpolation and eXtrapolation (TiX) [10] as therepresentative example for our second class of algorithms.
As before, the first step in TiX’s training phase is for eachAP to record the average RSSI of the other APs, i.e. build thefingerprint database shown in Table I. Each row consists of thedistance to the indicated AP as well as the mean and varianceof the signals between them (labeled D and B, respectively,in Figure 1(a)). In the second phase each AP fits these valuesto an exponential decay function of the form f(x) = a× ebx,where x is the RSSI, f(x) is the computed distance. Thisfunction gives each AP a mapping between RSSI and distancethat uses all of the information available to it during thetraining phase.
Once the algorithm is trained it can begin to locate devices.To locate a device it needs a packet to be heard by three APs.(We discuss the case where a client is heard by more than threeAPs in Section IV-C2.) It then applies the RSSI to distancefunction at each of the APs. Finally, it performs a modifiedversion of trilateration. The modified version of trilaterationhandles the case where noisy measurements cause there tonot be a single intersection of the radii of all the three circles.
For example, if TiX were to locate the client that producedthe fingerprint in Figure 1(b) it might use APs 2, 3, and 4. Fur-ther, if TiX used the training data in Table I it would find thatthe RSSI to distance curve for AP4 is f(x) = 69.2e−0.056x
and it would place the device 6.9 meters away from AP4. AP2
and AP3 compute similarly.
Probabilis2c Method
• Phase 1: Building a Radio Propaga2on Map
(a) AP4’s RPM (b) Final RPM derived from individual AP RPMs
Fig. 2. Example RPMs. Darker shading indicates higher probability.
APs that do not observe the strongest RSSI. To address thefirst issue probabilistic methods use data from the fingerprintdatabase, gathered during the training phase, to train a modelof decay that allows them to map an RSSI to locations otherthan the location of an AP. The probabilistic methods also useall n observed RSSIs when computing a final location.
Here, we consider the probabilistic algorithm proposed byde Moraes and Nunes [8] due to its reported mean errorof less than two meters and its foundation on a well-citedand understood RF propagation model [20]. The first step inthe client location phase of this algorithm is to estimate thedistance between each AP and the client. To do this each APthat hears the client builds a radio propagation map (RPM).An RPM is best thought of as a grid overlaid on the building’sfloor plan where each grid point, l = (x, y), has an associatedprobability, P [l|si], that the client is that location when signalstrength si is observed by AP i. P [l|si] is assumed to be aGaussian probability distribution, based on a large scale decaymodel where the model is trained on one entry in the offlinetraining database, Table I. An overly simple way to view theRPM is that the more similar an observed RSSI at location lis to the value the large scale decay model predicts, the morelikely P [l|si] will be large.
Once the probabilistic method has an RPM for each of then APs it combines them in the obvious manner, P [l|s] =Πn
i=1P [l|si] Finally a quick scan of the joint RPM is per-formed to find the point with the highest probability and thatpoint is declared as the client’s location. For example, if thealgorithm’s training phase produced Table I and subsequentlyobserved the client fingerprint in Figure 1(b), Figure 2(a)shows the RPM AP4 generates overlaid on our building’s floorplan while Figure 2(b) depicts the final RPM produced whenall of the four RPMs are combined into a final RPM.
Class II: The probabilistic algorithms leverages more infor-mation than our nearest-neighbor baseline: During the trainingphase, each AP observes (n−1) different RSSI-distance pairs,i.e., AP4 observes the RSSI values from AP1, AP2, andAP3. But, when an AP creates an RPM it uses only one
Device Distance (m) from sender Mean RSSI VarianceAP1 4.6 48.9 1.2AP2 10.0 42.3 0.8AP3 15.4 22.6 0.2
TABLE IAN EXAMPLE OF AP4’S OFFLINE TRAINING DATABASE.
RSSI/distance. Algorithms in our second class overcome thislimitation of probabilistic algorithms by using the reportedsignal strength at multiple APs simultaneously when attempt-ing to locate the client [10], [14], [18]. We have chosenTriangular Interpolation and eXtrapolation (TiX) [10] as therepresentative example for our second class of algorithms.
As before, the first step in TiX’s training phase is for eachAP to record the average RSSI of the other APs, i.e. build thefingerprint database shown in Table I. Each row consists of thedistance to the indicated AP as well as the mean and varianceof the signals between them (labeled D and B, respectively,in Figure 1(a)). In the second phase each AP fits these valuesto an exponential decay function of the form f(x) = a× ebx,where x is the RSSI, f(x) is the computed distance. Thisfunction gives each AP a mapping between RSSI and distancethat uses all of the information available to it during thetraining phase.
Once the algorithm is trained it can begin to locate devices.To locate a device it needs a packet to be heard by three APs.(We discuss the case where a client is heard by more than threeAPs in Section IV-C2.) It then applies the RSSI to distancefunction at each of the APs. Finally, it performs a modifiedversion of trilateration. The modified version of trilaterationhandles the case where noisy measurements cause there tonot be a single intersection of the radii of all the three circles.
For example, if TiX were to locate the client that producedthe fingerprint in Figure 1(b) it might use APs 2, 3, and 4. Fur-ther, if TiX used the training data in Table I it would find thatthe RSSI to distance curve for AP4 is f(x) = 69.2e−0.056x
and it would place the device 6.9 meters away from AP4. AP2
and AP3 compute similarly.
10%
Probabilis2c Method
• Phase 1: Building a Radio Propaga2on Map
(a) AP4’s RPM (b) Final RPM derived from individual AP RPMs
Fig. 2. Example RPMs. Darker shading indicates higher probability.
APs that do not observe the strongest RSSI. To address thefirst issue probabilistic methods use data from the fingerprintdatabase, gathered during the training phase, to train a modelof decay that allows them to map an RSSI to locations otherthan the location of an AP. The probabilistic methods also useall n observed RSSIs when computing a final location.
Here, we consider the probabilistic algorithm proposed byde Moraes and Nunes [8] due to its reported mean errorof less than two meters and its foundation on a well-citedand understood RF propagation model [20]. The first step inthe client location phase of this algorithm is to estimate thedistance between each AP and the client. To do this each APthat hears the client builds a radio propagation map (RPM).An RPM is best thought of as a grid overlaid on the building’sfloor plan where each grid point, l = (x, y), has an associatedprobability, P [l|si], that the client is that location when signalstrength si is observed by AP i. P [l|si] is assumed to be aGaussian probability distribution, based on a large scale decaymodel where the model is trained on one entry in the offlinetraining database, Table I. An overly simple way to view theRPM is that the more similar an observed RSSI at location lis to the value the large scale decay model predicts, the morelikely P [l|si] will be large.
Once the probabilistic method has an RPM for each of then APs it combines them in the obvious manner, P [l|s] =Πn
i=1P [l|si] Finally a quick scan of the joint RPM is per-formed to find the point with the highest probability and thatpoint is declared as the client’s location. For example, if thealgorithm’s training phase produced Table I and subsequentlyobserved the client fingerprint in Figure 1(b), Figure 2(a)shows the RPM AP4 generates overlaid on our building’s floorplan while Figure 2(b) depicts the final RPM produced whenall of the four RPMs are combined into a final RPM.
Class II: The probabilistic algorithms leverages more infor-mation than our nearest-neighbor baseline: During the trainingphase, each AP observes (n−1) different RSSI-distance pairs,i.e., AP4 observes the RSSI values from AP1, AP2, andAP3. But, when an AP creates an RPM it uses only one
Device Distance (m) from sender Mean RSSI VarianceAP1 4.6 48.9 1.2AP2 10.0 42.3 0.8AP3 15.4 22.6 0.2
TABLE IAN EXAMPLE OF AP4’S OFFLINE TRAINING DATABASE.
RSSI/distance. Algorithms in our second class overcome thislimitation of probabilistic algorithms by using the reportedsignal strength at multiple APs simultaneously when attempt-ing to locate the client [10], [14], [18]. We have chosenTriangular Interpolation and eXtrapolation (TiX) [10] as therepresentative example for our second class of algorithms.
As before, the first step in TiX’s training phase is for eachAP to record the average RSSI of the other APs, i.e. build thefingerprint database shown in Table I. Each row consists of thedistance to the indicated AP as well as the mean and varianceof the signals between them (labeled D and B, respectively,in Figure 1(a)). In the second phase each AP fits these valuesto an exponential decay function of the form f(x) = a× ebx,where x is the RSSI, f(x) is the computed distance. Thisfunction gives each AP a mapping between RSSI and distancethat uses all of the information available to it during thetraining phase.
Once the algorithm is trained it can begin to locate devices.To locate a device it needs a packet to be heard by three APs.(We discuss the case where a client is heard by more than threeAPs in Section IV-C2.) It then applies the RSSI to distancefunction at each of the APs. Finally, it performs a modifiedversion of trilateration. The modified version of trilaterationhandles the case where noisy measurements cause there tonot be a single intersection of the radii of all the three circles.
For example, if TiX were to locate the client that producedthe fingerprint in Figure 1(b) it might use APs 2, 3, and 4. Fur-ther, if TiX used the training data in Table I it would find thatthe RSSI to distance curve for AP4 is f(x) = 69.2e−0.056x
and it would place the device 6.9 meters away from AP4. AP2
and AP3 compute similarly.
10%
90%
Probabilis2c Method
• Phase 2: Combing RPMs for a final es2ma2on – Final probability is the product across all RPMS
(a) AP4’s RPM (b) Final RPM derived from individual AP RPMs
Fig. 2. Example RPMs. Darker shading indicates higher probability.
APs that do not observe the strongest RSSI. To address thefirst issue probabilistic methods use data from the fingerprintdatabase, gathered during the training phase, to train a modelof decay that allows them to map an RSSI to locations otherthan the location of an AP. The probabilistic methods also useall n observed RSSIs when computing a final location.
Here, we consider the probabilistic algorithm proposed byde Moraes and Nunes [8] due to its reported mean errorof less than two meters and its foundation on a well-citedand understood RF propagation model [20]. The first step inthe client location phase of this algorithm is to estimate thedistance between each AP and the client. To do this each APthat hears the client builds a radio propagation map (RPM).An RPM is best thought of as a grid overlaid on the building’sfloor plan where each grid point, l = (x, y), has an associatedprobability, P [l|si], that the client is that location when signalstrength si is observed by AP i. P [l|si] is assumed to be aGaussian probability distribution, based on a large scale decaymodel where the model is trained on one entry in the offlinetraining database, Table I. An overly simple way to view theRPM is that the more similar an observed RSSI at location lis to the value the large scale decay model predicts, the morelikely P [l|si] will be large.
Once the probabilistic method has an RPM for each of then APs it combines them in the obvious manner, P [l|s] =Πn
i=1P [l|si] Finally a quick scan of the joint RPM is per-formed to find the point with the highest probability and thatpoint is declared as the client’s location. For example, if thealgorithm’s training phase produced Table I and subsequentlyobserved the client fingerprint in Figure 1(b), Figure 2(a)shows the RPM AP4 generates overlaid on our building’s floorplan while Figure 2(b) depicts the final RPM produced whenall of the four RPMs are combined into a final RPM.
Class II: The probabilistic algorithms leverages more infor-mation than our nearest-neighbor baseline: During the trainingphase, each AP observes (n−1) different RSSI-distance pairs,i.e., AP4 observes the RSSI values from AP1, AP2, andAP3. But, when an AP creates an RPM it uses only one
Device Distance (m) from sender Mean RSSI VarianceAP1 4.6 48.9 1.2AP2 10.0 42.3 0.8AP3 15.4 22.6 0.2
TABLE IAN EXAMPLE OF AP4’S OFFLINE TRAINING DATABASE.
RSSI/distance. Algorithms in our second class overcome thislimitation of probabilistic algorithms by using the reportedsignal strength at multiple APs simultaneously when attempt-ing to locate the client [10], [14], [18]. We have chosenTriangular Interpolation and eXtrapolation (TiX) [10] as therepresentative example for our second class of algorithms.
As before, the first step in TiX’s training phase is for eachAP to record the average RSSI of the other APs, i.e. build thefingerprint database shown in Table I. Each row consists of thedistance to the indicated AP as well as the mean and varianceof the signals between them (labeled D and B, respectively,in Figure 1(a)). In the second phase each AP fits these valuesto an exponential decay function of the form f(x) = a× ebx,where x is the RSSI, f(x) is the computed distance. Thisfunction gives each AP a mapping between RSSI and distancethat uses all of the information available to it during thetraining phase.
Once the algorithm is trained it can begin to locate devices.To locate a device it needs a packet to be heard by three APs.(We discuss the case where a client is heard by more than threeAPs in Section IV-C2.) It then applies the RSSI to distancefunction at each of the APs. Finally, it performs a modifiedversion of trilateration. The modified version of trilaterationhandles the case where noisy measurements cause there tonot be a single intersection of the radii of all the three circles.
For example, if TiX were to locate the client that producedthe fingerprint in Figure 1(b) it might use APs 2, 3, and 4. Fur-ther, if TiX used the training data in Table I it would find thatthe RSSI to distance curve for AP4 is f(x) = 69.2e−0.056x
and it would place the device 6.9 meters away from AP4. AP2
and AP3 compute similarly.
Probabilis2c Method
• Phase 2: Combing RPMs for a final es2ma2on – Final probability is the product across all RPMS
(a) AP4’s RPM (b) Final RPM derived from individual AP RPMs
Fig. 2. Example RPMs. Darker shading indicates higher probability.
APs that do not observe the strongest RSSI. To address thefirst issue probabilistic methods use data from the fingerprintdatabase, gathered during the training phase, to train a modelof decay that allows them to map an RSSI to locations otherthan the location of an AP. The probabilistic methods also useall n observed RSSIs when computing a final location.
Here, we consider the probabilistic algorithm proposed byde Moraes and Nunes [8] due to its reported mean errorof less than two meters and its foundation on a well-citedand understood RF propagation model [20]. The first step inthe client location phase of this algorithm is to estimate thedistance between each AP and the client. To do this each APthat hears the client builds a radio propagation map (RPM).An RPM is best thought of as a grid overlaid on the building’sfloor plan where each grid point, l = (x, y), has an associatedprobability, P [l|si], that the client is that location when signalstrength si is observed by AP i. P [l|si] is assumed to be aGaussian probability distribution, based on a large scale decaymodel where the model is trained on one entry in the offlinetraining database, Table I. An overly simple way to view theRPM is that the more similar an observed RSSI at location lis to the value the large scale decay model predicts, the morelikely P [l|si] will be large.
Once the probabilistic method has an RPM for each of then APs it combines them in the obvious manner, P [l|s] =Πn
i=1P [l|si] Finally a quick scan of the joint RPM is per-formed to find the point with the highest probability and thatpoint is declared as the client’s location. For example, if thealgorithm’s training phase produced Table I and subsequentlyobserved the client fingerprint in Figure 1(b), Figure 2(a)shows the RPM AP4 generates overlaid on our building’s floorplan while Figure 2(b) depicts the final RPM produced whenall of the four RPMs are combined into a final RPM.
Class II: The probabilistic algorithms leverages more infor-mation than our nearest-neighbor baseline: During the trainingphase, each AP observes (n−1) different RSSI-distance pairs,i.e., AP4 observes the RSSI values from AP1, AP2, andAP3. But, when an AP creates an RPM it uses only one
Device Distance (m) from sender Mean RSSI VarianceAP1 4.6 48.9 1.2AP2 10.0 42.3 0.8AP3 15.4 22.6 0.2
TABLE IAN EXAMPLE OF AP4’S OFFLINE TRAINING DATABASE.
RSSI/distance. Algorithms in our second class overcome thislimitation of probabilistic algorithms by using the reportedsignal strength at multiple APs simultaneously when attempt-ing to locate the client [10], [14], [18]. We have chosenTriangular Interpolation and eXtrapolation (TiX) [10] as therepresentative example for our second class of algorithms.
As before, the first step in TiX’s training phase is for eachAP to record the average RSSI of the other APs, i.e. build thefingerprint database shown in Table I. Each row consists of thedistance to the indicated AP as well as the mean and varianceof the signals between them (labeled D and B, respectively,in Figure 1(a)). In the second phase each AP fits these valuesto an exponential decay function of the form f(x) = a× ebx,where x is the RSSI, f(x) is the computed distance. Thisfunction gives each AP a mapping between RSSI and distancethat uses all of the information available to it during thetraining phase.
Once the algorithm is trained it can begin to locate devices.To locate a device it needs a packet to be heard by three APs.(We discuss the case where a client is heard by more than threeAPs in Section IV-C2.) It then applies the RSSI to distancefunction at each of the APs. Finally, it performs a modifiedversion of trilateration. The modified version of trilaterationhandles the case where noisy measurements cause there tonot be a single intersection of the radii of all the three circles.
For example, if TiX were to locate the client that producedthe fingerprint in Figure 1(b) it might use APs 2, 3, and 4. Fur-ther, if TiX used the training data in Table I it would find thatthe RSSI to distance curve for AP4 is f(x) = 69.2e−0.056x
and it would place the device 6.9 meters away from AP4. AP2
and AP3 compute similarly.
85%
Probabilis2c Method
• Phase 2: Combing RPMs for a final es2ma2on – Final probability is the product across all RPMS
(a) AP4’s RPM (b) Final RPM derived from individual AP RPMs
Fig. 2. Example RPMs. Darker shading indicates higher probability.
APs that do not observe the strongest RSSI. To address thefirst issue probabilistic methods use data from the fingerprintdatabase, gathered during the training phase, to train a modelof decay that allows them to map an RSSI to locations otherthan the location of an AP. The probabilistic methods also useall n observed RSSIs when computing a final location.
Here, we consider the probabilistic algorithm proposed byde Moraes and Nunes [8] due to its reported mean errorof less than two meters and its foundation on a well-citedand understood RF propagation model [20]. The first step inthe client location phase of this algorithm is to estimate thedistance between each AP and the client. To do this each APthat hears the client builds a radio propagation map (RPM).An RPM is best thought of as a grid overlaid on the building’sfloor plan where each grid point, l = (x, y), has an associatedprobability, P [l|si], that the client is that location when signalstrength si is observed by AP i. P [l|si] is assumed to be aGaussian probability distribution, based on a large scale decaymodel where the model is trained on one entry in the offlinetraining database, Table I. An overly simple way to view theRPM is that the more similar an observed RSSI at location lis to the value the large scale decay model predicts, the morelikely P [l|si] will be large.
Once the probabilistic method has an RPM for each of then APs it combines them in the obvious manner, P [l|s] =Πn
i=1P [l|si] Finally a quick scan of the joint RPM is per-formed to find the point with the highest probability and thatpoint is declared as the client’s location. For example, if thealgorithm’s training phase produced Table I and subsequentlyobserved the client fingerprint in Figure 1(b), Figure 2(a)shows the RPM AP4 generates overlaid on our building’s floorplan while Figure 2(b) depicts the final RPM produced whenall of the four RPMs are combined into a final RPM.
Class II: The probabilistic algorithms leverages more infor-mation than our nearest-neighbor baseline: During the trainingphase, each AP observes (n−1) different RSSI-distance pairs,i.e., AP4 observes the RSSI values from AP1, AP2, andAP3. But, when an AP creates an RPM it uses only one
Device Distance (m) from sender Mean RSSI VarianceAP1 4.6 48.9 1.2AP2 10.0 42.3 0.8AP3 15.4 22.6 0.2
TABLE IAN EXAMPLE OF AP4’S OFFLINE TRAINING DATABASE.
RSSI/distance. Algorithms in our second class overcome thislimitation of probabilistic algorithms by using the reportedsignal strength at multiple APs simultaneously when attempt-ing to locate the client [10], [14], [18]. We have chosenTriangular Interpolation and eXtrapolation (TiX) [10] as therepresentative example for our second class of algorithms.
As before, the first step in TiX’s training phase is for eachAP to record the average RSSI of the other APs, i.e. build thefingerprint database shown in Table I. Each row consists of thedistance to the indicated AP as well as the mean and varianceof the signals between them (labeled D and B, respectively,in Figure 1(a)). In the second phase each AP fits these valuesto an exponential decay function of the form f(x) = a× ebx,where x is the RSSI, f(x) is the computed distance. Thisfunction gives each AP a mapping between RSSI and distancethat uses all of the information available to it during thetraining phase.
Once the algorithm is trained it can begin to locate devices.To locate a device it needs a packet to be heard by three APs.(We discuss the case where a client is heard by more than threeAPs in Section IV-C2.) It then applies the RSSI to distancefunction at each of the APs. Finally, it performs a modifiedversion of trilateration. The modified version of trilaterationhandles the case where noisy measurements cause there tonot be a single intersection of the radii of all the three circles.
For example, if TiX were to locate the client that producedthe fingerprint in Figure 1(b) it might use APs 2, 3, and 4. Fur-ther, if TiX used the training data in Table I it would find thatthe RSSI to distance curve for AP4 is f(x) = 69.2e−0.056x
and it would place the device 6.9 meters away from AP4. AP2
and AP3 compute similarly.
85% 95%
(a) AP4’s RPM (b) Final RPM derived from individual AP RPMs
Fig. 2. Example RPMs. Darker shading indicates higher probability.
APs that do not observe the strongest RSSI. To address thefirst issue probabilistic methods use data from the fingerprintdatabase, gathered during the training phase, to train a modelof decay that allows them to map an RSSI to locations otherthan the location of an AP. The probabilistic methods also useall n observed RSSIs when computing a final location.
Here, we consider the probabilistic algorithm proposed byde Moraes and Nunes [8] due to its reported mean errorof less than two meters and its foundation on a well-citedand understood RF propagation model [20]. The first step inthe client location phase of this algorithm is to estimate thedistance between each AP and the client. To do this each APthat hears the client builds a radio propagation map (RPM).An RPM is best thought of as a grid overlaid on the building’sfloor plan where each grid point, l = (x, y), has an associatedprobability, P [l|si], that the client is that location when signalstrength si is observed by AP i. P [l|si] is assumed to be aGaussian probability distribution, based on a large scale decaymodel where the model is trained on one entry in the offlinetraining database, Table I. An overly simple way to view theRPM is that the more similar an observed RSSI at location lis to the value the large scale decay model predicts, the morelikely P [l|si] will be large.
Once the probabilistic method has an RPM for each of then APs it combines them in the obvious manner, P [l|s] =Πn
i=1P [l|si] Finally a quick scan of the joint RPM is per-formed to find the point with the highest probability and thatpoint is declared as the client’s location. For example, if thealgorithm’s training phase produced Table I and subsequentlyobserved the client fingerprint in Figure 1(b), Figure 2(a)shows the RPM AP4 generates overlaid on our building’s floorplan while Figure 2(b) depicts the final RPM produced whenall of the four RPMs are combined into a final RPM.
Class II: The probabilistic algorithms leverages more infor-mation than our nearest-neighbor baseline: During the trainingphase, each AP observes (n−1) different RSSI-distance pairs,i.e., AP4 observes the RSSI values from AP1, AP2, andAP3. But, when an AP creates an RPM it uses only one
Device Distance (m) from sender Mean RSSI VarianceAP1 4.6 48.9 1.2AP2 10.0 42.3 0.8AP3 15.4 22.6 0.2
TABLE IAN EXAMPLE OF AP4’S OFFLINE TRAINING DATABASE.
RSSI/distance. Algorithms in our second class overcome thislimitation of probabilistic algorithms by using the reportedsignal strength at multiple APs simultaneously when attempt-ing to locate the client [10], [14], [18]. We have chosenTriangular Interpolation and eXtrapolation (TiX) [10] as therepresentative example for our second class of algorithms.
As before, the first step in TiX’s training phase is for eachAP to record the average RSSI of the other APs, i.e. build thefingerprint database shown in Table I. Each row consists of thedistance to the indicated AP as well as the mean and varianceof the signals between them (labeled D and B, respectively,in Figure 1(a)). In the second phase each AP fits these valuesto an exponential decay function of the form f(x) = a× ebx,where x is the RSSI, f(x) is the computed distance. Thisfunction gives each AP a mapping between RSSI and distancethat uses all of the information available to it during thetraining phase.
Once the algorithm is trained it can begin to locate devices.To locate a device it needs a packet to be heard by three APs.(We discuss the case where a client is heard by more than threeAPs in Section IV-C2.) It then applies the RSSI to distancefunction at each of the APs. Finally, it performs a modifiedversion of trilateration. The modified version of trilaterationhandles the case where noisy measurements cause there tonot be a single intersection of the radii of all the three circles.
For example, if TiX were to locate the client that producedthe fingerprint in Figure 1(b) it might use APs 2, 3, and 4. Fur-ther, if TiX used the training data in Table I it would find thatthe RSSI to distance curve for AP4 is f(x) = 69.2e−0.056x
and it would place the device 6.9 meters away from AP4. AP2
and AP3 compute similarly.
Probabilis2c Method
• Phase 2: Combing RPMs for a final es2ma2on – Final probability is the product across all RPMS
(a) AP4’s RPM (b) Final RPM derived from individual AP RPMs
Fig. 2. Example RPMs. Darker shading indicates higher probability.
APs that do not observe the strongest RSSI. To address thefirst issue probabilistic methods use data from the fingerprintdatabase, gathered during the training phase, to train a modelof decay that allows them to map an RSSI to locations otherthan the location of an AP. The probabilistic methods also useall n observed RSSIs when computing a final location.
Here, we consider the probabilistic algorithm proposed byde Moraes and Nunes [8] due to its reported mean errorof less than two meters and its foundation on a well-citedand understood RF propagation model [20]. The first step inthe client location phase of this algorithm is to estimate thedistance between each AP and the client. To do this each APthat hears the client builds a radio propagation map (RPM).An RPM is best thought of as a grid overlaid on the building’sfloor plan where each grid point, l = (x, y), has an associatedprobability, P [l|si], that the client is that location when signalstrength si is observed by AP i. P [l|si] is assumed to be aGaussian probability distribution, based on a large scale decaymodel where the model is trained on one entry in the offlinetraining database, Table I. An overly simple way to view theRPM is that the more similar an observed RSSI at location lis to the value the large scale decay model predicts, the morelikely P [l|si] will be large.
Once the probabilistic method has an RPM for each of then APs it combines them in the obvious manner, P [l|s] =Πn
i=1P [l|si] Finally a quick scan of the joint RPM is per-formed to find the point with the highest probability and thatpoint is declared as the client’s location. For example, if thealgorithm’s training phase produced Table I and subsequentlyobserved the client fingerprint in Figure 1(b), Figure 2(a)shows the RPM AP4 generates overlaid on our building’s floorplan while Figure 2(b) depicts the final RPM produced whenall of the four RPMs are combined into a final RPM.
Class II: The probabilistic algorithms leverages more infor-mation than our nearest-neighbor baseline: During the trainingphase, each AP observes (n−1) different RSSI-distance pairs,i.e., AP4 observes the RSSI values from AP1, AP2, andAP3. But, when an AP creates an RPM it uses only one
Device Distance (m) from sender Mean RSSI VarianceAP1 4.6 48.9 1.2AP2 10.0 42.3 0.8AP3 15.4 22.6 0.2
TABLE IAN EXAMPLE OF AP4’S OFFLINE TRAINING DATABASE.
RSSI/distance. Algorithms in our second class overcome thislimitation of probabilistic algorithms by using the reportedsignal strength at multiple APs simultaneously when attempt-ing to locate the client [10], [14], [18]. We have chosenTriangular Interpolation and eXtrapolation (TiX) [10] as therepresentative example for our second class of algorithms.
As before, the first step in TiX’s training phase is for eachAP to record the average RSSI of the other APs, i.e. build thefingerprint database shown in Table I. Each row consists of thedistance to the indicated AP as well as the mean and varianceof the signals between them (labeled D and B, respectively,in Figure 1(a)). In the second phase each AP fits these valuesto an exponential decay function of the form f(x) = a× ebx,where x is the RSSI, f(x) is the computed distance. Thisfunction gives each AP a mapping between RSSI and distancethat uses all of the information available to it during thetraining phase.
Once the algorithm is trained it can begin to locate devices.To locate a device it needs a packet to be heard by three APs.(We discuss the case where a client is heard by more than threeAPs in Section IV-C2.) It then applies the RSSI to distancefunction at each of the APs. Finally, it performs a modifiedversion of trilateration. The modified version of trilaterationhandles the case where noisy measurements cause there tonot be a single intersection of the radii of all the three circles.
For example, if TiX were to locate the client that producedthe fingerprint in Figure 1(b) it might use APs 2, 3, and 4. Fur-ther, if TiX used the training data in Table I it would find thatthe RSSI to distance curve for AP4 is f(x) = 69.2e−0.056x
and it would place the device 6.9 meters away from AP4. AP2
and AP3 compute similarly.
Probabilis2c Method
• Phase 2: Combing RPMs for a final es2ma2on – Final probability is the product across all RPMS
81%
Triangular Interpola2on and eXtrapola2on (Gwon and Jain)
• Phase I: Distance Es2ma2on – Exponen2al decay model
• – Least-‐squares fit over all RSSI-‐distance pairs
F(x) =! ! e("!x )
Triangular Interpola2on and eXtrapola2on (Gwon and Jain)
• Phase I: Distance Es2ma2on – Exponen2al decay model
• – Least-‐squares fit over all RSSI-‐distance pairs
Sender Distance (M)
Mean RSSI
Varia2on
AP 2 4.6 48.9 1.2
AP 1 10.0 42.3 0.8
AP 4 15.4 12.1 0.7
F(x) =! ! e("!x )
Triangular Interpola2on and eXtrapola2on (Gwon and Jain)
• Phase I: Distance Es2ma2on – Exponen2al decay model
• – Least-‐squares fit over all RSSI-‐distance pairs
– For AP3:
Sender Distance (M)
Mean RSSI
Varia2on
AP 2 4.6 48.9 1.2
AP 1 10.0 42.3 0.8
AP 4 15.4 12.1 0.7
F(x) =! ! e("!x )
F(x) = 78! e("0.09!x )
Triangular Interpola2on and eXtrapola2on (Gwon and Jain)
• Phase I: Distance Es2ma2on – Exponen2al decay model
• – Least-‐squares fit over all RSSI-‐distance pairs
– For AP3:
Sender Distance (M)
Mean RSSI
Varia2on
AP 2 4.6 48.9 1.2
AP 1 10.0 42.3 0.8
AP 4 15.4 12.1 0.7
technologies including Bluetooth signals [12], RFID proximity[19], and ultrasonic acoustic waves [3], among others.
Systems that use signal strength measurements, or, morespecifically in the case of WiFi, the received signal strengthindication (RSSI)—a unit-less measure of signal power asso-ciated with a packet as measured by a receiving device—canbe subdivided into two categories: those that require manualcalibration and those that do not. While the goal of this paperis to evaluate self-calibrating 802.11-based location systems,we briefly summarize the approaches based upon manualcalibration for context.
Many of the early (and still most accurate) 802.11-basedlocation systems require manual calibration. During this initialsite survey an installer collects client fingerprints (RSSI valuesfor packets broadcast by APs) at a large number of knownlocations to build up a fine-grained fingerprint database thatcan be used for future client localization. In general thesesystems have three major manifestations: point matching,probabilistic matching, and Bayesian networks.
Point matching, also known as nearest neighbor in signal
strength [2], starts by capturing a client fingerprint and com-paring it against the initial fingerprint database scanning for thefingerprint that minimizes a similarity metric, often Euclideandistance. The client’s location is determined to be the locationof the offline fingerprint that minimizes the metric.
The second set of methods once again starts with the usertaking a client fingerprint. Then the user computes the proba-bility of receiving the fingerprint at each location, l = (x, y),in the building. Assuming that the distribution of RSSIs in thefingerprint is Gaussian, it is straightforward to determine an lsuch that P (s|l) is maximized [1], [25].
Algorithms based on Bayesian networks are the most com-plicated. They depend on building a Hidden Markov Model(HMM) in which the RSSIs represent the observed state, andthe distance from the APs (and hence the x, y coordinatesof the client) are the hidden state. The relationship betweenvisible and hidden states is encoded via a signal decay modeland then the fingerprint database is used to train the parametersof the model [16], [21].
Chintalapudi et al. recently proposed EZ Localization as amethodology for calibration-free wireless localization in anindoor environment [6]. EZ uses WiFi-enabled smart phonesto take fingerprints at unknown locations in a building. Ratherthan using the location of APs to ground their model, however,they leverage the GPS units available on the client devices.When a phone is able to get a GPS lock they opportunisticallyuse the measurement as an anchor point. They then treat thefingerprints as constraints to be solved by a genetic algorithm.The authors report that EZ has low error in their environment,but remains unable to surpass the accuracy of systems thatrequire manual calibration.
III. SELF-CALIBRATING ALGORITHMS
Because manually calibrated systems may require takinghundreds of fingerprints per floor, a second class of systemsthat requires no manual calibration has been proposed. Instead,
1
3 4
2
(a) Simplified network diagram
Device Mean RSSIAP1 24.4AP2 47.8AP3 45.5AP4 40.6
(b) Client fingerprint
Fig. 1. An example network and fingerprint corresponding to signal A
these self-calibrating systems leverage information regardingthe location of WiFi access points to conduct a trainingphase in which they collect system fingerprints, or sets ofRSSI measurements taken of (and by) the stationary APs inthe network. After the self-calibration stage, the fingerprintdatabase is combined with pre-configured information aboutthe physical location of the APs to compute a model of therelationship between RSSI and physical location. This modelis then used to answer on-line queries about client location.
A. Overview
It is convenient to think of self-calibrating algorithms asfitting into three categories based on the amount of data usedto perform calibration. To assist the reader in understandingeach algorithm we give examples of how they work in a simplewireless network. Our explanatory network, diagrammed inFigure 1(a), has four APs and one client. Each AP knowsthe distance between it and every other AP, for example thedistance between AP3 and AP4 is ten meters and labeled D.The 802.11 signal emitted by the client when sending a packetis labeled A. The RSSI values for one transmission as seen bythe four APs, i.e. a client fingerprint, are documented in thetable in Figure 1(b). Conversely, the 802.11 signal generatedwhen an AP sends a packet into the network is labeled C.When an AP sends a packet it is heard by everyone withinrange, though all but the intended destination will usuallyignore it when the network is operational. In the systemtraining phase, however, APs do not ignore packets sent fromother APs, labeled B, but instead record their RSSI values.Table I shows an example of training data that might observedby AP4 in this network.
Baseline: In order to provide some context in which tointerpret the performance of the localization system classes, weconsider a simple, calibration-free baseline that leverages onlya single RSSI from a client transmission, labeled A in Figure1(a). This nearest-neighbor approach requires no calibration:to locate a client it simply determines which AP heard theclient’s transmission with the strongest RSSI and declares theclient to be co-located with that AP. For example if the systemobtained the fingerprint in Figure 1(b) it would estimate theclient to be at the same location as AP2.
Class I: Our baseline algorithm, while simple, has twomajor limitations. First, it can choose only locations occupiedby an AP. Second, it ignores all information from the (n− 1)
F(x) =! ! e("!x )
F(x) = 78! e("0.09!x )
Triangular Interpola2on and eXtrapola2on (Gwon and Jain)
• Phase I: Distance Es2ma2on – Exponen2al decay model
• – Least-‐squares fit over all RSSI-‐distance pairs
– For AP3: – For AP3: Meters
Sender Distance (M)
Mean RSSI
Varia2on
AP 2 4.6 48.9 1.2
AP 1 10.0 42.3 0.8
AP 4 15.4 12.1 0.7
technologies including Bluetooth signals [12], RFID proximity[19], and ultrasonic acoustic waves [3], among others.
Systems that use signal strength measurements, or, morespecifically in the case of WiFi, the received signal strengthindication (RSSI)—a unit-less measure of signal power asso-ciated with a packet as measured by a receiving device—canbe subdivided into two categories: those that require manualcalibration and those that do not. While the goal of this paperis to evaluate self-calibrating 802.11-based location systems,we briefly summarize the approaches based upon manualcalibration for context.
Many of the early (and still most accurate) 802.11-basedlocation systems require manual calibration. During this initialsite survey an installer collects client fingerprints (RSSI valuesfor packets broadcast by APs) at a large number of knownlocations to build up a fine-grained fingerprint database thatcan be used for future client localization. In general thesesystems have three major manifestations: point matching,probabilistic matching, and Bayesian networks.
Point matching, also known as nearest neighbor in signal
strength [2], starts by capturing a client fingerprint and com-paring it against the initial fingerprint database scanning for thefingerprint that minimizes a similarity metric, often Euclideandistance. The client’s location is determined to be the locationof the offline fingerprint that minimizes the metric.
The second set of methods once again starts with the usertaking a client fingerprint. Then the user computes the proba-bility of receiving the fingerprint at each location, l = (x, y),in the building. Assuming that the distribution of RSSIs in thefingerprint is Gaussian, it is straightforward to determine an lsuch that P (s|l) is maximized [1], [25].
Algorithms based on Bayesian networks are the most com-plicated. They depend on building a Hidden Markov Model(HMM) in which the RSSIs represent the observed state, andthe distance from the APs (and hence the x, y coordinatesof the client) are the hidden state. The relationship betweenvisible and hidden states is encoded via a signal decay modeland then the fingerprint database is used to train the parametersof the model [16], [21].
Chintalapudi et al. recently proposed EZ Localization as amethodology for calibration-free wireless localization in anindoor environment [6]. EZ uses WiFi-enabled smart phonesto take fingerprints at unknown locations in a building. Ratherthan using the location of APs to ground their model, however,they leverage the GPS units available on the client devices.When a phone is able to get a GPS lock they opportunisticallyuse the measurement as an anchor point. They then treat thefingerprints as constraints to be solved by a genetic algorithm.The authors report that EZ has low error in their environment,but remains unable to surpass the accuracy of systems thatrequire manual calibration.
III. SELF-CALIBRATING ALGORITHMS
Because manually calibrated systems may require takinghundreds of fingerprints per floor, a second class of systemsthat requires no manual calibration has been proposed. Instead,
1
3 4
2
(a) Simplified network diagram
Device Mean RSSIAP1 24.4AP2 47.8AP3 45.5AP4 40.6
(b) Client fingerprint
Fig. 1. An example network and fingerprint corresponding to signal A
these self-calibrating systems leverage information regardingthe location of WiFi access points to conduct a trainingphase in which they collect system fingerprints, or sets ofRSSI measurements taken of (and by) the stationary APs inthe network. After the self-calibration stage, the fingerprintdatabase is combined with pre-configured information aboutthe physical location of the APs to compute a model of therelationship between RSSI and physical location. This modelis then used to answer on-line queries about client location.
A. Overview
It is convenient to think of self-calibrating algorithms asfitting into three categories based on the amount of data usedto perform calibration. To assist the reader in understandingeach algorithm we give examples of how they work in a simplewireless network. Our explanatory network, diagrammed inFigure 1(a), has four APs and one client. Each AP knowsthe distance between it and every other AP, for example thedistance between AP3 and AP4 is ten meters and labeled D.The 802.11 signal emitted by the client when sending a packetis labeled A. The RSSI values for one transmission as seen bythe four APs, i.e. a client fingerprint, are documented in thetable in Figure 1(b). Conversely, the 802.11 signal generatedwhen an AP sends a packet into the network is labeled C.When an AP sends a packet it is heard by everyone withinrange, though all but the intended destination will usuallyignore it when the network is operational. In the systemtraining phase, however, APs do not ignore packets sent fromother APs, labeled B, but instead record their RSSI values.Table I shows an example of training data that might observedby AP4 in this network.
Baseline: In order to provide some context in which tointerpret the performance of the localization system classes, weconsider a simple, calibration-free baseline that leverages onlya single RSSI from a client transmission, labeled A in Figure1(a). This nearest-neighbor approach requires no calibration:to locate a client it simply determines which AP heard theclient’s transmission with the strongest RSSI and declares theclient to be co-located with that AP. For example if the systemobtained the fingerprint in Figure 1(b) it would estimate theclient to be at the same location as AP2.
Class I: Our baseline algorithm, while simple, has twomajor limitations. First, it can choose only locations occupiedby an AP. Second, it ignores all information from the (n− 1)
F(x) =! ! e("!x )
F(x) = 78! e("0.09!x )
F(47.8) =1.3
TiX
• Phase II: Loca2on Es2ma2on – Triangular Interpola2on and eXtrapola2on – Like trilatera2on but a bit smarter
TiX
• Phase II: Loca2on Es2ma2on – Triangular Interpola2on and eXtrapola2on – Like trilatera2on but a bit smarter
1.3 Meters
TiX
• Phase II: Loca2on Es2ma2on – Triangular Interpola2on and eXtrapola2on – Like trilatera2on but a bit smarter
1.3 Meters
TiX
• Phase II: Loca2on Es2ma2on – Triangular Interpola2on and eXtrapola2on – Like trilatera2on but a bit smarter
1.3 Meters
TiX
• Phase II: Loca2on Es2ma2on – Triangular Interpola2on and eXtrapola2on – Like trilatera2on but a bit smarter
1.3 Meters
TiX
• Phase II: Loca2on Es2ma2on – Triangular Interpola2on and eXtrapola2on – Like trilatera2on but a bit smarter
1.3 Meters
TiX
• Phase II: Loca2on Es2ma2on – Triangular Interpola2on and eXtrapola2on – Like trilatera2on but a bit smarter
1.3 Meters
TiX
• Phase II: Loca2on Es2ma2on – Triangular Interpola2on and eXtrapola2on – Like trilatera2on but a bit smarter
1.3 Meters
TiX
TiX
TiX
TiX
TiX
ZCFG (Lim et al)
• Phase 1: Distance Es2ma2on – ZCFG has APs share calibra2on data – AP 1’s distance es2ma2on is linear combina2on of all AP’s belief about distance between device and AP1
• Phase II: Mul2latera2on – Like trilatera2on but w/ more than 3 distances
Outline
• Mo2va2on • Algorithms
• Experimental set-‐up / results
• Causes of Error • Conclusion
Test Environment
• Test Environment – 30k Sq. Ft. per floor – 10 APs per Floor
• Channels 1,6, and 11
• Training the Algorithms – APs monitor bcast traffic for 10 seconds
Evalua2on • Experimental Evalua2on
– Located a laptop on a 1.5M ladder – Average 2 seconds worth of RSSI data per loca2on agempt
– 13 different loca2ons • Mul2ple localiza2ons in each loca2on
Method Accuracy (M)
Min Median Published Median
Nearest AP n/a
Probabilis2c (de Moraes & Nunes) 1.84
TiX (Gwon & Jain) 5
ZCFG (Lim et al.) 2.5
Evalua2on • Experimental Evalua2on
– Located a laptop on a 1.5M ladder – Average 2 seconds worth of RSSI data per loca2on agempt
– 13 different loca2ons • Mul2ple localiza2ons in each loca2on
Method Accuracy (M)
Min Median Published Median
Nearest AP n/a
Probabilis2c (de Moraes & Nunes) 1.84
TiX (Gwon & Jain) 5
ZCFG (Lim et al.) 2.5
Evalua2on • Experimental Evalua2on
– Located a laptop on a 1.5M ladder – Average 2 seconds worth of RSSI data per loca2on agempt
– 13 different loca2ons • Mul2ple localiza2ons in each loca2on
Method Accuracy (M)
Min Median Published Median
Nearest AP 0.9 n/a
Probabilis2c (de Moraes & Nunes) 0.7 1.84
TiX (Gwon & Jain) 2.9 5
ZCFG (Lim et al.) 2.7 2.5
Method Accuracy (M)
Min Median Published Median
Nearest AP n/a
Probabilis2c (de Moraes & Nunes) 1.84
TiX (Gwon & Jain) 5
ZCFG (Lim et al.) 2.5
Evalua2on • Experimental Evalua2on
– Located a laptop on a 1.5M ladder – Average 2 seconds worth of RSSI data per loca2on agempt
– 13 different loca2ons • Mul2ple localiza2ons in each loca2on
Method Accuracy (M)
Min Median Published Median
Nearest AP 0.9 n/a
Probabilis2c (de Moraes & Nunes) 0.7 1.84
TiX (Gwon & Jain) 2.9 5
ZCFG (Lim et al.) 2.7 2.5
Method Accuracy (M)
Min Median Published Median
Nearest AP 0.9 6.2 n/a
Probabilis2c (de Moraes & Nunes) 0.7 5.8 1.84
TiX (Gwon & Jain) 2.9 7.8 5
ZCFG (Lim et al.) 2.7 12.4 2.5
Outline
• Mo2va2on • Algorithms
• Experimental set-‐up / results
• Causes of Error – Height Disparity / Signal Reflec2on – Extrapola2on Failures – Interpola2on Failures
• Conclusion
Height Magers
• Access Points are mounted near the Ceiling – Tests took place at ‘chest’ height
• Is this height discrepancy significant?
Height Magers
• Access Points are mounted near the Ceiling – Tests took place at ‘chest’ height
• Is this height discrepancy significant? Method Floor 1.5 M Ceiling Difference
Nearest AP 6.7 6.7 4.8 1.9
Probabilis2c (de Moraes & Nunes) 6.6 3.37 2.65 0.72
TiX (Gwon & Jain) 16.8 11.6 11.2 0.4
ZCFG (Lim et al.) 9.5 14.2 12.9 1.3
Signal Reflec2on
• Can we correct for the induced error? – Simply add n RSSI to each measurement?
• RSSI vs. 2me – Ceiling; Ladder; Floor Height
Signal Reflec2on
• Can we correct for the induced error? – Simply add n RSSI to each measurement?
• RSSI vs. 2me – Ceiling; Ladder; Floor Height
Signal Reflec2on
• Can we correct for the induced error? – Simply add n RSSI to each measurement?
• RSSI vs. 2me – Ceiling; Ladder; Floor Height
Signal Reflec2on
• Can we correct for the induced error? – Simply add n RSSI to each measurement?
• RSSI vs. 2me – Ceiling; Ladder; Floor Height
Signal Reflec2on
• Can we correct for the induced error? – Simply add n RSSI to each measurement?
• RSSI vs. 2me – Ceiling; Ladder; Floor Height
Signal Reflec2on
• Can we correct for the induced error? – Simply add n RSSI to each measurement?
• RSSI vs. 2me – Ceiling; Ladder; Floor Height
Signal Reflec2on
• Can we correct for the induced error? – Simply add n RSSI to each measurement?
• RSSI vs. 2me – Ceiling; Ladder; Floor Height
Signal Reflec2on
• Can we correct for the induced error? – Simply add n RSSI to each measurement?
• RSSI vs. 2me – Ceiling; Ladder; Floor Height
• Different loca2ons different results
Beyond Reflec2on
• Signal reflec2on causes some degrada2on – 1-‐2 Meters
• Experiments found more than 2 meters of accuracy difference from published values
• What other issues could be responsible?
Calibra2on Data
• Does our calibra2on data accurately capture decay – Order of magnitude less AP/sq. r. than previous works
Study Square Feet # of Aps Aps/Sq. Ft.
Prob (de Moraes & Nunes) 1,722 4 0.0023
TiX (Gwon & Jain) 6,717 4 0.0006
ZCFG (Lim et al.) 598 6 0.01
Cisco Recommends 3,000-‐5,000 1 0.0002-‐0.0003
UCSD CSE Building 30,000 10 0.0003
Signal Degrada2on due to Walls
AP 3
AP 1
AP 2
Signal Degrada2on due to Walls
AP 3
AP 1
AP 2
Signal Degrada2on due to Walls
AP 3
AP 1
AP 2
Signal Degrada2on due to Walls
AP 3
AP 1
AP 2
Signal Degrada2on due to Walls
• Wall placement causes signal degrada2on
AP 3
AP 1
AP 2
Degraded Training Data
• APs have 4-‐6 signal-‐distance pairs for training – Distance ranges from 5-‐35 Meters – RSSI ranges from 45-‐10
– Expect to fit exponen2al decay
Degraded Training Data
• APs have 4-‐6 signal-‐distance pairs for training – Distance ranges from 5-‐35 Meters – RSSI ranges from 45-‐10
– Expect to fit exponen2al decay
Degraded Training Data
• APs have 4-‐6 signal-‐distance pairs for training – Distance ranges from 5-‐35 Meters – RSSI ranges from 45-‐10
– Expect to fit exponen2al decay
Effect of mul2ple walls
When Walls Get in the Way
• Common for AP to have wall between it and all other training APs
• Yet, device to locate has line of sight AP 3
AP 1
AP 2
AP 4
When Walls Get in the Way
When Walls Get in the Way
When Walls Get in the Way
Device was here
When Walls Get in the Way
• Probabilis2c method doesn’t handle this well
Device was here
When Walls Get in the Way
• Probabilis2c method doesn’t handle this well – We ins2tuted a filtering scheme to help
Device was here
When Walls Get in the Way
• Probabilis2c method doesn’t handle this well – We ins2tuted a filtering scheme to help
• TiX handles this scenario much beger
Device was here
When Walls Get in the Way
• Probabilis2c method doesn’t handle this well – We ins2tuted a filtering scheme to help
• TiX handles this scenario much beger
When Walls Get in the Way
Almost no error
• Probabilis2c method doesn’t handle this well – We ins2tuted a filtering scheme to help
• TiX handles this scenario much beger
When Walls Get in the Way
• Probabilis2c method doesn’t handle this well – We ins2tuted a filtering scheme to help
• TiX handles this scenario much beger
When Walls Get in the Way
• Probabilis2c method doesn’t handle this well – We ins2tuted a filtering scheme to help
• TiX handles this scenario much beger
When Walls Get in the Way
• Probabilis2c method doesn’t handle this well – We ins2tuted a filtering scheme to help
• TiX handles this scenario much beger
But not always
Extrapola2on Failures
• Called extrapola2on failures because they occur beyond training data
• Some2mes filtering is effec2ve – Median improved 2.1 M for Probabilis2c method
• Made two agempts to filter ZCFG – Increased median error by 0.5 Meters
• Possible to overcome with addi2onal APs – Excessive AP density impacts WiFi performance
Good Training Data Isn’t Enough
• Not all failures can easily be overcome
Good Training Data Isn’t Enough
• Not all failures can easily be overcome
Similar RSSIs at different distances
Good Training Data Isn’t Enough
• Not all failures can easily be overcome
Similar RSSIs at different distances
1 wall between APs
Decay Isn’t Always Exponen2al
• 10 RSSI varia2on in under 3 Meters due to walls
Decay Isn’t Always Exponen2al
• 10 RSSI varia2on in under 3 Meters due to walls
Decay Isn’t Always Exponen2al
• 10 RSSI varia2on in under 3 Meters due to walls
Decay Isn’t Always Exponen2al
• 10 RSSI varia2on in under 3 Meters due to walls
Line of sight to AP
Decay Isn’t Always Exponen2al
• 10 RSSI varia2on in under 3 Meters due to walls
Line of sight to AP
Mul2ple walls to AP
Interpola2on Failures
• Called interpola2on failures because they occur in-‐between training data
• This isn’t fixable with extra training data – Walls add unpredictable decay
• Even the sophis2cated algorithms can’t overcome these issues – In fact most do worse
Summary
• Indoor Localiza2on is in demand • Achieving less than 5m accuracy w/o site survey is s2ll an open problem – Improved filtering mechanisms
– Novel models that beger isolate signal from noise in training data