This document discusses device-free indoor localization using machine learning techniques at 28 GHz. The methodology uses ray tracing to generate fingerprint data and selects features from received power values. A random forest algorithm is used for classification and regression training on global and combined classifiers. Results show that combining independent classifiers from one or two transmitters reduces positioning error by at least 16-19% compared to global classification, and by at least 36-37% when combining two transmitters with classification-regression. The size and number of partition classes impacts error, and additional small improvements are achieved through classification-regression combination.
Machine Learning Based Device Free Indoor Localization Using 28 GHz Signals
1. Device Free Indoor Localization in the 28 GHz
band based on machine learning
The 14th International Conference on Ambient Systems, Networks and Technologies
March 15-17, 2023, Leuven, Belgium
www.elsevier.com/locate/procedia
Juan Avilés , Verónica Ojeda , Víctor Asanza
javiles@espol.edu.ec, vejaojed@espol.edu.ec, victor.asanza@sdas-group.com
3. Topics
❑ Introduction
❑ Methodology
❑ Results
❑ Conclusions
❑ Related Works
Device Free Indoor Localization in the 28 GHz band based on machine learning
3
www.elsevier.com/locate/procedia
4. Introducción
Introduction
3
• Device free localization (DFL) indoor position of a person.
• As opposed to active localization, DFL does not require the person to carry any active
electronic device.
• Variations in received power: ‘no presence’ versus ‘presence’ positioning.
• Potential applications: people monitoring, disaster situations, intelligent buildings, multisubject
counting and positioning.
Rx
Tx
Tx
Tx
Tx
Tx
Rx
Rx Rx
Rx
Rx
Rx
Posicionamiento activo Posicionamiento pasivo
www.elsevier.com/locate/procedia
5. Motivation
The current study explores the application of global-based classification,
combined classification, and the combination of classification and regression
techniques to estimate the position of a person indoors .
• Few studies on the combination of classification and regression techniques for
positioning.
• Positioning at 28 GHz? Frequency band candidate for 5G and 6G services
4
www.elsevier.com/locate/procedia
6. Problema?
16 clases
4 clases (Y)
4 clases (X)
Al aumentar la cantidad de
clases, disminuye la exactitud
Menor cantidad
de clases (aumenta la
exactitud del clasificador
Disminuye el error
de Posicion?
Centro
Geometrico
Asociado a
cada clase
(xi,yi)
www.elsevier.com/locate/procedia
8. Introducción
Methodology: Ray-tracing + Matlab: automate the use of the tool for propagation calculations and data recording.
Fingerprint data
generation
(Ray Tracing)
Data processing &
Feature selection Training algorithm Testing Error
6
• Txs equipped with 9-ULAs.
• Rxs operate with omnidirectional antennas.
• 380 person’s locations
• Angles per Tx: 21
Rx3
Rx4
3m
www.elsevier.com/locate/procedia
9. Introducción
Methodology: Ray-tracing + Matlab: automate the use of the tool for propagation calculations and data recording.
Tx antenna orientation angles:
-75o, -67.5o, -60o, -52.5o, -45o, -37.5o, -
30o, -22.5o, -15o, -7.5o, 0o, 7.5o, 15o,
22.5o, 30o, 37.5o, 45o, 52.5o, 60o, 67.5o,
75o relative to the perpendicular direction
of the array.
6
Fingerprint data
generation
(Ray Tracing)
Data processing &
Feature selection Training algorithm Testing Error
Interfase Matlab-Herramienta de calculo de
propagacion (15,960 simulaciones)
Data set global: 63840 valores de potencia
www.elsevier.com/locate/procedia
10. Indoor positioning is a challenging task!: 1) Small power difference. 2) Not all Rxs contribute for positioning
7
www.elsevier.com/locate/procedia
11. For one Tx, 4 Rxs, 290H and 21 angles (𝜃1, 𝜃2, … , 𝜃21 ), 84 features are selected as follows:
∆𝑃𝑇𝑥1𝜃𝑖𝑅𝑥𝑗
290𝑥1
= ∆𝑃𝑇𝑥1𝜃𝑖𝑅𝑥𝑗
1
∆𝑃𝑇𝑥1𝜃𝑖𝑅𝑥𝑗
2
… . ∆𝑃𝑇𝑥1𝜃𝑖𝑅𝑥𝑗
290
𝑇
∆𝑃𝑇𝑥1𝜃1 290𝑥4
= ∆𝑃𝑇𝑥1𝜃1𝑅𝑥1
∆𝑃𝑇𝑥1𝜃1𝑅𝑥2
… ∆𝑃𝑇𝑥1𝜃1𝑅𝑥4
∆𝑃𝑇𝑥1𝜃2 290𝑥4
= ∆𝑃𝑇𝑥1𝜃2𝑅𝑥1
∆𝑃𝑇𝑥1𝜃2𝑅𝑥2
… ∆𝑃𝑇𝑥1𝜃2𝑅𝑥4
…………………………………………
∆𝑃𝑇𝑥1𝜃21 290𝑥4
= ∆𝑃𝑇𝑥1𝜃21𝑅𝑥1
∆𝑃𝑇𝑥1𝜃21𝑅𝑥2
… ∆𝑃𝑇𝑥1𝜃21𝑅𝑥4
∆𝑃𝑇𝑥1 290𝑥84 = ∆𝑃𝑇𝑥1𝜃1
∆𝑃𝑇𝑥1𝜃2
. . . ∆𝑃𝑇𝑥1𝜃21
For 2 Txs, 4 Rxs, 290H, 21 angles (𝜃1, 𝜃2, … , 𝜃21 ), 168 features are chosen as follows:
∆𝑃𝑇𝑥1𝑇𝑥2 290𝑥168
= ∆𝑃𝑇𝑥1 290𝑥84 ∆𝑃𝑇𝑥2 290𝑥84
Fingerprint data
generation
(Ray Tracing)
Data processing &
Feature selection Training algorithm Testing Error
www.elsevier.com/locate/procedia
12. Figure 3. Data set: a) classification training, b) regression training, c) testing.
Methodology
Fingerprint data
generation
(Ray Tracing)
Data processing &
Feature selection Training algorithm Testing Error
Example 1 f1,1 f1,2 f1,3 ….. f1,w ClassID1
Example 2 f2,1 f2,2 f2,3 ….. f2,w ClassID2
a) ................ …..... …..... …..... …..... …..... ….....
Example 290 f290,1 f290,2 f290,3 ….. f290,w ClassIDn
fv1,1 fv1,2 fv1,3 ….. fv1,w X1 or Y1
fv2,1 fv2,2 fv2,3 ….. fv2,w X2 or Y2
b) …..... …..... …..... …..... …..... ….....
fvk,1 fvk,2 fvk,3 ….. fvk,w Xk or Yk
fu1,1 fu1,2 fu1,3 ….. fu1,v
fu2,1 fu2,2 fu2,3 ….. fu2,v
c) …..... …..... …..... …..... ….....
fu90,1 fu90,2 fu90,3 ….. fu90,v
k examples
90 examples
www.elsevier.com/locate/procedia
15. 75% 25%
Validation
72 Groups of
Recvd. power
values
training
218 Groups of
Recvd. power
values
Random Forest
Algorithm that generates several decision trees using random bootstrapped
training feature sets. It averages the prediction results to reduce the
variance of the statistical learning method.
(Examples)
Model Model Model
X
X X
Majority
Vote
Or
prediction
train
train train
n’ n’ n’
11
Methodology
Features
Fingerprint data
generation
(Ray Tracing)
Data processing &
Feature selection Training algorithm Testing Error
Set de datos
Features
Examples
www.elsevier.com/locate/procedia
16. ( )
,
i i i
p X Y
=
1
1 i
n
i ij
i j
X x
n =
=
1
1 i
n
i ij
i j
Y y
n =
=
( )
,
r r r
p X Y
=
2 2
i Xi Yi
PE PE PE
= +
r
Xi i
X
PE X
= − r
Yi i
Y
PE Y
= −
Technique a: estimated position depends on the classification result (any of 16, 32 or 64 classes) associated to
the center position of the predicted block.
Technique b: estimated position depends on the classification result (any of 16, 32 or 64 classes) associated to
the center position of the predicted block.
( )
,
r r r
p X Y
= r
Xi i
X
PE X
= −
1
1
xi
i
n
i ij
x j
X x
n =
= 1
1
yk
k
n
k kj
y j
Y y
n =
=
r
Yk k
Y
PE Y
= −
2 2
ik Xi Yk
PE PE PE
= +
k
Y
r
p
i
X
ik
PE
i
n
i
p r
p
i
PE
Technique c: each (independent) classifier predicts a particular strip (class) in X and Y which in turn identifies the
regressor to be applied. The is calculated as in technique b where is obtained from the regressors’ outputs.
ik
PE ( )
,
i k
X Y
www.elsevier.com/locate/procedia
23. Results: CDF of median values of position error
Application of global-based classifiers and combination of independent classifiers for Tx1 or Tx1 and Tx2.
• Combination of classifiers produces a decrease in the median value of the position error for all partition cases.
As # classes (blocks or strips) , examples per class . It affects classification accuracy. Position error
On the other hand, as # classes (blocks or strips) , dimensions of the blocks (strips) . Position error
Global
Combined
Global
Combined
www.elsevier.com/locate/procedia
27. Results
• Except for combined classification 1Tx and 4X cases (Tx1Class4X4Y, Tx1Class4X8Y), the integration of
regressors produces a relatively small an accuracy improvement. Such improvement is small in all Tx1Tx2 cases.
www.elsevier.com/locate/procedia
28. Results
For Tx1, Table 1 shows:
Reduction of at least 16.2% between global and combined Classification
Reduction of at least 26.9% between global and combined Classification-Regression
www.elsevier.com/locate/procedia
31. • Simulations suggest that the block (strip) size and number of classes (number of
examples) in which the study area is partitioned play a role in the position error
determination.
• Results also indicate that the statistical position error achieved with a global-based
classifier can be reduced by adding data from a second Tx and through the application of
independent classifiers using data corresponding to position coordinates
• Statistical localization error reduction of at least 16% over a global-based classification
technique can be obtained through the combination of two independent classifiers using
one transmitter and a reduction of at least 19% for 2 transmitters.
• An additional (small) improvement is achieved by combining each independent classifier
with a regression algorithm.
Conclusions:
18
www.elsevier.com/locate/procedia
32. Introducción
Related work
Environment
Indoor positioning.
Algorithm/ Method-Accuracy
Reference
Indoor area (5x8x3)m, 8 Txs and 8
Rxs, omni antenas, 433.1 and 909.1
MHz power values measurements.
Probabilistic classification based on
discriminant análisis. Accuracy of 97,2 %.
Xu C. et al. [2]
ULA, signal subspace vector, 2.4 GHz,
Indoor areas (7mx7m and 10mx11m)
(measurements)
Support Vector Machine
(2.06 m RMSE, 4.14 m RMSE)
Hong et al. [5]
Brick house of 84 square meters. omni
antenas, 2.4 GHz
DFL error rate (position accuracy degrades)
doubles, on average, for every six random
changes. Random forest classifier provided the
lowest error rate compared to KNN, SVM and LDA.
Mager et al. [7]
14 data sets (indoor Wi-Fi received
signal strength). Multibuilding,
multifloor. Active positioning.
Cascade of three models of building classification,
floor classification and 2D-localization regression.
Klus et al. [13]
19
www.elsevier.com/locate/procedia