image/svg+xml
Miquel Perello-Nieto, Raul Santos-Rodriguez, Peter Flach
Abstract
Real vs Synthetic
(rooms)
(rooms)
Real Walkaround
Synthetic Walkaround
Real houses
Synthetic houses
Leveraging synthetic data for indoor
localisation with missing sensors
Indoor Localisation
-70dB
-90dB
-34dB
Annotated
locations
Signal strenght
to all sensors
-Obstacles interfer in
the signal propagation
-35dB
air
-100dB
furniture
missing
wall
Missing Sensors
- In any real project with lots of sensors it is expected for some
of them to fail.
- It is necessary to design algorithms that are robust agains
unexpected sensor failures.
- Example of predictions in a real house in which some sensors failed during last month.
(The figure aesthetics and legends are different for non related reasons).
- Side by side slice
- An algorithm that
does not
expect sensor failures
only
predicts bedroom during last
month (all purple).
- An algorithm
expecting sensor
failures
indicates that the bedroom is
used mostly during the night periods, and
other rooms are used during the day.
Leveraging Synthetic Data
1. room locations
2. Perfect path
3. Add movement
1st floor
2nd floor
3rd floor
1st floor
2nd floor
3rd floor
- Its graph representation
- For each room we can perform random walkarounds
- In a matrix form
- We need large amounts of data to statistically validate if one method is more robust agains
sensor failure than another.
- This is a real house layout
- We can generate infinite house layouts with
different room sizes, locations,
sensors, signal quality and sensor failures.
- We can generate synthetic houses with realistic number of sensors and rooms.
Results
(work in progress)
Real
Synthetic
Synthetic (sensors broken)
- Example performance in one house
Real
Synthetic
Synthetic
(sensors broken)
Real
Synthetic
Synthetic
(sensors broken)
- Performance of one method in all real houses and
50 synthetic houses.
- Statistical comparison of different methods
in all real and 50 synthetic examples.
Technologies for
this presentation
inkscape.org
sozi.baierouge.fr
homestyler.com
- Free
- Open Source
- Vector
Graphics Editor
- Free
- Open Source
- Slides from a
Scalable Vector
Graphics image
- Create house
designs and
furniture
distribution in the
web browser
T
racking activities of daily living and daily routine provides clear indicators of
well-being. However, most research in Ambient Assisted Living is performed in
controlled environments. While in real-world scenarios sensors may fail, and
sensor signals may change due to external reasons (eg. redistribution of furniture).
We are investigating good Machine Learning algorithms that can predict indoor
locations even with sensor failures. In order to draw statistical conclusions
from different methods we generate synthetic data as close as possible to our
real deployments in which we can generate different types of failure. Finally,
the best methods are applied into the full set of real deployments
.
- It is possible to estimate the participants' location based on the "distance" (signal
strength) from the wearable to all the receivers.
- We use machine learning algorithms to learn the
correspondence between "distance" values and room locations.
- In order to get these pairs the technicians walk to every room
and annotate their location during the installation day.
1
All
abstract
indoor
indoor 1
indoor 2
missing
missing 1
missing 2
leveraging
leveraging 1
leveraging 2
real
real 1
real 2
results
results 1
results 2
results 3
technologies
technologies