Building Tomography:

1
Building Tomography Automatic Floor Plan
Generation for Mobile and Pervasive Applications
2
INTRODUCTION
Building tomography is to explore indoor
architectural layouts, usually represented as a
floor plan, from outside.
Floor plan illustrates a number of key spatial
elements like rooms, corridors, walls, and other
physical features at one level of a building.
3
INTRODUCTION
Floor plan plays an essential role in many indoor
pervasive and mobile applications
human localization
navigation
smart space
asset management
provides basic geographical information for
analyzing human mobility and social behavior
4
INTRODUCTION
the enlargement of applicable areas
strangled
pretty limited information of building interiors
leverage user mobility information for reduced
dependency of indoor locating systems on site
survey
floor plan, as the ground truth, is still
necessary.
5
INTRODUCTION
smartphones, with rich built-in sensors

human-centric sensing
compasses
gyroscopes
accelerometers
crowdsourcing
6
INTRODUCTION
three major challenges
erroneous sensor readings would corrupt user
moving trajectory
without ground truth information, how can the
data from multiple users be geographically
aggregated?
space regionalization and function recognition
based on multi-sensor readings are insufficiently
investigated, lacking of well-proven effective
methods
7
INTRODUCTION
solution
based on nowadays wireless and sensor technology
inexpensive and pervasive
no building knowledge is required and no extra
hardware needs
all sensor readings are collected by
off-the-shelf smartphones, and without user
interventions or even attention
8
RELATED WORKS
The idea of building tomography comes from
human-centric sensing and crowdsourcing.
Mobility-based Indoor Localization
Based on ambience features including sound,
light, color, Wi-Fi, etc. All these approaches
require site survey over areas of interests to
build fingerprint database.
Fingerprinting-based techniques
Mobile Crowdsourcing
9
RELATED WORKS
The idea of building tomography comes from
human-centric sensing and crowdsourcing.
Mobility-based Indoor Localization
In these methods, locations are calculated rather
than searched from known reference data.
Fingerprinting-based techniques
Model-based techniques
Mobile Crowdsourcing
10
RELATED WORKS
The idea of building tomography comes from
human-centric sensing and crowdsourcing.
Mobility-based Indoor Localization
Researchers have begun to leverage mobility
information assisting localization. Users motion
trajectories can be revealed through
accelerometers, gyroscopes, or compasses
Fingerprinting-based techniques
Model-based techniques
Mobility assisted techniques
Mobile Crowdsourcing
11
RELATED WORKS
The idea of building tomography comes from
human-centric sensing and crowdsourcing.
Mobility-based Indoor Localization
SLAM often relies on accurate control of a robot
or builds probabilistic models of environments.
Human mobility and activity, measured by
smartphones, are not the input of SLAM.
Fingerprinting-based techniques
Model-based techniques
Mobility assisted techniques
Simultaneous localization and mapping
Mobile Crowdsourcing
12
RELATED WORKS
Crowdsourcing is a distributed problem-solving
model.
Equipped with different kinds of built-in sensors
for various functions, mobile phones can be seen
as an information interface between users and
environments, as well as a client of
crowdsourcing.
In our solution, crowdsourcing is implicit and
unconscious. Except running a program for data
collection, users who contribute to building
tomography pay no attention and energy to this
process.
13
SYSTEM OVERVIEW
14
SYSTEM OVERVIEW
15
TRACE COLLECTION
User Data Collection
Three types of sensors accelerometer, gyroscope
and magnetometer
A trace reported from a client to the central
repository contains 1) timestamped sensor
readings 2) timestamped WiFi signatures 3)
last available GPS positioning result
16
TRACE COLLECTION
Dead-reckoning
Based on step counting
We observe that each step begins with a positive
acceleration, and reaches a summit soon. Then the
acceleration drops down and becomes negative, and
reaches a negative summit afterwards. Finally the
step ends when the acceleration comes back to
around 0. Based on this observation, we implement
a finite state machine to detect user steps.
17
TRACE COLLECTION
There are two main sources of errors.
The accelerometer itself generates noises which
are presented in outputs.
Residual component of gravity.
Track a users orientation change using gyroscope.
Use cellphones compass reading, which is derived
by the combination of acceleration and magnetic
field strength.
18
TRACE REALIZATION
outdoor part
indoor part
reference points
19
TRACE REALIZATION
We leverage WiFi fingerprint to match two
reference points. During the data collection
phase, a client keeps scanning WiFi APs and
records their signal strength.
20
TRACE REALIZATION
Drift Fixing
21
TRACE REALIZATION
22
MAP GENERATION
Space Regionalization
Once the trace map is constructed by aggregating
user traces together, the next step is to
recognize different areas and identify their
functionality, such as corridors, rooms,
stairs, or elevators.
K-means clustering
23
MAP GENERATION
24
MAP GENERATION
Space Regionalization
Once the trace map is constructed by aggregating
user traces together, the next step is to
recognize different areas and identify their
functionality, such as corridors, rooms,
stairs, or elevators.
K-means clustering
25
MAP GENERATION
Functionality Recognition
a Bayesian Model based approach for functionality
recognition.
For office buildings, four types of area
functionality are considered in our system room,
corridor, elevator, and stairs.
26
MAP GENERATION
All the features from a specific area form an
observation O (m, a, v, l, t, s), which
represents the signature of this area.
27
MAP GENERATION
28
MAP GENERATION
29
MAP GENERATION
30
APPLICATIONS
Indoor localization is a direct application of
the building tomography system.
The building tomography system also assists the
research of human mobility and social behavior,
in which data collection is hard and the
collected data are less well labeled. Our system
provides not only sufficient mobility
information, but also corresponding meaningful
comments. In our preliminary experiment, the
traces are collected from one specific user for 3
hours in a morning. We observe that he stayed at
his office 94.36 of time, while 5.64 in the
corridor. His total walking distance is 548.64m,
and the average speeds are 0.85m/s and 1.30m/s in
the office and corridor, respectively.
31
EXPERIMENT
32
EXPERIMENT
33
EXPERIMENT
34
EXPERIMENT
35
EXPERIMENT
The total size of the building can be derived as
well. From the trace map, the length and width of
the building are estimated as 50m and 17m,
leading to a total size of 850m2, which is
slightly smaller than the real indoor size of
about 900m2.
36
CONCLUSION
We find that many mobile and pervasive
applications rely on indoor floor maps, but it is
difficult to obtain a large collection of floor
maps worldwide. In this study, we implement a
remote building tomography system, which
discovers indoor layouts through utilizing user
motion information detected by off-the-shelf
smartphones. Preliminary results prove the
feasibility of our attempt. A direction of our
future work is to adapt the techniques to various
buildings and environments. We envision this
point in order to increase the robustness of the
building tomography system.
37
Thanks!