9/10/2015
Cyber-physical Computing Group
Urban Sensing Applications
Tarek Abdelzaher
University of Illinois at Urbana Champaign
History
Initial Drivers: Smartphones and
Connected Vehicles



J. Burke, D. Estrin, M. Hansen, A. Parker, N. Ramanathan, S. Reddy, M. B.
Srivastava. “Participatory sensing,” In Proc. World Sensor Web Workshop,
ACM Sensys, Boulder, Colorado, October 2006.
(Cited by 848)
Andrew T. Campbell, Shane B. Eisenman, Nicholas D. Lane, Emiliano Miluzzo,
and Ronald A. Peterson, “People-centric Urban Sensing,” In Proc. 2nd annual
international workshop on Wireless Internet (WICON), August 2006.
(Cited by 375)
Tarek Abdelzaher, Yaw Anokwa, Péter Boda, Jeff Burke, Deborah Estrin,
Leonidas Guibas, Aman Kansal, Sam Madden, Jim Reich, “Mobiscopes for
Human Spaces,” IEEE Pervasive, Vol. 6, No. 2, pp. 20-29, April 2007.
(Cited by 248)
1
9/10/2015
2006
Smart phones
2006
Smart phones
2
9/10/2015
Vision


Exploit mobile devices in the possession of
individuals to perform acts of sensing
Two competing flavors


Participatory sensing (UCLA)
Opportunistic sensing (Dartmouth College)
Two Competing Flavors

Participatory sensing: “the custodian
consciously opts to meet an application request
out of personal or financial interest”.



Places much demand on the user
Offers control at the cost of usability
Opportunistic sensing: “custodians configure
their devices to let [sensing] applications run
(subject to privacy and resource usage
restrictions), but they might not be aware
which applications are active at any given time”
6
3
9/10/2015
Application Types
1. Geo-tagging (participatory)

Phone-based geo-tagging
of events of interest
(UCLA)




Crowds/pollution on beach
Invasive species (weeds)
Trucks in residential
neighborhoods
Drinking fountains
Reprinted from UCLA/CENS
Application Types:
2. Statistics/Mapping (opportunistic)

Example: BikeNet
8
4
9/10/2015
People-centric Sensing
MetroSense Project

Where people are the focal point of sensing
9
CenceMe (2007)



A Facebook app
Sensors compute user context (or “sensing presence”).
Context is shared with social circle (e.g., Facebook
friends) according to specified privacy policies.
10
5
9/10/2015
CenceMe


A Facebook app
Sensors compute user context





Activities (sitting, walking, or meeting friends),
Disposition (happy, sad, or okay)
Locations (at the gym, coffee shop, or at work)
and
Surroundings (noisy, hot, or bright).
Context is shared with social circle (e.g.,
Facebook friends) according to specified
privacy policies.
11
Sharing in Virtual Worlds

Integrate second life with sensors in the real
world (e.g., on a phone) for various “cyberphysical” games
12
6
9/10/2015
Challenges in Opportunistic
Sensing




Determine the device’s sampling context
(what is it allowed to share in that context)
Adapt to the devices’ changing resource
availability and sampling context
Sustain custodian privacy
Achieve sufficient sensing coverage in the
face of mobility
13
Shortest and fastest
Applications:
3. Data Modeling
Most fuelefficient
Green GPS
Subscribers
The fuel
efficient option
Saves 6% over shortest path
and 13% over fastest path
Fuel Data
+
Physical Models
+
OBDII-WiFi
Adaptor ($50)
GPS Phone
Server
7
9/10/2015
Regression Modeling
Fuel consumption of
some cars driven on
some roads
Predict fuel consumption of
any car on any road
8
9/10/2015
Starting with the Basics:
1. Geo-tagging (participatory)

Phone-based geo-tagging
of events of interest
(UCLA)




Crowds/pollution on beach
Invasive species (weeds)
Trucks in residential
neighborhoods
Drinking fountains
Reprinted from UCLA/CENS
Geotagging Challenge: All Tags
Not Created Equal
Example: Find Free Parking Lots on UIUC Campus
Y: Free Parking lot - no
charge after 5pm
N: Not free parking lot
Y
N
18
9
9/10/2015
Geotagging Challenge: All Tags
Not Created Equal
Example: Find Free Parking Lots on UIUC Campus
Results of Extended EM vs Baselines
Experiment setup:
106 parking lots of interests, 46 indeed free
30 participants, 901 marks collected
19
Cleaning Noisy Multi-label Data
Average
Apollo
Ground
Truth
10
9/10/2015
Surrogate Sensing
Example:
Traffic Regulator Mapping
Cell phones in vehicles were used as the sources (whose reliability is
unknown)
Vehicles reported encountered stop signs and traffic lights together with
their locations
All reports were fed to the Apollo Fact-finder to determine their probability
of correctness
Apollo predictions were compared against ground truth
21
Traffic Regulator Detection
From GPS Data
Source Reliability
Traffic Light Location
Estimation
Detection
Experiment setup:
34 drivers, 300 hours of driving in Urbana-Champaign
1,048,572 GPS readings, 4865 claims generated by phone
(3033 for stop signs, 1562 for traffic lights)
22
11
9/10/2015
Traffic Regulator Detection
From GPS Data
3 outliers out of 34
Source Reliability
Traffic Light Location
Estimation
Detection
Experiment setup:
34 drivers, 300 hours of driving in Urbana-Champaign
1,048,572 GPS readings, 4865 claims generated by phone
(3033 for stop signs, 1562 for traffic lights)
23
Traffic Regulator Detection
From GPS Data
Source Reliability
Estimation
Traffic Light Location
Detection
24
12
9/10/2015
Traffic Regulator Detection
From GPS Data
Source Reliability
Estimation
Traffic Light Location
Detection
25
Traffic Regulator Detection
(Enhanced)
26
13
9/10/2015
Traffic Regulator Detection
(Enhanced)
27
Conclusion

A very rich space is emerging where even
the simplest applications (geotagging)
lead to interesting research problems.
More on that in the rest of the semester…
14