Microsoft Research
Faculty Summit 2007
Aman Kansal
Researcher
Networked Embedded Computing, MSR
APPLICATION
Upload
Pictures,
Video,
Audio
Stitched view
(Data centric
coverage model)
SenseWeb
PARK with people …and phones
GROUP
MEMBER
1. Is the court wet?
SMS: Click picture
of court.
Group Points: 400
2. What play structures
are there?
3. Which bird sounds
reported?
Community Fitness
• Runners: Where are sidewalks broken? Construction finished on 24th St?
Recreation
• Mountain Bikers: Average biker heart rate at Adams Pass on trail 320?
• Surfer: What do the waves look like now?
• Hikers: Did the storm block the trail?
Public initiated instant news coverage: ground truths
Shopping
• Which displays changed? What’s attracting most attention?
Urban Moods
• Where are people hanging out tonight?
Real time Virtual Earth street side imagery
Pollution updates to Scorecard.org
Business Intelligence
• What did customer add to our design at the last meeting?
2.14 billion phones and growing
Mobility
reach where static sensor cannot
increased spatial coverage
Phone exists for voice/data apps:
Piggybacking sensing is cost effective
Human assistance
Can sometimes help detect or aim at
interesting phenomenon
SenseWeb
Server
Client on phone
Indexes images by location
and time (SQL Server
database)
Web service API for
phones and apps.
Supports several sensor
types
Allows users to take
pictures
Automatically uploads data
to server
Location stamps using
inbuilt/Bluetooth GPS
Example App: Portal
Displays sensor data
by location and
sensor type
Publicly accessible at
http://atom.research.
microsoft.com/sensor
map
Web service API’s
allow building
other apps.
Information value
Which data to collect and share: battery and
bandwidth constraints
Coverage management
Which phone sensed where app needs
coverage
Sensor tasking for application demands
Incentive mechanisms
Data verifiability, user privacy
Entropy of a single image:
H(X) = -S(p.log(p))
[p: image histogram]
Value among multiple images
Consider common spatial coverage
H(X|Y) = -E[log2p(X|Y)]
H(X|Y1,…,Ym) = H(X|Z) (Z: common spatial coverage)
3
DataDataSize
Size (MB)(MB)
Commonality:
found using key
feature based
algorithm
Value based selection
2.5
Buildings
2
1.5
1
Kitchen
0.5
0
10
Details: ACM Sensys WSW 2006
20
30
40
50
60
Relevance Value Cut-off
70
80
Relevance Value Cutoff (%)
90
Which sensors does app access
Who sensed in required region during required
time window?
Mobile Sensor Swarm
Which sensors does app
access
Who sensed in required
region during required
time window?
Application n
Application 1
Data Centric
Abstraction
Solution: location
Samples are geostamped
Apps do not track device
Trajectory
Connectivity
Sharing preferences
Device ID anonymized
Mobile Sensor Swarm
Several location technologies
GPS: does not work everywhere
Cell tower: coarse
Wifi: coarse
Human entered tags:
approximate, high manual effort
Mij
j
i
Leverage camera data to
enhance location
Refine location granularity
Room within building, aisle within
store
Associate data when location not
available
Verify location
Details: ACM NOSSDAV 2007
Algorithm
Images within vicinity
organized as a graph
Edge weight by match
Relation R(i,j) by
highest weight
Refined location zone:
Transitive closure of R
Minimize sensing task overhead on phones
Sense to be most accurate on most used regions
Good model: determine where sensing needed
Learn most used: where apps need data
Task phones: battery, bandwidth, privacy, intrusion
costs
Phenomenon
Demand
Details: Andreas Krause, Intern project report
Sensing cost
Set V of possible observations
For each subset A of V, define utility
U(A) = Σi E[Di (Var(Si) – Var(Si | A)) ]
Theorem: U(A) is submodular
Theorem [Nemhauser et al]:
For submodular U:
U(greedy solution) >
(1-1/e) U(optimal)
Demand-weighted variance
Expectation over demand Di and observations A
0.03
0.025
0.02
Random selection
Optimized for
variance reduction
Optimized for
demand-weighted variance
0.015
0.01
0.005
0
0
20
40
60
80
Number of observations (out of 534)
100
Mobile phones enable many sensing apps
Architecture to use a highly volatile swarm
of mobile devices as a sensor network
Information value based data selection
Location based data centric abstraction
Coverage management and data addressing
Avoids burdening applications with managing device
motion, connectivity, sharing
Efficient sensor tasking
Contact: kansal@microsoft.com
© 2007 Microsoft Corporation. All rights reserved. Microsoft, Windows, Windows Vista and other product names are or may be registered trademarks and/or trademarks in the U.S. and/or other countries.
The information herein is for informational purposes only and represents the current view of Microsoft Corporation as of the date of this presentation. Because Microsoft must respond to changing market
conditions, it should not be interpreted to be a commitment on the part of Microsoft, and Microsoft cannot guarantee the accuracy of any information provided after the date of this presentation.
MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.