lecture 4: people-centric sensing · 2017-04-26 · vehicular system owners can provide information...
TRANSCRIPT
Lecture 4: People-centric Sensing Cristian Borcea
Department of Computer Science
NJIT
2
Enable accurate real-time monitoring of the physical world
Consist of homogeneous, static, tiny sensors
Sink Internet, Satellite
Task Manager
End user
Sensor node
Water pollution in the river Fire in the forest Structural integrity of the bridge
How can we benefit from sensing in our
daily routine?
Leverage existing “infrastructure”: smart phones and vehicular systems No deployment cost
Excellent coverage: billions of devices distributed everywhere
3
Tornado
approaching!
Highway
Parking Area
Shopping
mall
Traffic jam!
Local fog
patches!
Free parking spot!
50%
discount!
People-centric: people are both consumers and providers of sensed data
On phones: camera, microphone, GPS, accelerometer, digital compass, light sensor, Bluetooth, [future: pollution sensors, medical sensors]
On cars: all car sensors can be made available to the vehicular system
Owners can provide information as well (i.e., human sensors)
4
Monitoring pollution in urban areas
Beneficiary: environment protection agencies
Monitoring vehicular traffic to detect/prevent traffic jams
Beneficiary: departments of transportation for planning, all of us for
improved traffic conditions
Finding the locations of potholes on the roads
Beneficiary: municipalities or departments of transportation
Determining the effect of lifestyle (meeting people, level
of exercise) on people’s wellbeing
Beneficiary: medical/social researchers, all of us
Documenting public events in real-time
Beneficiary: news outlets, all of us 5
On phones: free SDKs for development & app stores for deployment Could easily deploy sensing applications at scale
On cars: no standard vehicular system yet Care must be taken to ensure apps do not interfere with
car safety 6
Introduction
Smart phone sensing
Micro-Blog @ Duke
CenceMe @ Dartmouth
Activity recognition on smart phones
Vehicle sensing
Future challenges
7
Internet
Micro-Blog vision: virtual information telescope [Ref. 4]
8
Users encouraged to blog on mobile phones Video, audio, pictures, text, etc.
Micro-Blog phone client geo-tags blog
Uploads to server over WiFi/GPRS
Micro-Blog server positions blog on Google Maps
Internet users zoom into maps
Witness blog streams across the world
9
Virtual Telescope
Cellular, WiFi Visualization Service
Web Service
Phones People
Physical Space
Some queries participatory Is beach parking available?
Others are not Is there WiFi at the beach café?
Phones reply to query & reply posted on Google Map as new micro-blog
10
Content generated at a certain region is automatically
downloaded on user smart phone when user enters region
Example
User X creates micro-blog about restaurant food
“Floats” micro-blog at the restaurant
User Y arrives at restaurant
X’s micro-blog downloaded onto Y’s phone
Y can modify content and “re-float”
11
Continuous GPS (8 hours battery life) WiFi, GSM localization improves energy (16, 40 hours)
Degrades localization accuracy (40, 500m)
Solution: switch between them as function of app needs
WiFi
GSM GPS
Time (in minutes) 12
Phones need to continuously update their location Poses privacy risks
Pseudonyms insufficient
Proposed 3 blogging modes
Public, Social, Private
Users set privacy policy
In social mode, only those in social network view blogs
For querying
Privacy feasible through K-anonymity based solutions
13
14
Leverage sensors on the phones and social networks to provide context/activity inferences
15
16
Introduction
Smart phone sensing
Activity recognition on smart phones
Darwin phones @ Dartmouth
Vehicle sensing
Future challenges
17
Machine learning techniques applied on phone sensor data to
determine user activity
E.g., sitting, walking, running, in a car, etc.
Done on phone or server (on-line or off-line) function of amount of
resources required
Data exchange between nearby phones could enable group activity recognition
Examples of applications:
Is the user exercising enough? (sensors: GPS & accelerometer)
Has the user a good social life? (sensors: Bluetooth and microphone)
Deriving higher-level context information
▪ Is the user in an important meeting? (sensors: Bluetooth,
localization + social network) 18
19
Calibration, uncontrolled phone context, computational efficiency, battery power
Example of uncontrolled phone context: one activity, two different patterns
Cycling while phone in pants pocket (A) and backpack (B)
20
microphone
camera
GPS/WiFi/
cellular
air quality
pollution
social context
audio / pollution / RF
fingerprinting
image / video
manipulation
classification model evolution
classification model pooling
collaborative inference
Darwin
Applies distributed computing and collaborative inference to activity recognition
Why Darwin? Same classifier model for all users doesn’t work; even same classifier
for same user doesn’t work when changing context
Creating one classifier model per user doesn’t scale
21
4 phases
initial training (derive model seed)
classification model evolution
classification model pooling
collaborative inference
supervised
unsupervised
Darwin was tested on voice recognition 22
23
phone: feature extraction (low computation)
backend
backend: model training (high computation)
24
phone determines when to evolve
match?
NO
evolve (train new model using
backend as before)
training sampled
Speaker A’s model
Phone A Phone B
Phone C
Speaker B’s model
Speaker C’s model
Speaker C’s model
Speaker A’s model
Speaker B’s model
Speaker B’s model
Speaker A’s model
Speaker C’s model
Re-use already available classifiers instead of training a new model for each speaker Saves battery power and reduces delay
25
Phone A Phone B
Phone C
User A speaking!!!
A’s LI results:
Prob(A speaking) = 0.65
Prob(B speaking) = 0.25
Prob(C speaking) = 0.10
C’s LI results:
Prob(A speaking) = 0.30
Prob(B speaking) = 0.67
Prob(C speaking) = 0.03
B’s LI results:
Prob(A speaking) = 0.79
Prob(B speaking) = 0.11
Prob(C speaking) = 0.10
Local inference can be misleading
Use collaborative inference to reach common decision
26
Collaborative inference & classification model evolution boost performance The more phones the better 27
• Indoor, quiet scenario • 8 people around a table
Introduction
Smart phone sensing
Activity recognition on smart phones
Vehicle sensing
Pothole Patrol @ MIT
ParkNet @ Rutgers
Future challenges
28
Determine which roads need to be fixed Challenges
Differentiate potholes from other road anomalies (railroad crossings, expansion joints)
Cope with variations in detecting same pothole (speed, sensor orientation)
Vehicles have GPS and 3-axis accelerometer <time,location,speed,heading,3-axis acceleration>
Detection based on accelerometer data
Embedded Linux system + WiFi/Cellular installed in 7 taxis
29
Pothole Record
Clustering Algorithm
Cab 1
GPS
3 Axis Accelerometer
Location Interpolator
Pothole Detector
Cab 2
GPS
3 Axis Accelerometer
Location Interpolator
Pothole Detector
Central Server
30
Use classifier Hand-labeled data for training
Extend training set with loosely-labeled training data (type & frequencies, but not exact locations)
Clustering (at least K events must happen at same location with small error)
31
On labeled data:
32
On unlabeled data (with spot verification):
Provide a high-level view of parking availability to drivers on the road Single sensor vehicle senses multiple parking spots
33
34
Sensor vehicle instrumented with Computer & WiFi/cellular communication
Ultrasonic sensor: 20 readings per sec; range: 6.5m
GPS: 5 readings per sec
Webcam: capturing ground truth for evaluation
Obstacles: ‘dips’ in ultrasonic sensor readings Use classifier to differentiate parked cars from
other objects
35
36
Slotted (marked) spots
Unslotted spots Difficult to estimate if a car fits in the sensed spot
Introduction
Smart phone sensing
Activity recognition on smart phones
Vehicle sensing
Future challenges
37
Dependability How can “clients” trust data received from mobile sensors?
▪ How to validate these data? Should use reputation mechanism?
▪ How to determine ground truth?
Sensor calibration?
Privacy How to provide anonymity?
▪ How to balance the need for anonymity with the need for reputation?
Incentives Users motivated by self-interest for some apps (e.g., traffic jams)
Why should they provide data if there is no clear self-interest?
▪ Get paid? How to get paid while maintaining anonymity?
▪ Need for fair exchange protocol 38
Programmability Current systems are application-specific; How to quickly develop and
deploy new applications?
How to decompose global distributed sensing tasks into individual tasks?
▪ E.g., ensure spatial and temporal distribution of sensed data?
How can device owners control where, when, and how much resources are consumed for sensing?
Scheduling and real-time constraints What parameters should be considered in task scheduling?
▪ Payment, real-time constraints, spatio-temporal distribution of data providers?
▪ Should a provider be paid if the answer comes late? 39
1. Urban sensing @ UCLA: http://urban.cens.ucla.edu/
2. MetroSense project @ Dartmouth:
http://metrosense.cs.dartmouth.edu/
3. Urbanets: http://cs.njit.edu/~borcea/papers/ieee-pervasive07.pdf
4. Micro-Blog: http://synrg.ee.duke.edu/microblog.html
5. CenceMe:
http://www.cs.dartmouth.edu/~sensorlab/pubs/cenceme_sensys08.pdf
6. Darwin phones: http://www.ists.dartmouth.edu/library/478.pdf
7. Pothole patrol: http://db.csail.mit.edu/pubs/mobisys08.pdf
8. ParkNet:
http://www.winlab.rutgers.edu/~suhas/SuhasMathur_Mobisys2010.pdf
40
1. Two decades of mobile computing
2. Infrastructure support for mobility
3. Mobile social computing
4. People-centric sensing
5. Programming mobile ad hoc networks Migratory Services: context-aware client-service programming
Spatial Programming: location-aware imperative programming
Contory: SQL-like declarative programming
6. Vehicular computing and networking
7. Privacy and security in mobile computing
41