Upload
santanu-sarma
View
239
Download
3
Tags:
Embed Size (px)
Citation preview
Sensemaking from Distributed and Mobile Sensing Data: A Middleware
Perspec;ve
S.Sarma, N. Venkatasubramanian, N. DuA
1
Overview
• Introduc0on to Crowdsensing and Sensemaking
• A Middleware Perspec0ve • Example Middleware Pla?orms and techniques
• Research Direc0ons
2
Mobile Phone Trends
• Mobile subscrip;on 5.96 billion 2011 es;mate
• Smartphones (487.7 million) exceeding PCs (414.6 million)
• More Mobile Internet Users Than Wireline Users in the U.S. by 2015
• Smartphone and bandwidth cost reduces
• Smart devices contribute to more than 90% of mobile data traffic
3
Sensors In Mobile Phones
• MEMS & sensors for cell phones, expanding from $ 3.5 bn in 2009 to $7.9 bn in 2015 [Yole Developpement]
• Smartphone sensors to be $ 6 bn business by 2016 [Juniper Research] • 44 % of the mobile phones will be smartphones in 2015 • 7x increase in mobile health apps from 2010 to 2011 • mo;on sensor in smartphones and tablets will expand to $ US 2.1 billion in
2015 with a 25.3 % CAGR, up from $1.19 billion in 2011 (IHS iSuppli) 4
Mobile Sensors Trends
Source: IHS Consumer & Mobile MEMS Market Tracker, April 2014. 5
Mobile Data Delivery Everywhere
6
Smart devices contribute to more than 90% of mobile data traffic
The exploding number of apps is driven by a huge up;ck in the number of smart devices
~55%
Cisco’s report 2014
Crowdsourcing and CrowdSensing
7
Pushing toward more interven0on
Power of the Crowd
• Using mobile crowdsensing to – Leverage already deployed smartphones
– Extend the ranges of exis0ng in-‐situ sensors
– Send mobile users to specific loca0ons
• Crowdsensing broad use cases – Disaster and emergency response – Personal health monitoring and wellness
– Smart spaces and their effec0ve u0liza0on
8 [YKL11] M. Yuen, I. King, and K. Leung. A survey of crowdsourcing systems. In Proc. of IEEE Interna0onal Conference on Social Compu0ng (SocialCom’11), pages 766–773, Boston, MA, October 2011.
• Earthquakes • Hurricanes • Tornadoes • Energy/utility outages • Fire hazards • Hazardous materials releases • Terrorism/
Emergency Use Cases
9
Emergency Response
During Fire accidents can cause electric power failure. Mobile broadcast can be used to provide direc;ons to the users about rescue opera;ons.
10
Emergency situa;on Automa;c Altering can be used to inform family, rescue teams, or nearby cars / passengers in case of accidents.
Emergency Response
11
Sensing -‐> Sensemaking
Alert System
Severity
Personal Sensing to indicate Fall detec0ons, injury severity, alerts in old age people to
provide scalable health care 12
Sensing -‐> Sensemaking Radia0on field near Fukushima
Crisis Map Showing Latest Informa0on
Hazardous gas in campus
Spa0al Field Sensing With Mobile Sensors
13
Sensing -‐> Sensemaking
• Avoiding congested streets in a city • Finding the most popular booth in a fair • Searching for the ride with shortest lineup in an amusement park
14
SenseMaking : Purpose & Goals
u Simple and Easy-‐to-‐Use Framework for Sensing, Actua0on and Collabora0on using mobile phone
u Powerful addi0onal sensing abili0es and features for community of users by community of users
u Understand user and group context efficiently u Building energy-‐efficient collabora0on apps
over exis0ng mobile pla?orms u Supported and empowered by community of
users for community of user
15
The Problem – A cross layer, end to end issue
§ Several barriers and huge investment of 0me to build collabora0ve smart applica0ons
§ Lack of a framework to ease and speed the development of applica0ons
§ Non-‐Scalable, Ad-‐hoc, non-‐standardized API § Unsupported network infrastructure, and
configura0ons
16
Solu0on to the Problem – Middleware Approach, Hierarchy for Scale…
• Design and Develop and Open source distributed middleware framework suppor0ng collabora0ve mobile sensing
• Provide API and libraries to perform: – Collabora0on – Virtual Sensing and Compressive Context Determina0on
– Computa0onal Offloading – Cloud interface for scalability
17
Middleware Pla?orms and Techniques for Sensemaking
• On phone, on broker (SenseDroid, SATWARE) • Techniques implemented in middleware
– Compressive and Collabora0ve Sensing – Virtual Sensing for Sensemaking – Seman0cs Driven Sensing and Actua0on
• Combining In-‐situ Sensors with Mobile Crowdsensing
18
Combining In-‐situ Sensors with Mobile Crowdsensing
Pushing toward more interven0on
• For sensing tasks not covered by any in-‐situ sensors – Try opportunis0c and par0cipatory sensing using nearby mobile users
• What if there are no nearby mobile users • Pushing toward even more interven0on à Crowdsourcing
19
Explosion of Contextual Data Delivery
20
Emergency response
Transporta0on
~2.5 M mobile apps
Entertainment Mobile social networks
Healthcare
Shopping
Apps have various performance needs (reliability, ;meliness, quality…)
Explosion of Contextual Data Delivery
21
Explosion of Contextual Data Delivery
22
SenseDroid Architecture
…
Mobile Users
…
…
Internet /Public Cloud
Middleware Broker
Wi-‐Fi AP
3G AP
Query/ Response
Cloud Users
• Use compressive sensing with computa0onal offloading for energy-‐efficiency
• Use collabora0on for addi0onal and efficient sensing abili0es
• Leverage reconstruc0on abili0es of compressive sensing to improve robustness and reliability
23
Mul0-‐0ered Hierarchical Architecture
24
SenseDROID Distributed Middleware
APPS$1$
Communica.on$
Sensing$&$Sampling$
Context$Processing$&$Fusion$$
Query$+$Storage$
Manager$
Privacy$&$se>ngs$
Communica.on$
Sensing$&$Sampling$
Context$Processing$&$Fusion$
Query$+$Storage$
Manager$
Privacy$&$se>ngs$
Query$&$$Response$
Analysis$&$Processing$
Query$+$Storage$
Communica.on$
Collabora.on$
Data$Collec.on&$Comp.$Sampling$
Infrastructure$Sensing$$
Manager$
S1$ S2$ Sm$…….$
Query$ Response$
…….$
Query$&$$Response$
Infrastructure$Sensors$
Mobile$Node$
Broker$
Mobile$Node$
APPS$2$
APPS$N$
Cloud
AP
S1$
Sn$
S1$
Sn$
25
Sensemaking Using Compressed Sensing
• A random sampling technique that can represent Sparse signal with few random measurements
• Represents a Sparse Signal with few salient coefficients in a transformed domain
• Integrates sensing, compression, processing based on new uncertainty principles
26
Collabora0ve Compressive Sensing
Sink Node(Broker) Mobile Node Sampled Mobile Sensor Legend No#of#Measurements##
Reco
nstruc
tion##Error#(M
SE)#
Number of Measurement Accuracy of Sensemaking Number of Measurement Energy Consumed in Sensing Accuracy of Sensemaking Scalability and Coverage
Traded-‐off
27
Sensemaking using Virtual Sensing
Ambient Light
3D Magnetometer
3D Accelerometer
Barometer
Processing( CompressedSensing andCalibration)
SensorFusion
3D Gyroscope
Ambient Light
Barometer
Thermometer
Accelerometer
Gyrometer
Inclinometer
Orientation
Compass
Physical Devices
IsDriving
IsRunning
IsWalking
IsSitting
AtHome
InOffice
IsIndoor
IsAlone
hasFallen
IsHappy
Virtual SensingProcessing
Sampling &Data
Collection(Compressive
Sampling,Adaptive
Sampling)
Location Contexts
Activity Contexts
Context Processing
Social Contexts
Emotional Contexts
EnvironmentalContexts
Health Contexts
28
Research Direc0ons
• Energy Efficiency – Exploit collabora0ve & compressive sensing for energy efficiency
• Incen0ve Mechanisms – Device incep0ves for par0cipa0on and collabora0on
• Privacy Regula0on – Facilitate privacy preserving incen0ves
• Heterogeneity in Mobile Cloud – Use and exploit heterogeneity of sensors and devices
29
30
RELATED WORK REVIEW • Energy-Efficient Smart Spaces - Smartphone
Augmented Infrastructure Sensing
• Optimizing Event Detection on Smartphones
• Spatial-temporal Information Gathering using Smartphones
Smart Spaces
• Difference scales of intelligent systems: such as ci0es, stadiums, airports, building, and roads
• Ci0zens of a smart space are not observers but ac0vely help the officials to make the space berer, e.g., – Safer – More entertaining – More energy efficient – More situa0on-‐aware
• Similar to smart home, but across mul0ple users
31
Pla?orm for Public Smart Spaces
• Goal: develop a pla?orm to provide safety with sustainability for smart spaces
• Detec0ng many events in an energy-‐efficient way – Security related events: fights riots, protests, and demonstra0ons
– Hazardous events: fires, chemical leaks, and stampedes
– High crowd levels for poten0al conflicts
London School of Economics’ app that monitors crowd safety at events 32
Limita0on of Current Approach
State-‐of-‐the-‐art: Infrastructure sensing using in-‐situ sensors – High installa0on and maintenance cost – Insufficient node coverage ß limited budget – Does not scale! ß for crowded events
33
Usage Scenario #1
• Task: Sensing temperature at CS building • What if there is no working thermometer at the CS building?
– Infer the temperature by nearby buildings – Infer the temperature provided by 3G/4G smartphone users walking by the CS building
34
Usage Scenario #2 • Task: Traffic surveillance for safety applica;ons • What if the fixed surveillance videos are insufficient ?
– Leverage videos from nearby in-‐situ cameras – Leverage videos captured by police officers, fire fighters, and EMTs
– Leverage large volume of user-‐generated, geo-‐tagged videos captured by ci0zens
35
Dashboard hrp://info.theomegagroup.com/blog/bid/134307
36
System Architecture
37
Current Prototype
38
Challenges
• How to efficiently carry out the sensing requests? • How does the broker assign the requests to workers? • How to guide workers to the correct sensing loca0on? • How to efficiently process the raw sensory data? • Where to process the raw sensory data? • Can we leverage mul0ple close-‐by sensors for higher accuracy?
39
40
RELATED WORK REVIEW • Energy-Efficient Smart Spaces - Smartphone
Augmented Infrastructure Sensing
• Optimizing Event Detection on Smartphones
• Spatial-temporal Information Gathering using Smartphones
Event Detec0on on Smartphones
• Each event may be detected by mul0ple subsets of sensors ß subop0mal sensor subsets? – E.g., traffic jam may be detected by GPS, accelerometer, or GPS + accelerometer
• Mul0ple events may be (par0ally) detected by the same sensors ß uncoordinated sensor usage leads to redundant sensor ac0va0on – E.g., earthquake may also be detected by accelerometer
• Problem: how to select efficient sensing strategies
41
Context-‐aware Mobile Applica0ons
• Increasingly more context-‐aware apps leverage the smartphone sensors for berer user experience
• What is context-‐aware? – Essen0ally inferred from sensor readings!
42
An Equivalent Research Problem
• Context-‐aware apps may – Infer the same context using various combina0ons (sets) of sensors
– Impose diverse accuracy requirements • How to select efficient sensing strategy?
– Sa0sfy all apps’ requirements – Minimize energy consump0on
• Proposal: OSM (Op0mal Sensor Management) middleware
43
OSM Middleware
OSM Middleware
• It sits between apps and hardware • Apps may register or unregister requests through an API at any 0me.
• Our middleware is response to – Maintain a database of ac0ve requests – Determine what sensors to ac0vate at what 0me
44
System Architecture
45
API: 1. Register()/Unregister() 2. Feedback()
Request Manager 1. Manages a Request
Queue 2. Preprocess the contexts
Context Analyzer 1. Context Updater 2. Model Trainer
Resource Manager 1. Barery Monitor 2. Scheduling Algorithm
System Model • Combina0on/Accuracy/
Energy
• Coordinated and efficient sensor usage! • Avoid redundant energy waste!
How to Op0mally Schedule Sensor Ac0va0ons?
• Tradeoff between accuracy and energy consump0on
• Our scheduling algorithms have to pick the best combina0on for all requests
• The already-‐on sensors have to be considered
46
What if WiFi is already on?
Our Proposed Scheduling Problems
Two op0miza0on criteria: – Energy Minimiza;on (EM) Schedule with the lowest energy to sa0sfy all the apps’ requirements – Accuracy Maximiza;on (AM) Schedule with the highest overall accuracy under an energy budget
47
Energy Minimiza0on (EM) Formula0on
48
Minimize energy
Sa0sfy accuracy requirements
Within energy budget
Maximize accuracy
Accuracy Maximiza0on (AM) Formula0on
49
Proposed Scheduling Algorithms
• Energy Minimiza;on Algorithm (EMA) Accuracy Maximiza;on Algorithm (AMA) • Good performance • Suitable for smaller problems due to high complexity
• Efficient Energy Minimiza;on Algorithm (EEMA) Efficient Accuracy Maximiza;on Algorithm (EAMA) • Shorter running 0me • More suitable for smartphones • Inspired by two approxima0on algorithms for the weighted set cover and 0/1 knapsack problems ß But the approxima0on factor proofs do not work in our problems
50
Our Simulator
• We developed an event-‐driven simulator in Java • Baseline algorithm
– Selects the sensors for the highest accuracy of each context
• We compare the scheduling algorithms: – Op0mal : EMA/AMA – Heuris0c : EEMA/EAMA – Baseline
• Collect running apps in Android ac0vity stack from 5 users for three weeks
• Measure power consump0on on a Samsung Galaxy S
8
Energy Saving
• Save at least 40%, compared to the baseline • EEMA achieves a small gap of ∼2% than EMA • EMA terminates in 50ms and EEMA terminates in 1ms
9
Accuracy Improvement
• Increase accuracy by up to 39.06% than the baseline • EAMA achieves a gap of ~1% than AMA • AMA terminates in 5000ms and EAMA terminates in 1ms 53
More Restricted Environments Lead to Higher Gains
54
Lower Accuracy Requirement Less Energy Budget
Save More Energy Higher Accuracy Boost
Larger Problems Result in Higher Gains
55
Save More Energy Higher Accuracy Boost
Real Prototype System
• Implement two heuris0c algorithms and the proposed OSM on Android
• EEMA – Prolongs barery life two 0mes – Achieves accuracy : 93.94%
• EAMA – Prolongs barery life 1.5 0me – Achieves accuracy : 94.85%
56
Summary
• We propose an Op0mal Sensor Management middleware
• Four algorithms with different op0mal criteria and complexity levels for sensor scheduling
• EEMA (EAMA) saves energy (boost accuracy) in real-‐0me • Real implementa0on on smartphone
• Designed for a single smartphone, but the same sensor management mechanisms may be used for event detec0on in smart spaces
57
58
RELATED WORK REVIEW • Energy-Efficient Smart Spaces - Smartphone
Augmented Infrastructure Sensing
• Optimizing Event Detection on Smartphones
• Spatial-temporal Information Gathering using Smartphones
Geospa0al Informa0on Gathering
• A new class of crowdsourcing systems • Requesters: companies and organiza0ons
• Submit geospa0al and temporal-‐dependent tasks (specific 0me and loca0on)
• Task: capturing videos/pictures or collec0ng sensor readings
• Workers: smartphone users • Report their des0na0on and deadline • They wouldn’t mind to take some detour routes for small rewards
59
Detour Planning Problem
• Sample scenario: A smartphone user who needs to get to the Chia-‐Yi HSR Sta,on at 7 p.m. may have a few hours to spare. Why not making some money? – But it’s hard for a person to come up with the detour path
• Our problem: How to find the best detour path for each worker – to maximize the profit (= rewards – costs) – while guaranteeing on-‐0me arrival at the des0na0on
60
System Architecture
61
Feasible Spots
62
Problem Formula0on
Maximize overall profits
Start and end points
No rep. feasible spots
Arrive des0na0on in 0me
Visit each request once Start 0me of each request
Finish 0me of each request
63
Orienteering Problem with Time Window
• A similar problem – Goal: maximize the score – Game: players go to specific spots, and finish the predetermined job for a reward – Not exactly the same: (1) mul0ple feasible spots and (2) travel cost (gas and car deprecia0on)
• We enhanced a dynamic programming based OPTW algorithm [GS09] for an op0mal Detour Planning (DP) algorithm – Runs in polynomial 0me: O( N3Z3 )
64
[RS09] Decremental state space relaxa0on strategies and ini0aliza0on heuris0cs for solving the orienteering problem with 0me windows with dynamic programming. Computers and Opera0ons Research, 36(4):1191–1203, April 2009.
Collec0ng Feasible Spots
• Find 25 landmarks in Taipei (hrp://taipeitravel.net) and Vancouver (hrp://hotels.com)
• Use Flickr API to download the pictures tagged with each landmark, and retrieve the longitude/la0tude
• Use hierarchical clustering algorithm to group these photos at the granularity of blocks (~100 m) ß gives us the feasible spots
• Employ Google map to compute the distance between any two feasible spots
65
Simulator Implementa0on
• We implement a trace-‐driven simulator in C++ • It supports five algorithms
– The proposed DP algorithm – Four heuris0c algorithms
• Highest-‐Reward (HR) ß mimic human behavior • Closest-‐Request (CR) ß mimic human behavior • Highest-‐Reward with On0me (HROT) • Closest-‐Request with On0me (CROT)
66
Simula0on Design
• Parameters – N: number of requests: {5, 10, 15, 20, 25} – T: deadline: {1, 2, 4, 8, 16} (hr) – C: travel cost: {0, 0.06, 0.12, 0.24, 0.48} ($/km)
• Metrics – Total rewards – Running-‐0me – On0me-‐ra0o
67
On0me Ra0o
HR and CR (mimicing humans) à low on0me ra0os! 68
Total Profits
• Although HROT and CROT guarantee on0me arrival, they suffer from low profits
• Compared to HROT and CROT, DP doubles the rewards with 25 requests – More requests à larger gap! 69
DP is Efficient
• Terminates in less than 60 ms • Slower for Vancouver (right) ß more feasible spots
70
Implica0on of Travel Cost
• Higher profits when per-‐km cost is lower
71
Summary
• Studies a new class of crowdsourcing problems – Geospa0al informa0on gathering
• Proposes an op0mal detour planning algorithm based on an OPTW algorithm
• Simula0on results are encouraging • Poten0al Extensions
– Implemen0ng a working prototype – Guide the workers to shoot photos using augmented reality – Quality assurance and cheat detec0on mechanisms
• Designed for collec0ng spa0al-‐temporal mul0media informa0on, but can be extended for event detec0on
72
Ques0ons?
73
Challenges to realize smart spaces • How to efficiently carry out the sensing requests? • How does the broker assign the requests to workers? • How to guide workers to the correct sensing loca0on? • How to efficiently process the raw sensory data? • Where to process the raw sensory data? • Can we leverage mul0ple close-‐by sensors for higher accuracy?