From Context Awareness to Socially Aware and Interactive Systems - - PowerPoint PPT Presentation

From Context Awareness to Socially Aware and Interactive Systems

Paul Lukowicz DFKI/University of Kaiserslautern Germany

Overview

• Part 1: Introduction of key concepts
• Part 2: Basic technologies

– Sensing – Reasoning

• Part 3:

– phenomenology of complex socio-technical systems – applications and outlook

Overview

• Standard Sensors

– Motion, Sound, Location, FSR,RFID, BlueTooth

• Exotic Sensors

Acceleration Sensor

• Acceleration sensors provides two types of infomration

Microphone GPS Compass Bluetooth Wireless LAN Accelerometer Gyroscope Camera

Phone Sensors

• Acceleration sensors provides two types of infomration

 Angle towards the gravity vector (tilt)

Acceleration Sensor

1g X Y Z

Acceleration Sensor

• Acceleration sensors provides two types of infomration

 Angle towards the gravity vector (tilt  Change of motion speed

X Y Z

Acceleration Sensor

• Acceleration sensors provides two types of infomration

 Angle towards the gravity vector (tilt)  Change of motion speed  These components can not be easily separated !

1g

Begrüssung

• 2g

+2g 0g +1g /-2g /+2g /0g

time

0g

Filterung und Bechleunigungskomponenten

Raw signal Low pass High pass High pass (fcut=2Hz) Low pass (fcut=2Hz)

Walking Up and Down Stairs

Acceleration on the upper leg Down Up

IMUs

• XSENSE inetrial Measurement Unit (IMU)

– 3 axis accelerometer – 3 axis magnetic – 3 axis gyro

Derives absolute orientation of the device with respect to

• gravity direction
• north

Application of IMUs

Magnetic Field and Light

Multisensor Signal

detect sitting down

Sound

FFT and LDA with kitche chene nett tte sounds ds

Water Pipe Sounds

Forgarty, Au, Hudson, UIST 2006

Water Pipe Sounds

Forgarty, Au, Hudson, UIST 2006

Water Pipe Sounds

Audio: Conversation

• Intensity difference reveals if the owner is

speaking or listening

– human voice distinguishable from other noises

Dietary Monitoring

Initial success in detecting and classifying chewing sounds

– How many bites ? – What type of food was eaten?

Oliver Amft, Mathias Stäger, Paul Lukowicz, and Gerhard Tröster Analysis of Chewing Sounds for Dietary Monitoring,

• Proc. UBICOMP 2005, Tokyo

microphone gets sounds through bone conduction

Microphone position

1 2 3 4 5 6

Bite Detection

4 bites of apple

FSR for Muscle Activity Monitoring

• thickness <0.5mm
• sampling rate: >100Hz
• power loss <1mW

typical force curve

FSRs in den Schuhsohlen

normal hoch runter

Muskelaktivität

• Bewegung wird durch das Zusammenspiel verschiedener Muskel

hervorgerufen

• Auch kleine Unterschiede in den Bewegungsmustern sind klar in der

Muskelaktivität zu erkennen

Long vs. Short Strides

short long

Up vs. Downstairs

up down

Up vs. Downstairs

up down

Squats Technique

correct asymetric cheat

RFID

RFID Example: Intel

Patterson, Fox, Kauz, Philipose, ISWC 2005

Bluetooth: Presence of People

BT scans can reveal identities of people who interact with each other

BT

Location

• Types

– absolute coordinates – semantic (e.g. which room) – proximity

• Parameters

– precision (30cm to some meters until one room) – reliability (how often am I right ?) – installation – price – API

• outdoor location trivial (GPS),
• indoor location unsolved

Beacon Based Principles

Beacon Based Principles

Distance Estimation Methods

• 1. Time of flight

– for indoor environments signal speed should be small – problems with multi-path

• 2. Signal strength

– problems with non distance dependent attenuation

Beacon Technical Issues

• Occlusions/Absorbtion

– E.g. metal absorbs radio signals

• Reflections

– signals bounce around providing false information

• Noise
• User Acceptance

– electro-smog “hysteria”

Beacon Technologies

• Ultrasound: Time Of Flight

– used with RF for communication and synchronization

• RF: Time Of Flight, Signal Strength

– WiFi (RADAR, Ekahu) – Bluetooth,Zigbee – Active RFID – Ultrawideband (UWB e.g. Ubisense)

Beacon Technologies

• Ultrasound: Time Of Flight

– used with RF for communication and synchronization

• RF: Time Of Flight, Signal Strength

– WiFi (RADAR, Ekahu) – Bluetooth,Zigbee – Active RFID – Ultrawideband (UWB e.g. Ubisense)

Wifi Signal Strength

WLAN Fingerprinting

UWB Location

Example of UWB

• Ubisense

– Ubisense delivers a precise, real-time location system (RTLS) utilizing ultra-wideband (UWB) technology that locates assets and people within 30cm / 12" in 3D.

15’000 Euro for a research set !

Performance UWB

Floor Sensors

• Capacitive tracking of objects on the floor

– www.future-shape.com

Overview

• 1. Standard Sensors

– Motion, Sound, Location, FSR,RFID, BlueTooth

• Exotic Sensors

Relative position information?

48

Many activites are determined by relative positions of two body parts. („Hand is near chest.“)

Relative Positioning

49

Acceleration, Gyroscopes and Earth Magnetic Field Sensors Example: Xsens Relative Position information

• nly indirectelyavailable (Euler

angles + knowledge about sensor positions) Ultrasound Example: Relate Brick Obstacles shield measurements. RF Based Example: Ubisense Low accuracy and infrastructure needed

Physical Principle

50

Transmitter generates a magnetic field at a certain resonant frequency f Receiver is calibrated to resonant frequency f

51

Magnetic Field is oscilating Coil 1 Coil 2 Coil 4 Coil 3 Coil 5

Trinken Magnetic

Poor Men’s X-Ray: Capacitors

Physical background

• Human body can be considered as part of a

capacitor

• change of distance or moving of inside organs

result in capacitance change

Sensing Setup

attached to fit tightly but remain comfortable

Signal Examples

Eye Tracking

Andreas Bulling, ETH Zürich

Augenverfolgung

Eye Activity when Reading

• Bulling Eth Zürich

Bread cutter Mixer Water boiler Running Cutting -> bread, salami Mixing level Consitency of fluid Amount of boiled water

Current Sensing

Current Sensing

iSensor:

• stream raw ADC values of the current used power

consumption

• stream recognized device activity events

Kitchen Scenario

Why we choose a kitchen scenario:

• Lot of kitchen devices provide different modes
• Kitchen devices can be used in different ways
• In general: Preparing food is still quite hard to

recognize

Analyzed kitchen devices:

Kitchen Scenario – Signal Analysis Device: Mixer – Raw ADC values

Kitchen Scenario – Signal Analysis Device: Bread cutter – Raw ADC values

Kitchen Scenario – Signal Analysis Device: Water boiler (cold water) – Raw ADC values

Kitchen Scenario – Signal Analysis Device: Fridge – Raw ADC values

Kitchen Scenario – Signal Analysis Device: Egg boiler, toaster, coffee machine

3 or 5 eggs boiled Soft, medium or hard boiled Small or big coffee Bright / brown or dark toasted

Kitchen Scenario – Data Processing

Objective: avoid data streaming!!!

• > send only events (e.g. mixer mode: level 2)

Jennic Zigbee microcontroller for data processing

• > just simple feature calculation methods and recognition

algorithms Features Min Max Sum Variance Average (take max over the last 600ADC and calc features using the last 21 max ADC) Time Duration of a device activity/mode

+

Kitchen Scenario – Data Processing

Objective: avoid data streaming!!!

• > send only events (e.g. mixer mode: level 2)

Jennic Zigbee microcontroller for data processing

• > just simple feature calculation methods and recognition

algorithms Thresholds could be sufficient!!!

Kitchen Scenario – Data Processing

Some examples:

Kitchen Scenario – Data Processing

Evaluation

Single device usage: Fridge – door open events: 35 hours monitored, 91% recognition rate, 9% missed

Evaluation

Water boiler

• boil 0.75l, 1.0l, 1.25l, 1.5l and 1.7l of cold water

deviation: between 0.006l and 0.076l

Evaluation

Bread cutter

• cut slices of bread and salami, 6 different people,

28 slices of bread, 29 slices of salami

93,10% 96,42% Classificatio n:

Cutting activites were recognized two times

Evaluation

Mixer - 7 users, mix both a liquid and a creamy fluid: chocolate milk (250ml milk, 2 spoons of chocolate powder) + quark-yoghurt dish (250g quark and 1 yoghurt)

Result: 100% of chocolate milks = fluid liquids 100% of quark-yoghurt dishes = creamy liquids 57% of quark-yoghurt dishes = change from medium consistency to a creamy one

Evaluation

6 persons performed:

Evaluation

Several devices at the same time:

+ + +

Overview

• Part 1: Introduction of key concepts
• Part 2: Basic technologies

– Sensing – Reasoning

• Part 3:

– phenomenology of complex socio-technical systems – applications and outlook

Overview

• 1. General Background
• 2. Examples

What is Context Recognition ?

Embedded Controllers: feedback control loop Artificial Intelligence: imitating human cognition

• often video based
• includes interpretation

Context Recognition: mapping signals from simple sensors onto a set of predefined, environment related states

Context Aware Systems 2000

• Gellersen, Schmidt, Beigl 1999, from the SmartIts project
• also Day, Abowd, Pentland, Starner, Schiele,... around 2000

acceleration, sound, light...

Context Aware Systems 2000

• Gellersen, Schmidt, Beigl 1999, from the SmartIts project
• also Day, Abowd, Pentland, Starner, Schiele,... around 2000

acceleration, sound, light... modes of locomotion

Microphone GPS Compass Bluetooth Wireless LAN Accelerometer Gyroscope Camera

....and 2012

Context Recognition

IEEE PAMI 06, PMC 07, Pattern Rec 08, ACM TIST 11, Springer PAA 11 ISWC, Pervasive,..

Basic Principle

gyro acc

Basic Principle

class membership inferred from coordinates only

gyro acc ???????

Basic Principle

gyro acc

class membership inferred from coordinates only

Separation Plane

gyro acc

recognition given through separation plane

complex separation plane

gyro acc

• ften complex and impossible to

represent in an analytic way

Recognition Issues

• 1. representation of the

separation plane – memory requirements !

• 2. locating a point with respect

to the separation place

• 3. Identifying the separation

plane – Generation from a sample – cost minimazation

gyro acc

Learning Separation Planes

gyro acc

Learning Separation Planes

separation plane is not perfect

• ambivalent sensor data
• size of the sample
• sensor noise

gyro acc

Begrüssung

• 2g

+2g 0g +1g /-2g /+2g /0g

time

0g

Analyzing Recordings

Standing up Sitting down Sitting Standing Sitting

Torso Wrist

Arm on table Arm resting on table Arm moving Arm moving Handshake time [s] acceleration [g]

Time Series

z„„1 z„„2 z„„3 z„„N … z„1 z„2 z„3 z„N …

Process P„ Process P„„

t

Time Series

z*1 z*2 z*3 z*N …

Process P„ Process P„„

t ?????

Context Rec. Challenges

• 1. Ambiguous sensors

– e.g. phone in a pocket tracking patient care

• 2. Lack of representative training data

– huge inter and intra person variations – training with real users often not practicable

• 3. Inconsistent state space

– e.g walking, having a meal and taking a pill – different granularities, temporal scales, and levels of abstraction

• 4. Dynamically changing sensor configurations

– phone moves around, different phones, new sensors

Context Problem Types

state action situation parameters

Context Problem Types

state action situation parameters

• physical characterization
• ‘classical’ signal processing

?

Context Problem Types

state action situation parameters

• separable in time invariant

feature space

• sliding window algorithms

size window shift

Signal stream

Context Problem Types

state action situation parameters

• characterized by signal sequence
• time series analysis
• ‘spotting’ problem

Does this section contain a gesture ?

• r this one?

Context Problem Types

state action situation parameters

• specific, sequence/distribution of actions
• PCFG, Bayesian networks, ontologies…

Context Problem Types

state action situation parameters

Context Recognition Example

does this section contain a gesture ?

• r this one?
• no model of NULL class
• long NULL class segments
• variable length ‘true positives’

Confusion Matrix

Spotting Errors

Confusion Matrix

Precission Recall Graphs

Spotting Errors

Overview

• 1. General Background
• 2. Examples

Example: Assembly Monitoring

Example: Assembly Monitoring

1,3 microphones 2,4,5 accelerometers 6 computer

• P. Lukowicz, J. Ward, H. Junker, M. Stäger, G. Tröster, A. Atrash, T. Starner

Recognizing Workshop Activity Using Body Worn Microphones and Accelerometers, Pervasive Computing, pages 18-22, Vienna, Austria, 18.-23. April 2004.

Drilling

∘Well defined, steady motion

• slowly turn handle to lower, then return
• constant machine noise

time

z-axis sound x-axis

Sawing

∘Well defined, repetitive motion

• back and forth
• loud correlated sound

time

x-axis z-axis sound

Approach

Sound classif. Motion classif. Sound Analysis Segment info. Sound data Motion data Final predictions IA Prediction Sequence Audio Ch2. FFT LDA Audio Ch1.

• Min. Dist.

Audio Ch2. jm smoothing

Signal al Seg egme mentation ntation

IA Audio Ch1. Audio Ch2.

1 » Þ »

chest wrist chest wrist

I I d d 1 >> Þ <<

chest wrist chest wrist

I I d d

hand on the switch

So Sound nd Classific ssification ation

FFT LDA IA Audio Ch1. Audio Ch2.

• Min. Dist.

Pre-computed LDA training vectors

Sound nd Cl Class ssification fication

Minimum distance

• r K-Nearest Neighbour

Test frame 256 point FFT reduced to 3D with LDA

Fram ame e by by frame me evalu luation ation

Audio Ch1. Audio Ch2.

• ground truth

Fram ame e by by frame me evalu luation ation

FFT LDA Audio Ch1. Audio Ch2.

• Min. Dist.
• ground truth
• LDA

Sawing

Fram ame e by by frame me evalu luation ation

IA Prediction Sequence Audio Ch2. FFT LDA Audio Ch1.

• Min. Dist.

Audio Ch2.

• ground truth
• LDA
• LDA+IA

• ground truth
• LDA+IA

Continuous ntinuous evalu luation ation

• ground truth
• LDA+IA
• jm(LDA+IA)

IA Prediction Sequence Audio Ch2. FFT LDA Audio Ch1.

• Min. Dist.

Audio Ch2. jm smoothing

Acceleration Modeling

HMM models ML classification 3x3 axis raw data prediction (sampled at 100Hz) single gaussion, mostly feed forward

Results

• Sound: 109 Events

– 97 correct – 94 insertions – 2 deletions – 10 substitutions

• Motion: von 109 Events

– 102 correct – 23 insertions – 7 substitutions

Sound classif. Motion classif. Sound Analysis Segment info. Sound data Motion data

Results

• Sound: 109 Events

– 97 correct – 94 insertions – 2 deletions – 10 substitutions

• Motion: von 109 Events

– 102 correct – 23 insertions – 7 substitutions

Sound classif. Motion classif. Sound Analysis Segment info. Sound data Motion data Final predictions

Sensor Fusion

• Sound: 109 Events

– 97 correct – 94 insertions – 2 deletions – 10 substitutions

• Motion: von 109 Events

– 102 correct – 23 insertions – 7 substitutions

• Combined 109 Events

– 92 correct – 0 insertions – 17 deletions – 0 substitutions

Sound classif. Motion classif. Sound Analysis Segment info. Sound data Motion data Final predictions

Classifier Fusion

• Methods

– Comparison – Highest Rank – Borda Count – Logistic Regression

• Aims

– improve person independent performance – deletion/insertion tradeoffs

Sound classif. Motion classif. Sound Analysis Segment info. Sound data Motion data Final predictions Fusion

NULL

Classifier Fusion

Sound classif. Motion classif. Sound Analysis Segment info. Sound data Motion data Final predictions Fusion

NULL

• Methods

– Comparison – Highest Rank – Borda Count – Logistic Regression

• Aims

– improve person independent performance – deletion/insertion tradeoffs

• Result

– logistic regression 10% better for person independent case 400 events in 20 sequences, from 5 subjects

Activity Recognition of Assembly Tasks Using Body-Worn Microphones and Accelerometers, J.A. Ward and P. Lukowicz and G. Troester and T. Starner, IEEE Trans. Pattern Analysis and Machine Intelligence, 28:10, October 2006

Ultrasound

Location classif. Motion classif. Hands Location Segment info. Ultrasound Motion data Final predictions Fusion

NULL

Ultrasound

Approach

drift away after a few sec 2Hz sampling rate and many errors

Approach

LSQ estimate from US, „plausability“ filtered

• ext. Kalman from

from US+IMU

• ext. Kalman from US

Approach

EM clustering

• n multivariate

Guassian mixture discretization into motion strings, creation of matching cost graph, minimum tresholding

• n location spotted segments

distance based spotting with temporal smoothing

Approach

• more detailed examination of output of trajectory spotting with respect to location (with

tresholds or SVM classification)

• concurrency based plausibility analysis

Results

location spotting trajectory spottingspotting SVM based post processing concurency resolution

Muscle Monitoring

Candidate class.. HMM classif. Segment info. Motion data Final predictions Similarity search Partitioning Candidate sel. HMM Features Hand muscles

Sensing Muscle Activities with Body-Worn Sensors, Amft, O., Junker, H., Lukowicz, P., Tröster, G., and Schuster, C., In BSN 2006: Proceedings of the International Workshop on Wearable and Implantable Body Sensor Networks., pp. 138-141, April 2006

Muscle Monitoring

Skoda Qaulity Control

2nd Generation Activitiess

Sensors: Location

Sensors: Location

Ubisense tags

Location Raw Data

Smoothed and Filtered Result

Sensors: Motion

Sensors: Inertial

Moption Signal

Sensors: FSR for Muscle Activity

FSR for Muscle Activity Monitoring

Sensing Muscle Activities with Body-Worn Sensors, Amft, O., Junker, H., Lukowicz, P., Tröster, G., and Schuster, C., In BSN 2006: Proceedings of the International Workshop on Wearable and Implantable Body Sensor Networks., pp. 138-141, April 2006

Combined Signal Examples

Algorithmic Approach

• Processing separation

– Activity-wise – Sensor- / feature-wise

• High recall spotting stage
• Incremental fusion of additional information

Spotter for class 1 Spotter for class N

Results

1 open hood 2 close hood 3 open trunk 4 check trunk 5 close trunk 6 fuel lid 7 open left door 8 close left door 9 open right door 10 close right door 11 open two doors 12 close two doors 13 mirror 14 check trunk gaps 15 lock check left 16 lock check right 17 check hood gaps 18 open the spare wheel box 19 close the spare wheel box 20 writing

Spotter for class 1 Spotter for class N Motion spotting FSR classification of spotted gestures Masking step FSR Masking step location Masking step FSR + location final

+

x

• 8 subjects,
• 3680 checking
• 560 minutes of data

Overview

• Part 1: Introduction of key concepts
• Part 2: Basic technologies

– Sensing – Reasoning

• Part 3:

– phenomenology of complex socio-technical systems (collective sensing and communication phenomena) – applications and outlook

Application Types

exploiting people‟s mobility to sense parameters over large areas sensing/recognizing collective phenomena collaborative sensing and reasoning to improve accuracy

Application Types

exploiting people‟s mobility to sense parameters over large areas sensing/recognizing collective phenomena collaborative sensing and reasoning to improve accuracy influencing behaviour in the large

Application Types

exploiting people‟s mobility to sense parameters over large areas sensing/recognizing collective phenomena collaborative sensing and reasoning to improve accuracy influencing behaviour in the large

Large Scale Sensing

exploiting people‟s mobility to sense parameters over large areas

Example

TomTom realtime traffic combining

• Speed/location data from TomTom systems
• Phone GSM cell data from

Crowd Density Classification

from number of discoverable devices to crowd density

Discoverable BlueTooth Devices

Wepner, Lukowicz accepted, PhoneSense 11

160

Experiment Area - Munich Oktoberfest

160

3 days at the Octoberfest 2010

162

Methods I

• Individual device based estimation

– Number of Bluetooth devices – Mean signal strength – Variance of the signal strengths

163

Methods II

• Collaborated device based estimation

– Average number of devices – Variance of the number of devices – Variance of all signal strengths

Crowd Density Classification

a) b) c) d) e) f) g) h) i) 0!% 30!% 60!% 90!% 81!% 82!% 63!% 76!% 78!% 63!% 58!% 60!% 50!% 64!% 40!% 41!% 40!% 64!% 39!% 29!% 39!% 34!%

multiple phones, collaborative

Application Types

exploiting people‟s mobility to sense parameters over large areas collaborative sensing and reasoning to improve accuracy

Inertial tracking

true path estimated

฀ r s  r v (t)dt  r a (t)dt



 



dt



Collaborative Location

Collaborative Location

Collaborative Location

Collaborative Location

Experimental Results

15 30 45 60 75 90 25 50 75 100 time min mean localisation error m

PDR error collaborative localisation error, 10 20

Based on 60km of traces from 10 people at a 3 day festival in Malta

Kloch, Lukowicz, ISWC 11

Application Types

exploiting people‟s mobility to sense parameters over large areas sensing/recognizing collective phenomena collaborative sensing and reasoning to improve accuracy

Simple Idea

= consumer confidence

Allianz Arena Experiments

3 second and 1 first league games at Allianz Arena recorded with between 8 and 12 participants

Approval Disapproval Cheering

Allianz Arena Experiments

• Classification of relevant audio segments works

with an accuracy of almost 100 %

• Automatically selecting relevant segments from

the continuous audio stream is still a work-in- progress cooperation between many phones !

Application Types

exploiting people‟s mobility to sense parameters over large areas sensing/recognizing collective phenomena collaborative sensing and reasoning to improve accuracy influencing behaviour in the large

As Coupling Gets Stronger, System Behavior Can Change Completely: Traffic Breakdowns

Thanks to Yuki Sugiyama

Socially Interactive Systems Research

scenario boundary conditions sensing, information processing and spread impact on human decision making and social dynamics impact on physical dynamics (evacuation or traffic)

System Simulations

scenario boundary conditions sensing, information processing and spread impact on human decision making and social dynamics impact on physical dynamics (evacuation or traffic)

Simulation Framework

Movement Helbing AmI guidance system Models:

• Trust
• …

Decision

?

PDR Collaborative localisation

Helbing Motion Simulation

Helbing +Collaborative PDR

Helbing +Collaborative PDR+Trust

[LPF12] P. Lukowicz, S. Pentland, and A. Ferscha. From context awareness to socially aware

• computing. Pervasive Computing, IEEE, 11(1):32–41, 2012.

Literature

[AL09] O. Amft and P. Lukowicz. From backpacks to smartphones: Past, present, and future of wearable computers. Pervasive Computing, IEEE, 8(3):8–13, 2009. [BLA08] D. Bannach, P. Lukowicz, and O. Amft. Rapid prototyping of activity recognition applications. Pervasive Computing, IEEE, 7(2):22–31, 2008. [CAL10] J. Cheng, O. Amft, and P. Lukowicz. Active capacitive sensing: Exploring a new wearable sensing modality for activity recognition. Pervasive Computing, pages 319–336, 2010. [GLBR10] H. Gellersen, P. Lukowicz, M. Beigl, and T. Riedel. Cooperative relative

• positioning. Pervasive Computing, IEEE, (99):1–1, 2010.g modality for activity
• recognition. Pervasive Computing, pages 319–336, 2010.

[GBLH11] A. Grunerbl, G. Bahle, P. Lukowicz, and F. Hanser. Using indoor location to assess the state of dementia patients: Results and experience report from a long term, real world study. In Intelligent Environments (IE), 2011 7th International Conference on, pages 32–39. IEEE, 2011. [KKL+10] K. Kloch, J. Kantelhardt, P. Lukowicz, P. Wu ̈chner, and H. de Meer. Ad- hoc information spread between mobile devices: A case study in analytical modeling of controlled self-organization in it systems. Architecture of Computing Systems-ARCS 2010, pages 101–112, 2010.

Literature

[KPLF11] K. Kloch, G. Pirkl, P. Lukowicz, and C. Fischer. Emergent behaviour in collaborative indoor local- isation: an example of self-organisation in ubiquitous sensing

• systems. Architecture of Computing Systems-ARCS 2011, pages 207–218, 2011.

[KWK+09] K. Kunze, F. Wagner, E. Kartal, E. Morales Kluge, and P. Lukowicz. Does context matter?- a quantitative evaluation in a real world maintenance scenario. Pervasive Computing, pages 372–389, 2009 [LTGLH07] P. Lukowicz, A. Timm-Giel, M. Lawo, and O. Herzog. Wearit@ work: Toward real- world industrial wearable computing. Pervasive Computing, IEEE, 6(4):8–13, 2007.. [LCG11] P. Lukowicz, T. Choudhury, and H. Gellersen. Beyond context awareness. Pervasive Computing, IEEE, 10(4):15–17, 2011. [LNN+12] P. Lukowicz, S. Nanda, V. Narayanan, H. Albelson, D.L. McGuinness, and M.I.

• Jordan. Qual- comm context-awareness symposium sets research agenda for context-aware
• smartphones. Per- vasive Computing, IEEE, 11(1):76–79, 2012.

[WLG11] J.A. Ward, P. Lukowicz, and H.W. Gellersen. Performance metrics for activity

• recognition. ACM Transactions on Intelligent Systems and Technology (TIST), 2(1):6, 2011.

[WLTS06] J.A. Ward, P. Lukowicz, G. Troster, and T.E. Starner. Activity recognition of assembly tasks using body-worn microphones and accelerometers. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 28(10):1553–1567, 2006.

Some Research Groups

• MIT Media Lab

– http://hd.media.mit.edu/

• Dartmouth College
• K-Lab Pisa

– http://www-kdd.isti.cnr.it/