64
Activity Recognition: Linking Low-level Sensors to High-level Intelligence Qiang Yang Hong Kong University of Science and Technology http://www.cse.ust.hk/~qyang/ 1

Activity Recognition: Linking Low-level Sensors to High-level Intelligence

  • Upload
    brinly

  • View
    50

  • Download
    0

Embed Size (px)

DESCRIPTION

Activity Recognition: Linking Low-level Sensors to High-level Intelligence. Qiang Yang Hong Kong University of Science and Technology http://www.cse.ust.hk/~qyang/. What’s Happening Outside AI?. Pervasive Computing Sensor Networks Health Informatics Logistics Military/security WWW - PowerPoint PPT Presentation

Citation preview

Page 1: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

Activity Recognition: Linking Low-level Sensors to High-level

Intelligence

Qiang YangHong Kong University of Science and

Technologyhttp://www.cse.ust.hk/~qyang/

1

Page 2: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

What’s Happening Outside AI?

• Pervasive Computing• Sensor Networks• Health Informatics• Logistics• Military/security• WWW• Computer Human

Interaction (CHI)• GIS…

2

Page 3: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

What’s Happening Outside AI?

3

Wii Apple iPhone

Ekahau WiFi LocationEstimation

Page 4: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

Theme of The Talk

• Activity Recognition:– What it is– Linking low level sensors to high level intelligence

• Activity recognition research: Embedded AI– Empirical in nature– Research on a very limited budget

4

Page 5: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

A Closed Loop

5

Eating, Resting, Cooking, Doing Laundry, Meeting, Using the telephone, Shopping, Playing Games, Watching TV, Driving …

Cooking: Preconditions: (…), Postconditions: (…), Duration: (…)

(From Bao and Intille, Pervasive 04)

Page 6: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

Activity Recognition: A Knowledge Food Chain

Action Model Learning◦ How to model user’s

actions?Activity Recognition

◦ What is the user doing / will do next?

Localization & Context◦ Where is the user?◦ What’s around her?

6

• Knowledge Food Chain• Output of each level acts as input

to an upper level in a closed feedback loop

Page 7: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

Basic: Knowing Your ContextLocations and ContextWhere are you?What’s around you?Who’s around you?How long are you there?Where were you before?Status of objects (door open?)What is the temperature like?…

7

Page 8: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

Knowing Your ContextLocations and ContextWhere are you?What’s around you?Who’s around you?How long are you there?Where were you before?Status of objects (door open?)What is the temperature like?…

8

Page 9: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

Focusing on locations

Input:Sensor Readings

Wifi, RFID, Audio, Visual, Temperature Infrared, Ultrasound, magnetic fieldsPower lines

[Stuntebeck, Patel, Abowd et al., Ubicomp2008]…

Localization ModelsOutput: predicted locations

9

Dr. Yin, Jie @ work (HKUST)

Page 10: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

Location-based Applications: Indoor

• Healthcare at home and in hospitals

• Logistics: Cargo Control

• Shopping, Security

• Digital Wall– Collaboration

with NEC China Lab

10

Page 11: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

How to obtain a localization model?

• Propagation-model based– Modeling the signal

attenuation

• Advantages: Less data collection effort

• Disadvantages:– Need to know emitter locations– Uncertainty

• Machine Learning based– Advantages:

• Modeling Uncertainty Better• Benefit from sequential info

– Disadvantages:• May require a lot of labeled dat

a

11

RADAR [Bahl and Padmanabhan, CCC2000]

Page 12: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

Using both labeled and unlabeled data in subspace learning

• LeMan: Location-estimation w/ Manifolds [J. J. Pan and Yang et al., AAAI2006]

• Manifold assumption: similar signals have similar labels

• Objective: Minimize the loss over labeled data, while propagating labels to unlabeled data

12

Page 13: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

LeMan [J.J. Pan and Yang et al., AAAI2006]

• Supervised vs. Semi-Supervised in a 4m x 5m testbed

• To achieve the same accuracy under 80cm error distance

13

Supervised Semi-supervised

RADAR LeMan

Percentage of labeled data used 100% 23%

Page 14: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

Adding sequences: Graphical Model

• Conditional Random Fields [Lafferty, McCallum, Pereira, ICML2001]– Undirected graph, a

generalization to HMM

14

State=locations

Observations = signals

Not using sequential information

Using sequential information

Support vector regression(supervised learning)

CRF(supervised learning)

SemiCRF(semi-supervised learning)

Accuracy 67.33% 83.67% 85.67%

CRF based localization [R. Pan, Zheng, Yang et al., KDD2007]

Page 15: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

What if the signal data distribution changes?

• Signal may vary over devices, time, spaces …

• A -> B: the localization error may increase

15

Transfer Learning!

Page 16: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

Our work to address the signal variation problems

• Transfer Learning– Problem 1: Transfer Across Devices• [Zheng and Yang et al., AAAI2008a]

– Problem 2: Transfer Across Time• [Zheng and Yang et al., AAAI2008b]

– Problem 3: Transfer Across Spaces• [S. J. Pan and Yang et al., AAAI2008]

16

Page 17: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

Transferring Localization Models Across Devices [Zheng and Yang et al., AAAI2008a]

Input: Input:

Output: Output: The localization model on the target device

17

S=(-30dbm, .., -86dbm), L=(1, 3)S=(-33dbm, .., -90dbm), L=(1, 4)…S=(-44dbm, .., -43dbm), L=(9, 10)S=(-56dbm, .., -32dbm), L=(15, 22)S=(-60dbm, .., -29dbm), L=(17, 24)

S=(-37dbm, .., -77dbm), L=(1, 3)S=(-41dbm, .., -83dbm), L=(1, 4)…S=(-49dbm, .., -34dbm), L=(9, 10)S=(-61dbm, .., -28dbm), L=(15,22)S=(-66dbm, .., -26dbm), L=(17, 24)

S=(-33dbm, .., -82dbm), L=(1, 3)…S=(-57dbm, .., -63dbm), L=(10, 23)

Source devices have plentiful labeled dataTarget device has onlyfew labeled data

D-Link Buffalo CISCO

Page 18: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

Transferring Localization Models Across Devices [Zheng and Yang et al., AAAI2008a]

Model: Model: Latent Multi-Task Learning [Caruana, MLJ1997]

Each device: a learning task minimize its localization error, and devices share some common constraints

in a latent space

Regression with signals x to locations y

18

Localization on each wireless adapter is treated as a learning task.

0, ( ) ,t t t t ty w x b w w v sharedshared

Page 19: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

Transferring Localization Models Over Time [Zheng and Yang et al., AAAI2008b]

20

S=(-30dbm, .., -86dbm), L=(1, 3)

S=(-44dbm, .., -43dbm)L=(9, 10)

S=(-60dbm, .., -29dbm)L=(17, 24)

S=(-33dbm, .., -82dbm), L=(1, 3)…S=(-57dbm, .., -63dbm), L=(10, 23)

S=(-42dbm, .., -77dbm)

S=(-43dbm, .., -52dbm)

S=(-71dbm, .., -33dbm)

S=(-49dbm, .., -41dbm)L=(1, 3)

Input: Input: The old time periodPlentiful labeled sequences:

The new time periodSome (non-sequential) labeled data + some unlabeled sequences

Output: Output: Localization model for the new time period.

PhD Student Vincent Zheng @ Work

Page 20: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

Transferring Localization Models Over Time [Zheng and Yang et al., AAAI2008b]

Model: Model: ◦ Transferred Hidden Markov Model

21

Reference points (RPs)

Radio map

Transition matrix of user moves

Prior knowledge on the likelihood of where the user is

Transfer No-transfer

Accuracy under 3m error distance 85% 73%

Page 21: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

Transferring Localization Models Across Space [S. J. Pan and Yang et al., AAAI2008]

22

Input: Input:

Output: Output: Localization model for Area B

B

AAccess Point

Area B:

Few labeled data &

Some unlabeled dataArea A:Plentiful labeled data(red dots in the picture)

Page 22: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

Summary: Localization using Sensors

• Research Issues– Optimal Sensor

Placement [Krause, Guestrin, Gupta, Kleinberg, IPSN2006]

– Integrated Propagation and learning models

– Sensor Fusion– Transfer Learning– Location-based social

networks

24

Locations◦ 2D / 3D Physical Positions◦ Locations are a type of context

Other contextual Information◦ Object Context: Nearby objects +

usage statusLocations and Context

Where you areWho’s around youHow long you are thereStatus of objects (door open?)What is the temperature like?

Page 23: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

Activity Recognition• Action Model Learning

– How do we explicitly model the user’s possible actions?

• Activity Recognition– What is the user

doing / trying to do?• Localization and context

– Where is the user?– What’s around her?– How long/duration?– What time/day?

25

Events

Page 24: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

Steps in activity recognition

26

ActionRecognition

sensorsensor sensor sensor

Loc/ContextRecognition

GoalRecognition

• Also,– Plan, Behavior, Intent, Project …

Page 25: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

Activity Recognition: Input & Output

• Input– Context and locations

• Time, history, current/previous locations, duration, speed, • Object Usage Information

• Trained AR Model• Training data from calibration• Calibration Tool: VTrack

• Output:– Predicted Activity Labels

• Running?• Walking?• Tooth brushing?• Having lunch?

27

http://www.cse.ust.hk/~vincentz/Vtrack.html

Page 26: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

Activity Recognition: Applications• GPS based Location-based services

– Inferring Transportation Modes/Routines• [Liao, Fox, Kautz, AAAI2004]

– Unsupervised, bridges the gap between raw GPS and user’s mode of transportation

– Can detect when user missed bus stops offer help• Healthcare for elders– Example: The Autominder System– [Pollack, et al. Robotics and Autonomous Systems,

2003.]– Provide users w/ reminders when they need them

• Recognizing Activities with Cell Phones (Video)– Chinese Academy of Sciences (Prof Yiqiang Chen and

Dr. Junfa Liu)

28

Page 27: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

Microsoft Research Asia: GeoLife Project [Zheng, Xie, WWW2008]

• Inferring Transportation Modes, and • Compute similarity based on itineraries and link

people in a social net: GeoLife Video

29

Segment[i-1]: Car Segment[i]: Walk Segment[i+1]: Bike

P(Car): 75%P(Bus): 10%P(Bike): 8%P(Walk): 7%

P(Bike): 62%P(Walk): 24%P(Bus): 8%P(Car): 6%

P(Bike): 40%P(Walk): 30%P(Bus): 20%P(Car): 10%

Segment[i].P(Bike) = Segment[i].P(Bike) * P(Bike|Car)

Segment[i].P(Walk) = Segment[i].P(Walk) * P(Walk|Car)

Page 28: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

Activity Recognition (AR): ADL

• ADL = Activities of daily living (ADLs) • From sound to events, in everyday life

• [Lu and Choudhury et al., MobiSys2009]

• iCare (NTU): Digital home support, early diagnosis of behavior changes

• iCare Project at NTU (Hao-hua Chu, Jane Hsu, et al.) http://mll.csie.ntu.edu.tw/icare/index.php

• Duration patterns and inherent hierarchical structures

• [Duong, Bui et al., AI Journal 2008]

30

Page 29: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

Early Work: Plan Recognition

• Objective [Kautz 1987]:– Inferring plans of an agent from (partial)

observations of his actions– Input:• Observed Actions (K,L)• Plan Library

– Output:• Recognized Goals/Plans

31

Page 30: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

Review: Event Hierarchy in Plan Recognition

• The Cooking Event Hierarchy [Kautz 1987]

• Some works:– [Kautz 1987]: graph

inference– [Pynadath and

Wellman, UAI2000]: probabilistic CFG

– [Geib and Steedman, IJCAI2007]: NLP and PR

– [Geib, ICAPS2008]: string rewriting techniques

32

Abstraction relationshipAbstraction relationship

ActionsActions

Step 2 of Make

Pasta Dish

Step 2 of Make

Pasta Dish

Page 31: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

A Gap?

33

Page 32: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

AR: Sequential Methods• Dynamic Bayesian Networks

– [Liao, Fox, Kautz, AAAI2004] [Yin, Chai, Yang, AAAI2004]

• Conditional Random Field [Vail and Veloso, AAAI2008]• Relational Markov Network [Liao, Fox, Kautz, NIPS2005]

34

Page 33: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

Intel [Wyatt, Philipose, Choudhury, AAAI2005] : Incorporating Commonsense

• Model = Commonsense Knowledge– Work at Intel Seattle Lab /

UW– Calculate Object Usage

Information from Web Data P(Obj | Action)

– Train a customized model• HMM: parameter learning

[Wyatt et al. AAAI2005]• Mine model from Web

[Perkowitz, Philipose et al. WWW2004]

35

Page 34: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

Datasets: MIT PlaceLab http://architecture.mit.edu/house_n/placelab.html

• MIT PlaceLab Dataset (PLIA2) [Intille et al. Pervasive 2005]

• Activities: Common household activities

36

Page 35: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

Datasets: Intel Research Lab

• Intel Research Lab [Patterson, Fox, Kautz, Philipose, ISWC2005]– Activities Performed:

11 activities– Sensors

• RFID Readers & Tags

– Length:• 10 mornings

37Picture excerpted from [Patterson, Fox, Kautz, Philipose, ISWC2005].

Now: Intel has better RFID wristbands.

Page 36: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

Complex Actions? Reduce Labels?

Complex Actions: For multiple activities with complex relationships

[Hu and Yang, AAAI2008]

◦ concurrent and interleaving activitiesLabel Reduction:

What if we are short of labeled data in a new domain? [Zheng, Hu, Yang, et al. Ubicomp 2009]◦Use transfer learning to borrow knowledge from a source

domain (where labeled data are abundant)◦ For recognizing activities where labeled data are scarce

38

Page 37: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

Concurrent and Interleaving Goals [Hu, Yang, AAAI2008]

39

Concurrent Activities

Concurrent Activities

Interleaving Activities

Interleaving Activities

Page 38: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

Concurrent and Interleaving Goal and Activity Recognition [Hu, Yang, AAAI2008]

40

Use the long-distance dependencies in Skip-Chain Conditional Random Fields to capture the relatedness between interleaving activities.

Factors for linear chain edges

Factors for linear chain edges

Factors for skip edges

Page 39: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

Concurrent and Interleaving Goal and Activity Recognition [Hu, Yang, AAAI2008]

Our Approach Only Concurrent Only Interleaving

MIT PlaceLab Dataset

86% 73% 80%

41

1

0.32 1

0.93 0.27 1

0.48 0.13 0.72 1

S

Concurrent Goals:• correlation matrix between different goals learned from training data

Example: “attending invited talk” and “browsing WWW”.

Page 40: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

Cross Domain Activity Recognition [Zheng, Hu, Yang, Ubicomp 2009]

• Challenges:– A new domain of

activities without labeled data

• Cross-domain activity recognition– Transfer some available

labeled data from source activities to help training the recognizer for the target activities.

42

CleaningIndoor

Laundry

Dishwashing

Page 41: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

Calculating Activity Similarities

How similar are two activities?◦Use Web search results◦ TFIDF: Traditional IR

similarity metrics (cosine similarity)

◦ Example Mined similarity between

the activity “sweeping” and “vacuuming”, “making the bed”, “gardening”

Calculated Similarity with the activity "Sweeping"

Similarity with the activity "Sweeping"

43

Page 42: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

How to use the similarities?

44

Source Domain Labeled Data

Source Domain Labeled Data

Similarity MeasureSimilarity Measure

<Sensor Reading, Activity Name>

Example: <SS, “Make Coffee”>

<Sensor Reading, Activity Name>

Example: <SS, “Make Coffee”>

Example: sim(“Make

Coffee”, “Make Tea”) = 0.6

Example: sim(“Make

Coffee”, “Make Tea”) = 0.6

Example: Pseudo Training Data: <SS, “Make Tea”, 0.6>

Target Domain Pseudo Labeled

Data

Target Domain Pseudo Labeled

Data

Weighted SVM Classifier

Weighted SVM Classifier

THE WEB

Page 43: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

Cross-Domain AR: PerformanceMean Accuracy with Cross Domain Transfer

# Activities (Source Domain)

# Activities (Target Domain)

Baseline (Random Guess)

MIT Dataset (Cleaning to Laundry)

58.9% 13 8 12.5%

MIT Dataset (Cleaning to Dishwashing)

53.2% 13 7 14.3%

Intel Research Lab Dataset

63.2% 5 6 16.7%

45

Activities in the source domain and the target domain are generated from ten random trials, mean accuracies are reported.

Page 44: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

How Does AR Impact AI?

• Action Model Learning– How do we explicitly

model the user’s possible actions?

• Activity Recognition– What is the user doing /

trying to do?• Localization– Where is the user?

46

Page 45: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

Relationship to Localization and AR

• From context– state description from

sensors• From activity recognition

– activity sequences

• Learning action models

• Motivation:– solve new planning

problems– knowledge-engineering

effort– for Planning

• Can even recognize goals using planning

• [Ramirez and Geffner, IJCAI2009]

47

Page 46: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

What is action model learning?Input: activity sequences

◦ Sequences of labels/objects: Example: pick-up(b1) ,

stack(b1,b2)…etc◦ Initial state, goal, and partial

intermediate states Example: ontable(b1),clear(b1), …

etcOutput: Action models

◦ preconditions of actions: Example: preconditions of “pick-

up”: ontable(?x) , handempty, …etc.

◦ effects of actions: Example: effects of “pick-up”:

holding(?x), …etc

TRAIL [Benson, ICML1994]: learns Teleo-operator models (TOP) with domain experts’ help.

EXPO [Gil, ICML1994]: learns action models incrementally by assuming partial action models known.

Probabilistic STRIPS-like models [Pasula et al. ICAPS2004]: learns probabilistic STRIPS-like operators from examples.

SLAF [Amir, AAAI2005]: learns exact action models in partially observable domains.

48

Page 47: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

ARMS [Yang et al. AIJ2007]An overview

Build constraintsBuild constraints

Activity SequencesActivity SequencesSensor states, object usage

Sensor states, object usage

Information constraintsInformation constraints

Plan constraintsPlan constraintsSolved w/ Weighted MAXSAT/MLNSolved w/ Weighted MAXSAT/MLN

Action modelsAction models

•what can be in the preconditions/Postcond•what can be in the preconditions/Postcond

Each relation has a weight that can be learned

Each relation has a weight that can be learned

49

Page 48: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

Evaluation: by Students @ HKUST

execute learned actions Lego-Learning-Planning (LLP) System Design

50

Control Command

Robot Status/ Data

Internet

Activity recognition & planningRobot PDA

Web Server

Notebook

Bluetooth

Page 49: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

A Lego Planning DomainRelations

given by sensors/phy. map◦ (motor_speed ) ◦ (empty ) ◦…◦ (across x-loc y-loc z-loc)

ActionsKnown to the robot◦ (Move_forw x-loc y-loc z-loc)◦…◦ (Turn_left x-loc y-loc z-loc) ◦

Initial state: ◦ (empty ) (face grid0)…

Goal: ◦…(holding Ball)

Collection of Activity Sequences (Video 1: robot) (video 2: human)

51

Page 50: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

Activity SequencesHuman manually achieves

goal◦ 0: (MOVE_FORW A B C) ◦ …◦ 4: (MOVE_FORW D E F)◦ 5: (MOVE_FORW E F W)◦ 6: (STOP F)◦ 7: (PICK_UP F BALL)◦ …◦ 10: (STOP D)◦ 11: (TURN_LEFT D W E)◦ 12: (PUT_DOWN BALL D)◦ 13: (PICK_UP D BALL)

52

Activity Recognizer

ARMS: Action Model Learning

Page 51: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

Learned Action Models•(:action Stop:parameters (?x - loc):precondition (and (motor_speed) (is_at ?x) ):effect (and (empty) (face ?x) (not

(motor_speed)) ))•…

53This is an error

Dr. Hankz Hankui Zhuo @ HKUST

Page 52: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

LLP Solves a new Lego Planning Problem

• (:init (empty) (face B) (is_at A) (at ball F) …(across C D W) (across D E F) (across E F W) )

• (:goal (and (is_at F) (holding Ball) )

54

F

Goal

A B C D

E

F

W

Init

Ball

Lego

Ball

Page 53: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

Solving the Planning Problem

Generate a plan using a planner◦ 0: (MOVE_FORW A B C) ◦ …◦ 4: (MOVE_FORW D E F)◦ 5: (MOVE_FORW E F W)◦ 6: (STOP F)◦ 7: (PICK_UP F BALL)◦ …◦ 10: (STOP D)◦ 11: (TURN_LEFT D W E)◦ 11: (PUT_DOWN BALL D)◦ 12: (PICK_UP D BALL) ◦ 13: (MOVE_FORW D E F) ◦ 14: (MOVE_FORW E F W)

Execution◦ sometimes it will◦ Succeed!!!◦ But, sometimes it will◦ Fail…

More feedback on learning required

55

Page 54: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

Closing the Loop in the Knowledge Food Chain

Close the feedback loopLoop

1. Signal traces collected2. Location, Context, Activities

predicted 3. Action models learned4. New plan generated and

executed5. Errors found6. Human intervention to

correct plans7. New Plans

End loop

56

Page 55: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

Activity Recognition: Truly Multidisciplinary

57

•Computer Networks•Pervasive/Ubiquitous Computing•Logistics•Intelligent Transportation•Urban Design and Planning

•Health Informatics/Public Health•Mobile Commerce/Services•Mobile Social Nets•Geographical Information Sys•HCI, Data Mining

Software EngineeringComputer GraphicsHCI

Page 56: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

Open issues in Activity Recognition

• User privacy– [Klasjna, Consolvo,

Choudhury, et al., Pervasive2009]

• False Positives– Cost-sensitive appl.

• Market study– Business Models?

• Many users– mobile social networks– Multi-person AR

(cooperation? Competition?)

• Transfer Learning– Between users– Between activities– Between different types

of sensors

58

Page 57: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

ConclusionsFuture– Cheaper and more

ubiquitous sensors will bring a new era for AI through activity recognition research and applications

• Acknowledgement …

Theme of the Talk• Activity Recognition:– What it is– Linking low level sensors

to high level intelligence– Closed loop

• Activity recognition as Embedded AI– Empirical in nature– Research on a very

limited budget

59

Page 58: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

Theme of The Talk

60

Page 59: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

Acknowledgement

• Students (Former) @ HKUST– Vincent W. Zheng, Derek H. Hu, Hankz H. Zhuo,

Sinno J. Pan– Jie Yin, Dou Shen, Jeffrey J. Pan, Rong Pan

• Collaborators– Drs. Junhui Zhao and Yongcai Wang (NEC China

Lab)– Drs. Yiqiang Chen and Junfa Liu (CAS)– Drs. Xing Xie and Yu Zheng (MSRA)

61

Page 60: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

References (Localization)• [Bahl and Padmanabhan, CCC2000] RADAR: An in-building RF-based user location and

tracking system.• [Caruna, MLJ1997] Multi-task Learning.• [Ferris, Fox and Lawrence, IJCAI2007] WiFi-slam using Gaussian process latent variable

models.• [Fox and Hightower et al., PervasiveComputing2003] Bayesian filtering for location

estimation.• [Krause, Guestrin, Gupta, Kleinberg, IPSN2006] Near-optimal sensor placements: maximizing

information while minimizing communication cost. • [Ladd et al., MobiCom2002] Robotics-based Location Sensing using Wireless Ethernet.• [Ni et al., PerCom 2003] LANDMARC: indoor location sensing using active RFID.• [J. J. Pan and Yang et al., IJCAI2005] Accurate and low-cost location estimation using kernels.• [J. J. Pan and Yang et al., AAAI2006] A Manifold Regularization Approach to Calibration

Reduction for Sensor-Network Based Tracking.• [R. Pan, Zheng, Yang et al., KDD2007] Domain-Constrained Semi-Supervised Mining of

Tracking Models in Sensor Networks.• [S. J. Pan and Yang et al., AAAI2008] Transferring Localization Models Across Space.

62

Page 61: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

References (Localization)• [Stunteback and Patel et al., Ubicomp2008] Wideband powerline positioning for indoor

localization.• [Zheng and Yang et al., AAAI2008a] Transferring Multi-device Localization Models using

Latent Multi-task Learning.• [Zheng and Yang et al., AAAI2008b] Transferring Localization Models Over Time.

63

Page 62: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

References (Activity Recognition) [Bao and Intille, Pervasive 2004] Activity Recognition from User-Annotated Acceleration

Data. [Bui et al., AAAI 2008] The Hidden Permutation Model and Location-Based Activity

Recognition. [Chian and Hsu, IJCAI2009] Probabilistic Models for Concurrent Chatting Activity

Recognition [Choudhury and Basu, NIPS 2004] Modeling Conversational Dynamics as a Mixed-Memory

Markov Process. [Kautz 1987] A Formal Theory of Plan Recognition. [Geib and Steedman, IJCAI 2007] On Natural Language Processing and Plan Recognition. [Geib et al., ICAPS 2008] A New Probabilistic Plan Recognition Algorithm Based on String

Rewriting. [Hu, Yang, AAAI2008] CIGAR: Concurrent and Interleaving Goal and Activity Recognition. [Klasnja, Consolvo, Choudhury et al. Pervasive2009] Exploring Privacy Concerns about

Personal Sensing. [Liao, Fox, Kautz, AAAI2004] Learning and Inferring Transportation Routines.

64

Page 63: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

References (Activity Recognition) [Liao, Fox, Kautz, NIPS2005] Location-Based Activity Recognition. [Lu and Choudhury et al., MobiSys2009] SoundSense: scalable sound sensing for people-

centric applications on mobile phones. [Patterson, Fox, Kautz, Philipose, ISWC 2005] Fine-Grained Activity Recognition by

Aggregating Abstract Object Usage. [Pollack, 2003] Autominder: an intelligent cognitive orthotic system for people with memory

impairment. [Pynadath and Wellman, UAI 2000] Probabilistic State-Dependent Grammars for Plan

Recognition. [Vail and Veloso, AAAI 2008] Feature Selection for Activity Recognition in Multi-Robot

Domains. [Wyatt, Philipose and Choudhury, AAAI 2005] Unsupervised Activity Recognition Using

Automatically Mined Common Sense. [Yin, Chai, Yang, AAAI2004] High-level Goal Recognition in a Wireless LAN. [Zheng, Hu, Yang, Ubicomp 2009] Cross-Domain Activity Recognition. [Zheng, Xie, WWW 2008] Learning transportation mode from raw GPS data for geographic

applications on the web.

65

Page 64: Activity Recognition:  Linking Low-level Sensors to High-level Intelligence

References (Action Model Learning)

• [Amir, IJCAI2005] Learning Partially Observable Deterministic Action Models.• [Benson, ICML1994] Inductive Learning of Reactive Action Models.• [Gerevini, AIPS2002] A Planner Based on Local Search for Planning Graphs with Action

Costs.• [Gil, ICML1994] Learning by Experimentation: Incremental Refinement of Incomplete

Planning Domains.• [Pasula et al., ICAPS2004] Learning Probabilistic Planning Rules.• [Yang et al., AIJ2007] Learning action models from plan examples using weighted MAX-

SAT.• [Zhuo et al. PAKDD-09] Transfer Learning Action Models by Measuring the Similarity of

Different Domains.

66