a smart walker to understand walking abilities
TRANSCRIPT
1
A Smart Walker to Understand Walking Abilities
Pascal PoupartAssociate Professor
Cheriton School of Computer Science University of Waterloo
Monday, February 15, 2010University of Kentucky
2
Multidisciplinary Team• Computer Science
– Farheen Omar, Mathieu Sinn, Jakub Truskowski, Richard Hu, Allan Caine, Hao Chen, Adam Hartfiel, Adam Fourney, Omar Zia Khan, Jay Black, Richard Mann, Pascal Poupart
• Systems Design Engineering– Samantha Ng, Adel Fakih, John Zelek
• Kinesiology– James Tung, Matt Snyder, Tracy McWhirter, Rhiannon
Rose, Bill McIlroy, Eric Roy,
• Partners– UW Schelgel Research Institute for Aging, Village of
Winston Park, Toronto Rehabilitation Institute, CIHR, NSERC
3
Outline
• Motivation• Smart Walker• User study at the
Village of Winston Park• Behaviour recognition• Limb tracking
4
Disability Statistics• US National Business Service Alliance (July 2006)
5
Disability Statistics• US National Business Service Alliance (July 2006)
Types of disabilities
6
Mobility aids
Encourage walkingIncrease safety
Discourage walking
pp3
Slide 6
pp3 What are the walking aids available to seniors?Pascal Poupart, 11/21/2008
7
Smart Walker
walker devices
+ caregivers
users
Force sensorsLoad sensorsVideo camerasMicrophoneSpeech synthesizerServo-brakesetc.
assist
Smart walker
8
Current Prototype• Designed by James Tung and Bill McIlroy at the
Toronto Rehabilitation Institute
• Monitoring device• Collect information about
– gait (posture, mobility, walking abilities)– walker usage (speed, brakes, load)– environment (obstacles)
• Why?– Detect falls– Monitor users – Ensure safety
9
iWalker (Tung & McIlroy at TRI)
Single-axis load cells (each caster)
Wheel Rotation Counter
(L&R Rear Wheels)
Brake Sensor (x2)
Propulsion forces (x2)
Palm PC with A / D unit
Panoramic hemispherical camera
Battery Unit
Under-Seat Components
2-D accelerometer
Single-axis load cells (each caster)
Wheel Rotation Counter
(L&R Rear Wheels)
Brake Sensor (x2)
Propulsion forces (x2)
Palm PC with A / D unit
Panoramic hemispherical camera
Battery Unit
Under-Seat Components
2-D accelerometervideo camerapointing backward
Video recorder3-D accelerometer
and A / D unitPDA
10
Research questions• How do people use walkers?
– How is balance control affected?– What are the challenges?
• Can we assess walking abilities in a naturalistic environment?– Automated behaviour recognition– Track the 3D pose of lower limbs
• Benefits– Continuous and quantitative monitoring– Improve treatment plans
11
User Study• Village of Winston Park (retirement community)
– 11 walker users (average age: 89) – 10 non-walker users (average age: 83)
• Battery of physical and neuro-psychological tests• Obstacle course
– Straight walking– Turns– Reaching task:
• Press elevator button• Pick up a remote• Open a door
– Go up/down a ramp and curb
12
Prescription of Walkers
• No standardized procedure
• Common clinical measures:– Timed Up and Go (TUG)– Berg Balance Scale (BBS)– Center of pressure variability
13
Timed Up and Go (TUG)• Time to get up, walk 3m, come back and sit
– Normal: < 12 seconds
14
TUG Comparison• Longer time with walker for both groups• Problem: walker appears to decrease
performance
15
Lower Limb Balance Control
16
Upper Limb Balance Control• Center of Pressure (CoP)• Vertical Load (Fz)
17
Center of Pressure Variability
Anterior/PosteriorMedio/Lateral
Hands OFFHands ON
18
Behavioural context and instability triggersInstability triggers:
a: elevator door collisionsb: foot collisions
Behavioural Context:W: straight walkingS: standingR: resting on chair
19
Automated Behaviour Recognition• Segment time by behaviours:
• Manually label behaviours during walking course:– 9 walker users– 8 controls (non-walker users)
1. Not touching walker2. Stop/standing3. Walking forward4. Turning left5. Turning right6. Walking backward
7. Transfer8. Sitting on walker9. Going up ramp10. Going down ramp11. Going up curb12. Going down curb
20
Automated Behaviour Recognition• Problem: infer behaviour from sensor
measurements
• Supervised learning:– Hidden Markov model (HMM)– Conditional random field (CRF)
• Unsupervised learning:– HMM
• Expectation Maximization (EM)• Bayesian Learning
21
Hidden Markov Model (HMM)
• Each sensor:– 50 Hz (50 measurements per second)– 216 values discretized in 20 buckets
• Pr(Oit|Bt) = relative frequency counts
• Pr(Bt+1|Bt) ≈ 0.95 for Bt+1 = Bt
Bt
O1t O2
t O8t…
Bt+1
O1t+1 O2
t+1 O8t+1…
Bt+2
O1t+2 O2
t+2 O8t+2…
22
Conditional Random Field (CRF)
• Features: thresholded average and std– Speed, total load, CoP (m/d, a/p), accel. (x, y, z)– windows of 1, 5, 25 and 50 measurements
• Training: conjugate gradient descent
Bt
F1t F2
t Fnt…
Bt+1
F1t+1 F2
t+1 Fnt+1…
Bt+2
F1t+2 F2
t+2 Fnt+2…
Sensor measurements
23
Behaviour Recognition AccuracyPrediction Accuracy
24
Behaviour Transitions
Precision =# of correct transitions
# of predicted transitionsRecall =
# of correct transitions
# of labelled transitions
TransitionPrecision
TransitionRecall
25
Limb Tracking• Estimate 3D pose• Tapered cylinders
defined by 21 parameters
• HMM: particle filtering– infer pose P from cues Ci
Pt
C1t C2
t Cnt…
Pt+1
C1t+1 C2
t+1 Cnt+1…
Pt+2
C1t+2 C2
t+2 Cnt+2…
26
Appearance Model (Cues)• For each limb segment:
– Texture: histogram of oriented gradients (HOG)– Colour: histogram of colours (RGB)
Ex: histogram of oriented gradients Ex: histogram of colours
27
Uncertainty in Pose Estimate
28
Challenge
Sagittal plane Coronal planeeasy hard
• Motion capture
• Need depth information– Stereo-vision– Time-of-flight camera
29
Conclusion• Smart walker
– Automated behaviour recognition– Limb tracking
• Benefits– Naturalistic gait analysis– Continuous quantitative monitoring
• Next steps– Stereo-vision / time-of-flight camera– Unified model (behaviour recognition and limb tracking)
30
Vision: Robotic Nurse• Navigation assistance
– Obstacle avoidance – Route to some destination
• Personal trainer– Monitor and coach exercises
• Reminder system– Medication, appointments, location
• Alert system– Notify caregiver when in need or fallen
• Conversational companion– Speech recognizer/generator
31
PascalPascal’’s Research Projectss Research Projects• Policy explanation for MDP recommender systems
– Omar Zia Khan (PhD)• Planning as inference for decentralized POMDPs
– Igor Kiselev (PhD)• Bayesian Reinforcement Learning
– Ricardo Salmon (PhD)• Behaviour recognition with a smart walker
– Farheen Omar (PhD), Mathieu Sinn (Postdoc)• Limb tracking with a smart walker
– Richard Hu (MMath)• Automated feature generation in temporal models
– Adam Hartfiel (PhD), Mathieu Sinn (Postdoc)• Voice, Location and Activity Monitoring for Alzheimer
– James Tung (Postdoc)• Unsupervised cluster labelling with HDPs
– Ting Liu (MMath)• Hand-written equation recognition with HDP-HMMs
– Mazen Melibari (MMath)
32
Thank You
Questions?
33
CRF Confusion MatrixNot touching
walkerStanding
/StopWalking Forwards
Turn Left
Turn Right
Walking Backwards
Transfers /Chair
Reaching Tasks
Sitting On Walker
Going Up Ramp
Going Down Ramp
Going Up Curb
Going Down Curb
Predicted Incorrectly
Predicted Correctly
Not touching walker 7916 1265 0 0 0 0 0 0 364 0 0 0 0 1629 7916Standing/Stop 465 44727 1627 914 300 0 0 2483 744 106 17 206 49 6911 44727Walking Forwards 2 2688 60558 2581 2661 0 0 1823 20 578 179 420 78 11030 60558Turn Left 0 1279 4237 10397 523 0 0 238 80 80 19 14 0 6470 10397Turn Right 0 1135 5208 312 7976 0 0 566 14 29 236 65 120 7685 7976Walking Backwards 0 157 230 94 166 0 0 327 0 9 33 29 0 1045 0Transfers/Chair 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0Reaching Tasks 0 5278 1882 911 294 0 0 3876 12 26 8 0 0 8411 3876Sitting On Walker 111 674 67 107 45 0 0 0 70330 0 0 0 0 1004 70330Going Up Ramp 0 154 350 129 8 0 0 52 21 1980 158 137 9 1018 1980Going Down Ramp 0 39 620 0 156 0 0 0 0 0 2517 80 10 905 2517Going Up Curb 0 330 178 20 129 0 0 146 0 85 10 2122 67 965 2122Going Down Curb 0 69 268 0 85 0 0 0 0 0 161 68 821 651 821
Not touching walker
Standing with
Walking Forwards
Turn Left
Turn Right
Walking Backwards
Transfers /Chair
Reaching Tasks
Sitting On Walker
Going Up Ramp
Going Down Ramp
Going Up Curb
Going Down Curb
Predicted Incorrectly
Predicted Correctly
Not touching walker 8514 836 0 0 0 0 0 0 195 0 0 0 0 1031 8514Standing Stop 239 48558 579 431 102 0 0 1352 315 22 17 23 0 3080 48558Walking Forwards 0 1356 67580 959 741 0 0 740 20 16 66 100 10 4008 67580Turn Left 0 710 2191 13531 229 0 0 59 63 52 18 14 0 3336 13531Turn Right 0 702 3342 225 10915 0 0 319 5 15 21 65 52 4746 10915Walking Backwards 0 157 230 94 166 0 0 327 0 9 33 29 0 1045 0Transfers/Chair 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0Reaching Tasks 0 4013 1287 685 185 0 0 6071 12 26 8 0 0 6216 6071Sitting On Walker 71 220 54 107 45 0 0 0 70837 0 0 0 0 497 70837Going Up Ramp 0 87 152 9 8 0 0 52 0 2386 158 137 9 612 2386Going Down Ramp 0 15 216 0 26 0 0 0 0 0 3159 6 0 263 3159Going Up Curb 0 133 37 20 89 0 0 14 0 0 10 2784 0 303 2784Going Down Curb 0 17 168 0 46 0 0 0 0 0 10 19 1212 260 1212
Window with: 0
Window with: 40