Ubiquitous Care System to Support Independent Living : Preliminary Results
Machine Learning (2009)
Available from dis.ijs.si
or
Abstract
The European FP7 project CONFIDENCE - Ubiquitous Care System to Support Independent Living is developing a system that will monitor the health conditions of the elderly in real-time based on radio tags placed on the human body. In case of health problems, the system will issue a warning to the user and an alarm to the caregiver if necessary. This way the elderly should gain the needed con dence and security to continue their independent participation in society, thus reducing costs for medical services and burden to the working age pop- ulation. The paper describes work in progress of the reconstruction and interpretation subsystem and presents encouraging results in complexfall detection scenario.
Available from dis.ijs.si
Page 1
Ubiquitous Care System to Support Independent Living : Preliminary Results
Ubiquitous Care System to Support
Independent Living: Preliminary
Results
Boštjan Kaluža1,
Violeta Mirčevska1, Mitja Luštrek1, Igone Vélez2,
and Matjaž Gams1
1 Department of Intelligent Systems, Jožef Stefan Institute, Slovenia
2 Centro de Estudios e Investigaciones Técnicas de Gipuzkoa, Spain
Independent Living: Preliminary
Results
Boštjan Kaluža1,
Violeta Mirčevska1, Mitja Luštrek1, Igone Vélez2,
and Matjaž Gams1
1 Department of Intelligent Systems, Jožef Stefan Institute, Slovenia
2 Centro de Estudios e Investigaciones Técnicas de Gipuzkoa, Spain
Page 2
Outline
1. The CONFIDENCE project
2. Reconstruction and interpretation
subsystem
3. Evaluation and results
4. Conclusion
1. The CONFIDENCE project
2. Reconstruction and interpretation
subsystem
3. Evaluation and results
4. Conclusion
Page 3
The CONFIDENCE project
• FP 7 project
• 10 partners
• FP 7 project
• 10 partners
Page 4
The CONFIDENCE project
• Primary goal: create care system able to
detect abnormal situations
– Short term, e.g., fall
– Long term, e.g., gait disorders
• When hazardous situation
– Trigger alarm
– Call relatives or help center
• Primary goal: create care system able to
detect abnormal situations
– Short term, e.g., fall
– Long term, e.g., gait disorders
• When hazardous situation
– Trigger alarm
– Call relatives or help center
Page 5
The CONFIDENCE hardware
• Equipment
– Small body-worn tags
• Wireless
• 3D coordinates
– Central device
• Equipment
– Small body-worn tags
• Wireless
• 3D coordinates
– Central device
Page 6
The CONFIDENCE project
• General structure of the system
Localization
Subsystem
Reconstruction
and Interpretation
Subsystem
System Interface
Subsystem
• General structure of the system
Localization
Subsystem
Reconstruction
and Interpretation
Subsystem
System Interface
Subsystem
Page 7
The CONFIDENCE project
• General structure of the system
Localization
Subsystem
Reconstruction
and Interpretation
Subsystem
System Interface
Subsystem
• General structure of the system
Localization
Subsystem
Reconstruction
and Interpretation
Subsystem
System Interface
Subsystem
Page 8
Reconstruction and Interpretation Subsystem
(RIS)
• Main tasks:
– Posture reconstruction
– Activity recognition
– Fall detection
– Disability/disease
detection
V. Mirčevska and B. Kaluža. Towards Intelligent Home Caregiver. In Proceedings of the 1st Jožef Stefan International
Postgraduate School Student’s Conference, Ljubljana, Slovenia, May 2009.
(RIS)
• Main tasks:
– Posture reconstruction
– Activity recognition
– Fall detection
– Disability/disease
detection
V. Mirčevska and B. Kaluža. Towards Intelligent Home Caregiver. In Proceedings of the 1st Jožef Stefan International
Postgraduate School Student’s Conference, Ljubljana, Slovenia, May 2009.
Page 9
RIS: Architecture
Page 10
RIS: Preprocessing
• Noisy data
• Three steps:
– Median filter
– Anatomic
constraints
– Kalman filtering
B. Kaluža and E. Dovgan. Glajenje trajektorij gibanja človeškega telesa zajetih z radijsko tehnologijo.
IS 2009, Ljubljana, Slovenia, October 2008.
• Noisy data
• Three steps:
– Median filter
– Anatomic
constraints
– Kalman filtering
B. Kaluža and E. Dovgan. Glajenje trajektorij gibanja človeškega telesa zajetih z radijsko tehnologijo.
IS 2009, Ljubljana, Slovenia, October 2008.
Page 11
RIS: Attribute Computation
Reference coordinate system
• z coordinates
• absolute velocities, z velocities
• absolute distances between tags,
z distances
Reference coordinate system
• z coordinates
• absolute velocities, z velocities
• absolute distances between tags,
z distances
Page 12
RIS: Attribute Computation
Reference coordinate system
Body coordinate system
• z coordinates
• absolute velocities, z velocities
• absolute distances between tags,
z distances
• x, y, z coordinates
• absolute velocities,
x, y, z velocities
Reference coordinate system
Body coordinate system
• z coordinates
• absolute velocities, z velocities
• absolute distances between tags,
z distances
• x, y, z coordinates
• absolute velocities,
x, y, z velocities
Page 13
RIS: Attribute Computation
• z coordinates
• absolute velocities, z velocities
• absolute distances between tags,
z distances
• x, y, z coordinates
• absolute velocities,
x, y, z velocities
Reference coordinate system
Body coordinate system
Angles
• z coordinates
• absolute velocities, z velocities
• absolute distances between tags,
z distances
• x, y, z coordinates
• absolute velocities,
x, y, z velocities
Reference coordinate system
Body coordinate system
Angles
Page 14
RIS: Activity Recognition
• Two approaches:
– Machine learning
– Expert rules
• Two approaches:
– Machine learning
– Expert rules
Page 15
RIS: Activity Recognition with Machine
Learning
t
Snapshot
Snapshot
Learning
t
Snapshot
Snapshot
Page 16
RIS: Activity Recognition with Machine
Learning
t-1 t t-2 t-9 ...
Activity
Feature vector
Learning
t-1 t t-2 t-9 ...
Activity
Feature vector
Page 17
RIS: Activity Recognition with Machine
Learning
t-1 t t-2 t-9 ...
Activity
t+1
t+2
t+3
...
Learning
t-1 t t-2 t-9 ...
Activity
t+1
t+2
t+3
...
Page 18
RIS: Activity Recognition with Machine
Learning
t-1 t t-2 t-9 ...
Activity
t+1
t+2
t+3
...
Feed machine learning algorithm
M. Luštrek and B. Kaluža. Fall Detection and Activity Recognition with Machine Learning. Informatica, 33(2):197-204, 2009.
Learning
t-1 t t-2 t-9 ...
Activity
t+1
t+2
t+3
...
Feed machine learning algorithm
M. Luštrek and B. Kaluža. Fall Detection and Activity Recognition with Machine Learning. Informatica, 33(2):197-204, 2009.
Page 19
RIS: Activity Recognition with Machine
Learning
M. Luštrek, B. Kaluža, E. Dovgan, B. Pogorelc and M. Gams. Behavior Analysis Based on Coordinates of Body Tags. AmI 2009.
Learning
M. Luštrek, B. Kaluža, E. Dovgan, B. Pogorelc and M. Gams. Behavior Analysis Based on Coordinates of Body Tags. AmI 2009.
Page 20
RIS: Activity Recognition Example
No noise: sitting down Noise: falling
No noise: sitting down Noise: falling
Page 21
RIS: Activity Recognition with Expert Rules
• Common sense rules
• Rule induction from decision trees
IF
DistZ(ankle,
chest)
<
0.3
AND
DistXY(ankle,
chest)
>
1.2
AND
ABS(Vz(chest)
)
<
0.02
THEN
LYING
IF
DistZ(ankle,
chest)
<
0.2
AND
ABS(Vz(chest))
<
0.02
THEN
LYING
IF
DistZ(ankle,
chest)
>
1.2
AND
DistXY(ankle,
chest)
<
0.35
AND
Vz(chest)
>
-‐0.07
THEN
STANDING
IF
Vz(chest)
<
-‐0.15
THEN
FALLING
V. Mirčevska, M. Luštrek, M. Gams. Combining Machine Learning and Expert Knowledge for Classifying Human Posture.ERK 2009, September, Portoroz-Slovenia.
V. Mircevska, M.Lustrek, I.Velez, N. González, M.Gams. Classifying Posture Based on Location of Radio Tags. AMIF-Ambient Intelligence Forum-Czech Republic, September, 2009.
• Common sense rules
• Rule induction from decision trees
IF
DistZ(ankle,
chest)
<
0.3
AND
DistXY(ankle,
chest)
>
1.2
AND
ABS(Vz(chest)
)
<
0.02
THEN
LYING
IF
DistZ(ankle,
chest)
<
0.2
AND
ABS(Vz(chest))
<
0.02
THEN
LYING
IF
DistZ(ankle,
chest)
>
1.2
AND
DistXY(ankle,
chest)
<
0.35
AND
Vz(chest)
>
-‐0.07
THEN
STANDING
IF
Vz(chest)
<
-‐0.15
THEN
FALLING
V. Mirčevska, M. Luštrek, M. Gams. Combining Machine Learning and Expert Knowledge for Classifying Human Posture.ERK 2009, September, Portoroz-Slovenia.
V. Mircevska, M.Lustrek, I.Velez, N. González, M.Gams. Classifying Posture Based on Location of Radio Tags. AMIF-Ambient Intelligence Forum-Czech Republic, September, 2009.
Page 22
RIS: Detection of Critical Situations
• Arise from sudden health problems
– Tripping
– Fainting
– Falling from chair
• Reflected by activity sequence
• Implemented with expert rules
• Arise from sudden health problems
– Tripping
– Fainting
– Falling from chair
• Reflected by activity sequence
• Implemented with expert rules
Page 23
RIS: Detection of Disability/Disease Detection
• Illness reflected in daily life activities
• Statistical methods
– Monitor walking characteristics
– Behavioral changes
• Illness reflected in daily life activities
• Statistical methods
– Monitor walking characteristics
– Behavioral changes
Page 24
Evaluation
• 4 tags attached to person
– Belt, chest, left/right ankle
– Ubisens RTLS
• Complex scenario including
– Three falls
– Three false alarms
• Data
– 5 people
– Each preformed scenario 5 times
• Evaluation: leave-one-person-out
• 4 tags attached to person
– Belt, chest, left/right ankle
– Ubisens RTLS
• Complex scenario including
– Three falls
– Three false alarms
• Data
– 5 people
– Each preformed scenario 5 times
• Evaluation: leave-one-person-out
Page 25
Evaluation
Youtube: http://www.youtube.com/watch?v=r9gSUn9RPgk or keywords confence prototype.
Youtube: http://www.youtube.com/watch?v=r9gSUn9RPgk or keywords confence prototype.
Page 26
Results
Case Machine learning
accuracy [%]
Expert rules accuracy
[%]
Tripping 100 96
Fainting standing 88 92
Sliding from the chair 56 40
Jumping in bed 100 100
Sitting down quickly 100 96
Searching under a table/bed 96 72
Case Machine learning
accuracy [%]
Expert rules accuracy
[%]
Tripping 100 96
Fainting standing 88 92
Sliding from the chair 56 40
Jumping in bed 100 100
Sitting down quickly 100 96
Searching under a table/bed 96 72
Page 27
Conclusion
• Aim:
Prolong independency of the elderly
– Detecting falls, health problems
• RIS
– Several modules in pipeline
– Machine learning + expert knowledge
• Preliminary results promising
– Detecting complex fall scenarios
• Aim:
Prolong independency of the elderly
– Detecting falls, health problems
• RIS
– Several modules in pipeline
– Machine learning + expert knowledge
• Preliminary results promising
– Detecting complex fall scenarios
Sign up today - FREE
Mendeley saves you time finding and organizing research. Learn more
- All your research in one place
- Add and import papers easily
- Access it anywhere, anytime
Start using Mendeley in seconds!
Readership Statistics
2 Readers on Mendeley
by Discipline
by Academic Status
100% Ph.D. Student
by Country
50% Slovenia
50% United States