3. … no, seriously
healthcare applications
is a monitoring device on the body?
Where is it? (pick it up before you go)
Supporting elderly and cognitive impaired
people
MonAMI 3
12 million Euro budget over 4 years
4. derived requirements
infrastructure-less
cheap/simple sensors
symbolic location interesting
the two most likely ‘hits’ help
4
5. approach
A mobile phone ringing or
vibrating sounds differently
depending on where it is. 5
6. approach
-abstraction-
mechanical stimuli
in this case:
– narrow frequency ‘beeps’ (high frequency)
– vibration (low frequency)
analysis
high frequency stimulus
-high frequency response (over microphone)
low frequency stimulus
– low frequency response (over accelerometer)
– high frequency response (over microphone)
6
7. low frequency stimulus
-vibration-
vibration acceleration:
coupled directly to surface
absorption <=> resonance
information:
fixed vs. free
hard vs. elastic
vibration sound:
sound of device
hitting surface
depends not only on
surface,
7
but on overall structure
8. High frequency stimulus
-sound beeps-
structure specific
closed vs. open
material specific
absorption
well understood
in construction and music
8
Table from Olson, H.: Music, Physics and Engineering. (1967)
9. applying the approach
two distinct modes:
specific location mode
+ exact location information
- need for training data
abstract location class
+ training problem avoided
- only fuzzy location information
9
10. issues to consider
• microphone and speaker placement
– speakers and mics are cheap
• variations inside a symbolic location
– though luck
• number of relevant locations
– Room-level location (RF)
• sensor requirements
– cheap sensors sufficient
• complexity
– procedure performed seldom
10
11. recognition method
-features used-
From over 40 features calculated the following 10 are used:
• zero crossing rate
• median
• variance
• 75% percentile
• inter quartile range
• root mean square
• frequency range power
• sums power wavelet determinant coefficient
• number of peaks
• peak height
sound fingerprint vibration sound vibration acceleartion 11
12. recognition method
• sliding window feature extraction
– Over 30 standard features calculated, 10 used
• 2 frequency features
• separate classifiers for each stimulus response
– C 4.5, Naïve Bayes, KNN, HMMs etc.
– comparable results
fp sound vib sound vib accel
• fusion techniques: extract features extract features
extract features
sliding window sliding window
sliding window
– majority decision
classification classification
classification
– lookup table (using Naïve Bayes) (using Naïve Bayes)
(using Naïve Bayes)
best two classifications from each
lookup table
(created by
training data)
12
Result
13. fusion
fp sound vib sound vib accel
extract features extract features
extract features
sliding window sliding window
sliding window
classification classification
classification
(using Naïve Bayes)
(using Naïve Bayes) (using Naïve Bayes)
best two classifications from each
lookup table
(created by training data)
Result
13
14. experiments
data acquisition:
Nokia 5500 Sport
recognition method: batch processing
2 distinct experimental setups:
specific location scenarios:
office, living room, apartment
abstract location class driven data collection:
furniture store 14
15. scenarios 30 samples per location
10 for training 20 for testing
living room office apartment 15
9 locations 12 locations 11 locations
21. frequency range power
and
sums power wavelet
determinant coefficient
parameters adjusted for
each frequency ‘beep’
feature vectors
…
rms1 frp 500 wavelet 500
rms1 frp 500 wavelet 500
21