Context Modulation of Sensor Data Applied to Activity Recognition in - - PowerPoint PPT Presentation

▶

Aug 29, 2023 210 likes •481 views

Context Modulation of Sensor Data Applied to Activity Recognition in Smart Homes LMCE @ ECML 19-th September 2014 Niall Twomey Peter Flach SPHERE S ensing P latform for HE althcare in R esidential E nvironments Long-term goals Emergency

SLIDE 1

Context Modulation of Sensor Data Applied to Activity Recognition in Smart Homes LMCE @ ECML

19-th September 2014

Niall Twomey Peter Flach

SLIDE 2

SPHERE

Sensing Platform for HEalthcare in Residential Environments

Long-term goals

Emergency detection
Detection of health deterioration
Rehabilitation assessment
etc.

http://www.irc-sphere.ac.uk

SLIDE 3

SPHERE Many research areas

Ethics
Privacy
User acceptance
User centred design
Video, environment, on-body

sensors

Healthcare Ecosystem

Equip home with sensors
Learn models to quantify the

goals

Big, complex data

Eventually

SLIDE 4

SPHERE

Machine learning, data mining,

data fusion etc.

SPHERE data

○ Video ○ On-body sensors ○ Environment sensing ○ … Key idea

Long-term trends in activity may

tell us something about health

Learn activity recognition models

with SPHERE data Don’t have SPHERE data yet...

Using surrogate datasets

SLIDE 5

CASAS

Centre for Advanced Studies in Adaptive Systems

Approx 20 datasets
Variety of sensors

○ Motion ○ Door ○ Light ○ Temperature ○ Power ○ ...

Activities of daily living
Multiple inhabitants

http://ailab.wsu.edu/casas/

SLIDE 6

Sensor data file

SLIDE 7

SPHERE: 100 Homes Deployment

Difficulties

○ 100 different houses ○ Over 100 residents ○ Over 100 habits/behaviours

Want to

○ Reuse resources/models ○ Learn global “lessons”

Question:

What can we learn from the CASAS datasets now to help us with our deployment later?

Not necessarily as models, but

knowledge about domain

SLIDE 8

Context Modulation

Context modulation is: “...the act of modifying and adapting datasets in such a way that new topological representations of the original dataset are obtained...” Context modulation yields: “...new representations are illustrative of original (user- defined) contexts...”

Can we learn anything about the problem from the performance of models learnt from modulated data?

SLIDE 9

Original context: 80 sensors approx

Example: Topological Overview of Different Sensing Contexts

SLIDE 10

Original context: 80 sensors approx Context 1: 50 sensors approx

Example: Topological Overview of Different Sensing Contexts

SLIDE 11

Original context: 80 sensors approx Context 1: 50 sensors approx Context 2: 7 sensors approx

Example: Topological Overview of Different Sensing Contexts

SLIDE 12

Why? Sensors

Simulate breaking sensors

(corrupt dataset)

Replace 6 sensors with one of

different capabilities

Effect of blind spots
Optimise sensor distribution

throughout home (density/cost)

Translation

Labels

Multi-resident environments

activities are similar; sharing variables

Represent database as a

multitask problem

Transfer/multi-task

SLIDE 13

Experiments Performed

CASAS twor.2009 dataset

Presented here:

Assess effect of anonymising activity labels

Multi-resident environment
Labels specify the resident

Example of this in next slide

Presented in the paper:

Remove door/light sensors
Anonymous labels + remove

door/light sensors

Room-level motion sensing
Anonymous labels + room-level

motion sensing

SLIDE 14

Sensor data file - ambiguity due to overlapping labels, multi class predictions

SLIDE 15

Learning Algorithm

Conditional Random Field (CRF) Some properties of CRFs

Sequence classifier
Learn conditional distribution
Calibrated probability estimates
Handles imbalanced data well

Cross validation for parameter selection

SLIDE 16

Performance assessment

Surprisingly ambiguous

"Precision":

“...the probability that a predicted label of the i-th instance is in the set

f candidates...”

Bigger is better

"False Negative Rate":

“...the probability of never classifying an activity once over its entire life cycle...” Lower is better

SLIDE 17

Results

SLIDE 18

Ground truth; multilabel ground truth; multiclass predictions

SLIDE 19

Co-occurring activities

SLIDE 20

Quick-changing predictions

SLIDE 21

False positive prediction

SLIDE 22

False negative prediction (study)

SLIDE 23

Overall results - show expected trends; still interesting

SLIDE 24

Simulate broken sensors
Reform dataset to multi-

task/hierarchical classification problem

Reform label space

What else can we do?

Learn about importance of

sensors

Discover redundant sets of

sensors

Learn robust classification

models

Assess cost/benefit of sensor

topology

SLIDE 25

Results show light/door sensors

don’t change results too much

Many motion sensors appear to

be very valuable

Can test new scenarios without

gathering or annotating new data

Discussion

Few researchers seem look at

sensor importance/selection

Potentially can increase utility of

all sensors in network

SLIDE 26