Applications: Activity Sensing Spencer Kaiser, Laurel Khaleel, and - - PowerPoint PPT Presentation

Applications: Activity Sensing

Spencer Kaiser, Laurel Khaleel, and Jake Drew Ubiquitous Computing Southern Methodist University

Want To Play A Game ???

Researcher or Criminal ?

Researcher or Criminal ?

Boštjan Kaluža Ted Bundy Criminal Researcher

Researcher or Criminal ?

Researcher or Criminal ?

Criminal Researcher Radoje Milic Frank Abagnale

Researcher or Criminal ?

Researcher or Criminal ?

Researcher Criminal Sydney Biddle Barrows Matjaž Gams

Ensembles of Multiple Sensors for Human Energy Expenditure Estimation

Hristijan Gjoreski Boštjan Kaluža Matjaž Gams Radoje Milic Mitja Luštrek

Authors:

Sports faculty interested in measuring expenditure for athletes…

Research Overview

• Human Energy Expenditure (EE) directly reflects the level of

physical activity.

• Actual EE is unpractical to measure outside of the laboratory.
• Find better ways to estimate EE measuring physical activity

with various accelerometers and sensors.

• Previous regression models were created by activity.
• This research uses an ensemble of regression models for EE

estimation which are trained using the output of multiple sensors.

Question How might sports faculty affordances for wearable sensor types and the number of sensors worn be different from other individuals researching energy expenditure

• r activity sensing?

Energy Expenditure

Calorimetry derives the heat transfer associated with changes of its state due for example to chemical reactions, physical changes, or phase transitions under specified

• constraints. Calorimetry is typically performed with a calorimeter.

Direct Calorimetry measures the total heat output of a person. (Not practical outside of the laboratory) Indirect Calorimetry analyzes respiratory gasses which requires a breathing mask. Another approach uses doubly labeled water with marked isotopes which are used to trace the water’s movement throughout the body. (less accurate, but more convenient) http://en.wikipedia.org/wiki/Calorimetry

Cardiopulmonary and Metabolic Testing

Metabolic systems measure breath by breath or mixing chamber oxygen consumption, carbon dioxide production, minute ventilation, anaerobic threshold detection, flow volume loops, lung subdivisions, and maximal voluntary ventilation.

http://www.urmc.rochester.edu/physiology-exercise-lab/equipment/cardio-testing.cfm

Maximal Oxygen Uptake (VO2Max)

• Measures the maximum rate of oxygen consumption. It is used to measure

the functional capacity of the heart, which is a strong indicator of cardiorespiratory fitness.

• Determined by measuring oxygen and carbon dioxide content in

inhaled/exhaled air while continually increasing the intensity of an exercise.

• Occurs at the point at which oxygen consumption no longer increases

despite additional increases to intensity.

http://www.urmc.rochester.edu/physiology-exercise-lab/equipment/cardio-testing.cfm

Resting Energy Expenditure (REE)

Resting Energy Expenditure (REE) is the number of calories burned at rest per day. REE provides the total caloric expenditure in 24 hours and is calculated from the gas exchange data collected by the VIAYSIS VMAX metabolic system.

http://www.urmc.rochester.edu/physiology-exercise-lab/equipment/cardio-testing.cfm

Pulmonary Function Tests (PFT)

Spirometric pulmonary function tests include tests such as forced vital capacity (FVC), forced expiratory volume in one second (FEV1), flow volume loops, and maximal voluntary ventilation (MVV).

http://www.urmc.rochester.edu/physiology-exercise-lab/equipment/cardio-testing.cfm

Six Minute Walk Test (6MWT)

VO2max can be estimated from 6MWT results using multivariate equations. This test can be performed by elderly and patients who can not be tested with standard treadmill exercise equipment.

http://www.urmc.rochester.edu/physiology-exercise-lab/equipment/cardio-testing.cfm

Quark CPET

• State-of-the-art breath by breath gas exchange data analysis (VO2, VCO2)
• Fast response paramagnetic O2 sensor
• Optional 7-liter mixing chamber (either for low or high ventilation ranges)
• Fully integrated 12-lead ECG for Stress Testing (option)
• Nutritional assessment with face mask or optional canopy hood
• Full spirometry and optional exercise SpO2 monitoring

http://www.cosmedusa.com/en/products/cardio-pulmonary-exercise-testing/quark-cpet-stationary-cpet

Actigraphy

Actigraphy is a non-invasive method of monitoring human rest/activity cycles. A small actigraph unit, also called an actimetry sensor,[1] is worn by a patient to measure gross motor activity. The data can be later read to a computer and analyzed offline. In some applications, such as the Fitbit, the data is transmitted and analyzed in real time Moderately accurate and cost is an issue.

$99.00

FitBit

Experiment Sensors

Shimmer three-axis accelerometer BodyMedia SenseWear sensor Zephyr BioHarness sensor Cosmed indirect calorimeter Current State of the Art Baseline EE Benchmark (METs)

Question Why do these sensors limit the application

• f their results only to athletes and people

highly into the quantified self movement? How they could have made the experiments generalizable to a wider audience?

Sensor Features

Activities Measured By Cosmed Indirect Calorimeter

Features for Training Ensemble Models

• Activity (A) – Class label that was being predicted, including a MET value for training.
• Acceleration Peaks (AP) – Count of changes in the direction of acceleration numbers
• Heart Rate (HR) – Discretized raw sensor data
• Breath Rate (BR) - Discretized raw sensor data
• Chest Skin Temperature (CST) - Discretized raw sensor data
• Galvanic Skin Response (GSR) - Discretized raw sensor data
• Arm Skin Temperature (AST) - Discretized raw sensor data
• Ambient Temperature (AT) - Discretized raw sensor data

Values are divided into similar sized bins. Called “Context Components”

Creating Model Ensembles

• 1,000,000 Raw Data Samples Per Volunteer
• Four Machine learning methods were tested: linear regression, Gaussian processes,

multilayer perceptron (artificial neural network) and SMOReg (support vector regression)

• Four models were created for each of the seven Context Components or Features

which were divided into discretized bins each.

• 4 Methods * 4 Discretized Bins * 7 Context Components = 112 Total Models!
• Each time a prediction was made, 7 models were selected for the prediction based

upon the same discretized bins created during training.

• Each of the 7 model’s MET predictions were combined (using average, median, or
• ther) to estimate the final EE.

Model Validation

• A leave-one-person-out cross-validation technique was used
• Models were trained on the data of nine people and tested on the remaining person
• Procedure was repeated ten times, for each person
• Mean Absolute Percentage Error (MAPE) - The mean absolute error divided by the

true value was used for evaluation.

• MAPE is the most common metric in the EE estimation domain

Results

• To evaluate results, they compared against four standard regression methods:
• Linear Regression
• Gaussian Processes
• SMOReg (support vector regression)
• Multilayer Perceptron (artificial neural network)
• The Ensemble significantly outperformed the baseline and the SenseWear across all

four regression methods

Results

• The authors also chose to demonstrate results on a context basis
• Overall, the MAPE of the Ensemble across all contexts greatly out-performed the

MAPE of individual contexts

• “The whole is greater than the sum of its parts.” - Aristotle

Results

• Lastly, the authors compared estimated MET values against true MET values for a

variety of activities and placed the values on a scatterplot

• The results indicate that the Ensemble’s estimations better match the actual Cosmed

MET values than those of the other two approaches

• The SenseWear is intended to be used for “dynamic activities” which is likely why it

performed better for those activities and worse for “day-to-day” activities

• The results in this paper seem pretty straight-forward to me. More sensors and better

algorithms mean that the estimation will be better. Is that conclusion common sense? Is writing a paper to prove that valuable to ubicomp research?

Discussion Questions

• Is it practical to include this many sensors to measure EE?

Discussion Questions

Dog’s Life: Wearable Activity Recognition for Dogs

Thomas Ploetz Patrick Oliver Nils Hammerla Cassim Ladha Emma Hughs

Brief Overview

• “communicative behaviors” – specific moods, desires, or

intentions of the animals

• Comprised of movements that demonstrate the body language of the dog
• “response behaviors” – last for a short period of time and are

usually a response to stimuli

• Certain behaviors have been shown to be indicative of disease and pain
• A dog’s head plays a key role in most of its normal activities

Activity Sensing Platform

• Using a wearable accelerometry sensing platform
• Platform:
• AX3 accelerometer – Axtivity
• 3 axis accelerometer
• Contains an accelerometer and microcontroller
• Sensor attached to a dog collar
• Records a continuous stream of data

• 18 dogs
• 6 female/12 male
• 2 small/5 med/9 large
• 13 breeds
• 17 activities (70% classification accuracy) and 16 behavior

traits

• Participants (dog owners) put the collar with the sensor on

their dog to record daily activities

• Owners also given a video camera to help validate the results
• f the collar
• Dogs mostly left alone except where stimulation was needed
• Experiment in compliance with the UK’s ASPA regulations

Experiment

Behaviors and Triggers

Results Overview

• Overall recognition activity
• f 68.6%
• Much of the inaccuracy is

for walking

• Stationary activities easier

to detect

• Results generally more

accurate for smaller dogs

Confusion Matrices

Voyce

• Basically a “fitbit” for dogs
• Measures steps, rest, heart, and

respiratory rates

• Can help to show if a dog is sick
• Comes in a collar
• $300 plus subscription fee
• video

http://www.techhive.com/article/2085011/meet-voyce-the-sensor-packed-wearable-tech-wellness-monitor-for- dogs.html

Interpretation

• All the dogs in the study were “healthy”; how would results

differ from unhealthy dogs?

• Is ~70% accuracy high enough to detect a noticeable difference

in abnormal activity?

• Why did they choose these activities?
• Did gender affect results in any way?
• Did the ages of the dogs vary significantly? (ie. puppy vs. old

dog)

Weaknesses

• Well being not detected, only activities
• Results not very clearly explained
• Experiments were short term
• Activities collected could have been presented in a tabular

format

• Study heavily dependent on owners could potentially bias the

data

• Is 70% a high enough reliability for an emerging device? Does the reliability threshold

change when the user is a dog and not a human?

Discussion Questions

• What would a system like this cost? Would most people be willing to pay for a system

like this for their dog?

Discussion Questions

Thank You!