18
Wearable Computing Alexander Nelson November 2, 2020 University of Arkansas - Department of Computer Science and Computer Engineering

Wearable Computing - University of Arkansascsce.uark.edu/~ahnelson/CSCE5013/lectures/lecture16.pdf · Wearable Computing Alexander Nelson April 1st, 2019 University of Arkansas -

  • Upload
    others

  • View
    3

  • Download
    0

Embed Size (px)

Citation preview

Page 1: Wearable Computing - University of Arkansascsce.uark.edu/~ahnelson/CSCE5013/lectures/lecture16.pdf · Wearable Computing Alexander Nelson April 1st, 2019 University of Arkansas -

Wearable Computing

Alexander Nelson

November 2, 2020

University of Arkansas - Department of Computer Science and Computer Engineering

Page 2: Wearable Computing - University of Arkansascsce.uark.edu/~ahnelson/CSCE5013/lectures/lecture16.pdf · Wearable Computing Alexander Nelson April 1st, 2019 University of Arkansas -

Review: Applications of Wearable Computing

1

Page 3: Wearable Computing - University of Arkansascsce.uark.edu/~ahnelson/CSCE5013/lectures/lecture16.pdf · Wearable Computing Alexander Nelson April 1st, 2019 University of Arkansas -

Fitness & Wellness Tracking

Page 4: Wearable Computing - University of Arkansascsce.uark.edu/~ahnelson/CSCE5013/lectures/lecture16.pdf · Wearable Computing Alexander Nelson April 1st, 2019 University of Arkansas -

Fitness & Wellness Tracking

One of the more well-known applications of wearable computing

Application Specific – Detect physiological parameters & report

Examples:

• Inertial Sensors

• Pedometers/Step Counters

• Sleep (w/ microphone)

• Accelerometers/Gyroscopes/Magnetometers (IMU)

• Evoked Potential (Electrodes)

• Heart-rate

• Sweat sensors

• Location (GPS/Wi-Fi/Cellular)

• Air Quality

• Respiratory rate

2

Page 5: Wearable Computing - University of Arkansascsce.uark.edu/~ahnelson/CSCE5013/lectures/lecture16.pdf · Wearable Computing Alexander Nelson April 1st, 2019 University of Arkansas -

Function Determines Form

Many Fitness & Wellness applications require specific locations

e.g. Heart-rate/Pulse monitor requires skin contact near pulse

location

Respiratory rate around chest or near mouth/nose

Some of these sensors can be moved off-body

Only works under certain circumstances

3

Page 6: Wearable Computing - University of Arkansascsce.uark.edu/~ahnelson/CSCE5013/lectures/lecture16.pdf · Wearable Computing Alexander Nelson April 1st, 2019 University of Arkansas -

Example: Wi-Breathe

Wi-Breathe – Uses phase

change of Wi-Fi signals to

detect respiratory rate

Requires person to be in

the home, & not occluded

by another individual

4

Page 7: Wearable Computing - University of Arkansascsce.uark.edu/~ahnelson/CSCE5013/lectures/lecture16.pdf · Wearable Computing Alexander Nelson April 1st, 2019 University of Arkansas -

Function Determines Form

The chosen application determines what sensors/data are needed

• Pulse – Voltage Amplifier + A/D Converter

• Step Counter – Mechanical Switch/Accelerometer

• Air Quality – Chemical concentrations and Particulate

Counters

Also determines how accurate the sensors and data need to be

• Step Counter – ±5% average sensor – No need to be exact

• Pulse – Depends on application

e.g. medical vs. activity and amateur vs. professional

5

Page 8: Wearable Computing - University of Arkansascsce.uark.edu/~ahnelson/CSCE5013/lectures/lecture16.pdf · Wearable Computing Alexander Nelson April 1st, 2019 University of Arkansas -

Measuring/Evaluating Accuracy

With real valued numbers, “accuracy” often refers to both

accuracy and precision

1

1By Pekaje at English Wikipedia - Transferred from en.wikipedia to Commons.,

GFDL, https://commons.wikimedia.org/w/index.php?curid=1862863

6

Page 9: Wearable Computing - University of Arkansascsce.uark.edu/~ahnelson/CSCE5013/lectures/lecture16.pdf · Wearable Computing Alexander Nelson April 1st, 2019 University of Arkansas -

Accuracy/Precision/Resolution

Precision – Description of random errors/statistical variability

Accuracy – Description of systematic errors/statistical bias

These two parameters are independent of each other

Resolution – Smallest change in a quantity that can represent a

change in the measurement device

Properly evaluating these quantities in your system is important

7

Page 10: Wearable Computing - University of Arkansascsce.uark.edu/~ahnelson/CSCE5013/lectures/lecture16.pdf · Wearable Computing Alexander Nelson April 1st, 2019 University of Arkansas -

Example: Step Counting

How do you measure accuracy of a step counter?

• How do you define a step?

• MW – “a movement made by lifting your foot and putting it

down in a different place”

• Orendurff – “unweighted, moved to a new location, and then

re-weighted, in the load path of the body ”

• Binary classification accuracy of a single step?

• Steps counted in a given time frame? (e.g. 79 counted, 81

actual)

8

Page 11: Wearable Computing - University of Arkansascsce.uark.edu/~ahnelson/CSCE5013/lectures/lecture16.pdf · Wearable Computing Alexander Nelson April 1st, 2019 University of Arkansas -

Example: Step Counting

Figure 2: Raw Acceleration Signal2

2Mladenov et al 20099

Page 12: Wearable Computing - University of Arkansascsce.uark.edu/~ahnelson/CSCE5013/lectures/lecture16.pdf · Wearable Computing Alexander Nelson April 1st, 2019 University of Arkansas -

Example: Step Counting

3

3Mladenov et al 2009 10

Page 13: Wearable Computing - University of Arkansascsce.uark.edu/~ahnelson/CSCE5013/lectures/lecture16.pdf · Wearable Computing Alexander Nelson April 1st, 2019 University of Arkansas -

Example: Step Counting

4

4Mladenov et al 2009

11

Page 14: Wearable Computing - University of Arkansascsce.uark.edu/~ahnelson/CSCE5013/lectures/lecture16.pdf · Wearable Computing Alexander Nelson April 1st, 2019 University of Arkansas -

Example: Step Counting

The above example used the following approach:

1. Convert and filter 3-dimensional raw timeseries to

1-dimensional time series

2. Hill-climbing and thresholding to count steps

This method means that classifying steps only works in the context

of other steps (1-step motions would not be counted)

12

Page 15: Wearable Computing - University of Arkansascsce.uark.edu/~ahnelson/CSCE5013/lectures/lecture16.pdf · Wearable Computing Alexander Nelson April 1st, 2019 University of Arkansas -

Reporting Sources of Error

It is important to document your assumptions that could represent

sources of error

Example: Bassett Jr. et al 2016 show that undercounting steps is

a systemic problem with step counters

“... and at 26.8 m/min (1 mph) most devices will record only

50-75% of actual steps”

Overcounting can occur with extra mechanical motions

Example: Fitbit wrist-worn users would trigger steps while brushing

their teeth

13

Page 16: Wearable Computing - University of Arkansascsce.uark.edu/~ahnelson/CSCE5013/lectures/lecture16.pdf · Wearable Computing Alexander Nelson April 1st, 2019 University of Arkansas -

Reporting Error

In an estimator/predictor system (like counts of steps), the Mean

Squared Error (MSE) is often used as a descriptor

MSE (θ) = Eθ[(θ − θ)2]

MSE = 1n

∑ni=1(Yi − Yi )

2

Error parameter that penalizes larger errors, regardless of sign

May also use Mean Absolute Error

MAE = 1n

∑ni=1 |yi − xi |

14

Page 17: Wearable Computing - University of Arkansascsce.uark.edu/~ahnelson/CSCE5013/lectures/lecture16.pdf · Wearable Computing Alexander Nelson April 1st, 2019 University of Arkansas -

Binary Classification

Binary Classification Errors

15

Page 18: Wearable Computing - University of Arkansascsce.uark.edu/~ahnelson/CSCE5013/lectures/lecture16.pdf · Wearable Computing Alexander Nelson April 1st, 2019 University of Arkansas -

Precision vs. Recall

16