20
Low-Power Ambient Sensing in Smartphones for Continuous Semantic Localization Sinziana Mazilu, Alberto Calatroni, Ulf Blanke, Gerhard Tröster Wearable Computing Lab Swiss Federal Institute of Technology (ETH) Zürich

AmI2013 Talk

Embed Size (px)

Citation preview

Low-Power Ambient Sensing in Smartphones for Continuous Semantic Localization

Sinziana Mazilu, Alberto Calatroni, Ulf Blanke, Gerhard Tröster

Wearable Computing Lab

Swiss Federal Institute of Technology (ETH) Zürich

2

Where (Usually) Am I?

GPS: Latitude, Longitude Google Maps: Gloriastrasse 12, Zürich Google Services: IfE, ETH, GloriaBar

Physical Location

Logical Location

Office

Cafeteria

Semantic Location

3

Routinely Visited Semantic Locations

Why Semantic Localization?

• Location is a powerful cue for daily-life habits [Patridge et al. - On using existing time-use study data for ubiquitous computing applications - Ubicomp08]

• Activity recognition [Liao et al. - Location-based activity recognition – NIPS05]

• Learning daily routines automatically [Liao et al. - Learning and inferring transportation routines – Artif. Intellig. 2007]

4

Data for Semantic Localization?

GPS/WiFi/GSM - GPS does not work all the time (e.g., indoors, or phone in the pocket) - Energy expensive

I am actually in the Restaurant B.

Restaurant A, GPS says I am here.

I am here.

Where WiFi/GSM

locates me.

I am in the car, tram, or

outside?

+ Accelerometer, Gyroscope

WiFi/GSM very coarse

5

Data for Semantic Localization?

Audio

- Energy expensive - Complex processing algorithms - No privacy regarding activities

Ambient data?

Temperature, humidity, pressure and light

6

Ambient Sensors in Smartphones

http://blog.gsmarena.com/samsung-describes-what-each-of-the-nine-sensors-on-the-galaxy-s4-does/ 7

Two Research Questions:

8

Is ambient sensing informative for continuous sematic localization?

Is it energy efficient?

Data gathering app

9

Does It Look Promising?

10

5 days of data collection with the smartphone

Does It Look Promising?

11

More Users, More Data

250 hours of collected data 49 days

Does It Still Look Promising?

12

Does It Still Look Promising?

13

Evaluation

14

10s

Mean Std Min Max

Latitude Longitude Velocity …

13 MFCC Mean Std

Supervised Classifier

Detected Location

Labeled Location

15

Average accuracies (%) over all categories for each subject dataset, for semantic-location detection

Is Ambient Sensing Data Working?

Is It Energy-Efficient?

Is It Energy-Efficient?

16

Theory ...

Is It Energy-Efficient?

17

... versus practice

80% battery saving with THPL

Two Research Questions:

18

Is ambient sensing informative for continuous sematic localization?

Is it energy efficient?

Routinely visited places

Is That Enough? - Few users?

- Season dependent?

- Labels given by the users (not robust at all)

Label transfer from GPS/WiFi to other sensors

Use additional meteo station data

19

Thank you!

[email protected]