indoor localization without infrastructure using the acoustic
TRANSCRIPT
S. P. Tarzia - P. A. Dinda - R. P. Dick - G. Memik
Indoor Localization without Infrastructure using the
Acoustic Background Spectrum
Advanced Computer Architecture - Politecnico di Milano - 2012
Vincenzo Baldini - Matteo Torcoli
Presentation outline
● starting problem: localization● ABS concepts● how to get a good ABS fingerprint?● accuracy issues● an implementation for mobile phone● putting the application to the test● future works
reference: http://stevetarzia.com
Indoor Localization
i.e. which room am I in? ● outdoor localization -> well solved by GPS ● indoor localization
without an infrastructure specifically deployed for this scope?
Wi-Fi localization
fingerprint = WiFi signals intensity
fingerprint -> roomlabelroom "1"
room "3"
room "2"
room "4"
Beyond Radio
fingerprint -> roomlable
fingerprint = Acoustic Background Spectrum
● always available resource● surprisingly distinctive! ● compact and easily computed● robust
ABS Goalindependently from any specialized hw
Allow even a basic mobile device to
cheaply and quickly determine its location
by matching
previously-learned, specific location labelsto recordings
2 steps
1. Calculate the room's ABS2. Room's Classification
1. ABS fingerprint extraction
● Record and Windowing● Power Spectrum (through FFT)● Filter the freq. band of interest● Reject transient noise
Transient noise rejection
DistinctiverEsponsive Compact Efficiently-computable Noise-robust Time-invariant
Filter out transient
components
HOW?
5th-percentile value
(P05) for each frequency
ABS fingerprint extraction
2. The classification problem
to label the room
DISTANCE METRIC
VECTOR EUCLIDEAN DISTANCE
Current ABS room fingerprint compared with
previously-observed ones
● training pairs ● testing fingerprint● the classifier chooses:
Trace collection and simulation
Simulation to evaluateABS-localization accuracy
● 33 rooms in the Northwestern University● 30 seconds mono recordings
(24 bit - 96 kHz wav files)● 2 visits in different weeks ● 4 positions per room● Zoom H4n handheld recorder:
ABS from 2 rooms using the optimal parameter values
Simulation results
Confusion matrix for the 33 room simulation
(optimal values)
Noise-robustness problem (noisy room samples)
- Recordings in a lecture hall before, during and after lectures - 3 occupancy noise states:
1. quiet times - in the training set 2. conversation times3. chatter times
- in the test set
AUTOMATIC BAND SWITCHING
EXPERIMENT: frequency bands and occupancy noise
How to improve the accuracy
different fingerprint distances in a linear combination:● wi-fi● cellular radios ● accelerometer● camera data● ABS...
● = distance for each fingerprint types● = weighting constant● = match range
Mobile application: BatphoneMain features:
● it's free!● localization in real-time● capture and export fingerprints● ABS compact fingerprint + Wifi coordinates● almost the same extraction steps
Apple Ipod Touch (iOS 4.0 or later)
with the Batphone's GUI
Hardware limitations:1. microphone (speech frequencies)2. computation power (=> rectangular windows)3. memory (=> compact fingerprints)4. battery
Batphone experiment
● Ipod Touch 4g● 43 rooms ● 2 room positions =>● less training data ● fingerprint = ABS + Wifi
Neverthless good results!
Our Batphone experiment
* Strong change in conditions** Chatter problem
*
* *
Future works
fingerprint sharing/matching infrastructures - online DB
others kinds of fingerprint: - camera data
- accelerometer..
chatter problem: an open issue
automatic switching of frequency bands