acquiring traces from random walks project final presentation by: yaniv sabo aviad hasnis...

19
Acquiring traces from random walks Project final presentation By: Yaniv Sabo Aviad Hasnis Supervisor: Daniel Vainsencher

Post on 19-Dec-2015

220 views

Category:

Documents


0 download

TRANSCRIPT

Acquiring traces from random walks

Project final presentation

By:Yaniv Sabo

Aviad Hasnis

Supervisor:Daniel Vainsencher

Project GoalCreating traces of indoor walks

using signals collected by different agents.

Possible problem: GPS is not feasible indoors.

Possible SolutionThe agents will be cellular

devices (HTC) using Android platform.

Using the following signals:◦Built-in accelerometer producing the

device acceleration in each direction. ◦Built-in magnetometer measuring

the magnetic field at the device’s location.

◦WIFI signal levels received from multiple routers, combined using triangulation.

BackgroundAndroid is a mobile operating

system currently developed by Google.

AccelerometerMagnetometerWIFITriangulation

Background - Cont’d

1p

2p

3p*p

Possible Solution – Cont’dEach one of the signals we used

isn’t accurate enough on it’s own.◦The direction received from the

Accelerometer accumulates error very fast.

◦The WiFi measurements have a high variance because of significant noise.

Integrating the former inputs in order to create an accurate path.

WiFi signal strength to distanceWe wanted to find a function that

converts the WiFi signal level received to the distance from the AP.

According to articles, the function should behave like:

where a, b and c are constants.We did some experiments and using

Least Square Error we found that these constants should be:

10b p

cd p a

4.2427 , 64.995 , 30a m b dBm c dBm

System Flow

Accelerometer

WIFI receiver

Magnetometer

Algorithm

3

Algorithm - Cont’dWiFi regularization

◦Limiting the velocity to normal walking speed and limiting the acceleration.

Algorithm - Cont’dCombining WiFi measurements

◦Measurements are received from different WiFi APs at different times.

◦Requires an algorithm to combine these signals.

◦Two approaches: ring and a circle.

Algorithm - Cont’dCombining Accelerometer data with

Magnetometer data◦ The Accelerometer accumulates error in a

high rate.◦ We use the Magnetometer to get the

device’s direction and the Accelerometer to get the device’s acceleration.

ˆ ˆaccv n v n

sinˆ

cos

tdy dxn t

dx dy t

Algorithm - Cont’dAfter we have the WiFi processed

measurements as well as the Accelerometer and Magnetometer data combined we integrate these signals.

We use loss functions to denote the distance of the solution from the signals and we attempt to minimize these loss functions.

We also perform regularization on the solution.

Algorithm - Cont’dWe tried different loss functions

for the WiFi measurements:

We also tried different loss functions for the Accelerometer and Magnetometer measurements:

211.1 1.1

,0

guess real guess realg real guess real

otherwise

2, ,acc magg real guess real guess

Some Results

Some Results – Cont’d

Some Results – Cont’d

ConclusionsEach one of the signals we used

isn’t accurate enough on it’s own. We combined these signals to get a more accurate solution.

We have found that the quality of the solution depends heavily on the loss function used.

The methods used to collect data from the device have great effect on the precision of this data.

Future WorkInvestigating additional loss

functions, as well as faster functions for unconstrained optimization than MATLAB’s fminunc.

Using additional signals, such as Bluetooth, to improve the solution.

Combining different traces to build a map of the interior of a building.