Open in TheBrain

Dieter Fox Keynote Location-Based Activity Recognition

<p>Mar 25, 2009 2:15 PM

&#160;

Dieter is now at Univ of Wash Seattle... &#160;robotics.

prev at CMU.

He was&#160;

Chairman of AAAI in 2008.

&#160;

Assisted Cognition for eg Alzheimers... make sure patients don't get loss... try to guess their goals,

when they're lost.

&#160;

figure out what people are trying to do by oberving their routes.

&#160;

Get location info from GPS; &#160;infer where he is and mode of xport... bike, foot, car, bus.

&#160;

use street mapes, bus routes, etc.

&#160;

Also extend to indoor environ using &#160;&#160;WIFI signal strenth instead of GPS

(triangulate with wifi instead of GPS.

&#160;

Map is a &#160;directed graph... &#160;using Kalman filtering to handle noise.

&#160;

He uses &#160;DBN &#160;Dynamic Bayesian Networks &#160;to handle uncertainty.

to combine know sources... &#160;main source is GPS &#160;but also &#160;prior routes, &#160;expectations.

&#160;

Once routes are learned can turn-off (or turn down sampling freq of GPS) &#160;espec if person is on

a route the system has already learned.

&#160;

Its trying to figure out where HOME is &#160;(where you sleep) &#160;where you WORK &#160;where you visit, shop

Can infer mode of xport by velocity.

but &#160;cars stop in traffic; &#160;buses only go on certain routes.

&#160;

Goal is to find and label trip segments... changes in mode of xport.

&#160;

see &#160;&#160;Liao-Fox-Kautz &#160;&#160;AAAI04 &#160;or &#160;AIJ-07.

&#160;

Can see if person gets on wrong bus

&#160;

also uses &#160;conditional random field &#160;(similar to DBN).

&#160;

a1 thru aN &#160;is an activity sequence.... &#160;&#160;&#160;work, drive, sleep work..

&#160;

p1 &#160;thru ;N &#160;are PLACE nodes &#160;home work etc

&#160;

in 2 weeks system gathers &#160;&#160;50,000 &#160;GPS &#160;data points.

&#160;

INDOORS &#160;use &#160;WIFI.

&#160;

Gaussian Process Regression to handle noise. &#160;&#160;first order Markov. &#160;particle filter.

&#160;

They built a little box worn around the neck &#160;with camera, light meter, accelerom, barometric pres to

help assess location; &#160;soldiers wore it;

&#160;

Also used &#160;RFID tags &#160;to monitor breakfast;;;;

&#160;

tried &#160;to do route finding for a robot using a maximize cost/benefit function....

desiderata: &#160;go with the flow of the crow, &#160;avoid collisions, &#160;stay to the RIGHT, &#160;? &#160;max distance to other people.

&#160;

AAAI-SSS-09.

&#160;

Eric Horvitz sitting in front... has 6th grade son.</p>