Did You See Bob? Human Localization using

advertisement
Did You See Bob?
Human Localization using Mobile Phones
Ionut Constandache
Duke University
Presented by: Di Zhou
Slides modified from Nichole Stockman
The Issue
• Finding someone in a public place can be difficult

Might not know their location

Might be unfamiliar with the area

Maps and floor plans are not always available
Hypothetical Scenario
• Mobicom conference – in a big hotel
• Alice wants to meet her colleague Bob…
How?
– Can walk around
– Can ask “Did you see Bob?”
• But these can take a long time and he may be moving
– Can call him
• But he may be in a meeting already
• Best to have someone escort you…
Current Solutions
• GPS
– Drains the battery (and not very precise)
– Outdoor scheme
• WiFi/GSM schemes
– Not enough accuracy than GPS
– Require special infrastructure, RF transmitter or
war-driving
Escort
• Guide a user to the vicinity of a desired person
– mobile phone sensors
– Opportunistic encounters
– Client/server model
• Does not require:
- Physical coordination
– GPS/Wi-Fi
– War-driving
– Maps or floor plans
• Can be easily installed and provides localization
service when GPS/WiFi not available
Outline
• Motivation Recap:
– We need efficient and accurate software for
people localization: Escort
• Outline
– Basic Design of Escort
– Challenges
– Solutions
– Experiment Specifics
– Results
– Future Work
Basic Design

Escort consists of two parts:
•
Navigation
- Get directions to the person
Visual Identification
- End to end human localization
•
Part 1 - Navigation
• Clients report “Movement trail” periodically
–
–
–
–
–
A(t), B(t), C(t)
<displacement, direction, time>
Use accelerometer and compass measurements
Displacement = # of steps multiply step size
Direction read from compass
• Also report “Encounters”
B(t)
B(t)
– T_AC, T_BC
– <movement trace, encounter time>
T_BC
– Definition ->Audio signals in inaudible frequencies within 5m
T_AC

Server builds a virtual graph

C(t)
Global view of users’ positions andC(t)
paths
A(t)
Challenges
1. Accelerometers & Compasses are noisy
– Measured path error
over time
2. No global reference frame
– How to correct errors?
3. Even if correct user position, difficult to
correct entire trail
- Yet is necessary! (routing)
4. Trail graph grows over time
Solutions - 1
1. Accelerometers & Compasses are noisy
– Use: (step size) * (step count)
Signature from up and down bounce
of human body while walking
Displacement error using double
integration
Solutions - 1
1. Accelerometers & Compasses are noisy
– Use: (step size) * (step count)

Take into account varying step size
(Vary step size with error factor drawn from Gaussian distribution centered
on 0 and standard deviation 0.15m)
Error with step count method (avg 4%)
Solutions - 2
2. No global reference frame
– Use a fixed beacon transmitter
•
Beacon Transmitter
– Location is origin of a virtual coordinate system
– Location diffusion ( single point updates )
– Drift cancellation ( path correction )  sol’n 3
Solutions - 3
• Drift Cancellation
– Amortize the correction vector over time
– Assume that user’s projected path
deviates from the true path linearly over
time
User encounters beacon at t_r1,
and another beacon or recently
updated user at t_r2.
Solid Line: actual path
Dotted Line: user-computed trail
Dashed Line: corrected user trail
Solutions 4
•
Graph computation done by Server
–
–
Pruning heuristic -> eliminate duplicates
Floyd-Warshall algorithm -> shortest paths
Graph – 4 users, 10 min
After pruning
After Floyd-Warshall alg
Basic Design

Escort consists of two parts:
1. Navigation
2. Visual Identification

Help you identify the person if she is someone you
have not met before
(e.g. first-time meeting
with a professional
colleague at a conference)
Basic Design – Visual Identification

Perhaps Alice has not met Bob before…


In a trusted environment like Mobicom, can



Picture of face not enough
Take many photos of user to generate a “fingerprint”
Alice takes photos of surroundings. Image processing is
done to identify Bob from the photos
Totally works in theory!

Currently only implemented
offline and requires user input…
Experiment Specifics
• Parking Lot Experiment
– 4 users, 13 min, phones in hand,
40 routing exp’mts
– Used parking spot lines &
markers for ground truth
(GPS not fine-grained enough)
• Indoor Experiment
– 2 users, 6 min, 10 routing exp’mts
Results – Parking Lot
• Instantaneous location error over time
Results – Parking Lot

Instantaneous location error < 10m:
Intertial ------------------------ ~6% of cases
 Beacon & Encounter ------- ~ 68% of cases
 Drift Cancellation ----------- ~ 84% of cases


Final destination distance error 8.2m on average
Results – Indoor

Instantaneous location error:

Better overall (due to indoor structure and user guesswork)

Intertial --------------------- (N/A because no GPS indoors)
Beacon & Encounter ------ ~ 85% of cases
Drift Cancellation ---------- ~ 90% of cases



Final destination distance error was 7m on average
Results – Visual Identification
• Is this a good result?
– 80% accurate with 8 people in surroundings
Future Work
• “Escort is not designed for energy efficiency”
– Turn off sensors (except accelerometer) when user not
moving
– Less audio signaling when many users or beacons nearby
– Less frequent uploading of data to server
• Routing through physical obstacles
– People are smart
– Visual representation of path may help
• Long routing paths
– Include truedirection arrow
to Bob’s location
Future Work
• Routing instructions under low location accuracy
– If update too far in past, prompt user to approach beacon
– Update path as more info becomes available
• Phone placement
– Need advancement of phone
gyroscopes to infer orientation
• Behavior under heavy user load
– Scalability needs to be explored but should improve
performance (more beacons/users = more updated info, etc)
Thank You
Any
Questions?
Download