APT: Accurate Outdoor Pedestrian
Tracking with Smartphones
TsungYun 20130401
1
Outline
•
•
•
•
•
Introduction
Preliminary Experiment
System and Mechanism
Evaluation
Conclusion
2
Introduction
• Motivation
– Want to build a system to assist the blind people
with smartphones by providing accurate location
information
– GPS measurements show error up to 15 meters in
a clear-sky-view environment
3
Introduction
• Observations
– Pedestrians have regular movements patterns
– Although GPS is unsatisfactory, it works well in
distinguishing between distant routes
– Can easily generate augmented maps on a
smartphone
• Dead-Reckoning algorithm
• Map-Matching algorithm
4
Introduction
• Dead-Reckoning algorithm
– Accelerometer: walking step
– Gyroscope: walking direction
• Consume much less energy than GPS
• Map-Matching algorithm
– Match a walking trace to a route on the map
• Challenges
– Placement of the smartphone
– Error-tolerant
5
Preliminary Experiment
• Limitation of GPS system
– GPS system achieved error up to 15 meter
– GPS readings cannot be improved by itself solely
• First issue
– If the GPS coordinate stabilizes, then it will not
change for at least several hours
– staying in one place longer does not help improve
GPS accuracy
6
Preliminary Experiment
• Collect 15-20 GPS
coordinates at three
locations at seven different
days
– Clear view of the sky
– Do not mention how far
between these locations
7
Preliminary Experiment
• Results show that
– GPS readings at the
same location can differ
up to 15 meters
– hard to find any
obvious temporal or
spatial correlation
8
Preliminary Experiment
• Walks along a route 5
times
– a large portion of this
route is covered by trees
• Result shows
– the error can still be
more than 20 meters
– no obvious error pattern
9
Preliminary Experiment
• Conclusion
– We find that it is unlikely to improve localization
accuracy based solely on GPS
• In this work, the use of GPS is limited to help
reduce route ambiguity in the Map-Matching
algorithm
10
Mechanism
11
Mechanism I
• Dead-Reckoning – estimating distances
– taking the double integral of acceleration results
in large error
– a common approach is to count the number of
walking steps and then multiply it by the stride
length
• By finding the recurring patterns of accelerometer
readings
12
Mechanism I
• Different placement
of the phone has a
large impact on the
accuracy of each step
counter
– 6 recurring patterns
– 3 recurring patterns
13
Mechanism I
• No matter how the phone is placed, we find
that acceleration always shows some recurring
patterns
– define an up-down pattern as a step
– A pattern ‘10’ or ‘1 ∧ 0’ is defined as a step
14
Mechanism I
• Using acceleration magnitude, instead of
acceleration in a certain direction, can tolerate
different ways pedestrians carry the phone
• Step length can be measured or trained in
advance
15
Mechanism I
• Dead-Reckoning – estimating direction
– two Cartesian frame of reference
– xyz axes V.S. XYZ axes
– We can obtain
• x y z data
– We need
• Z data
16
Mechanism I
• straight line -> 90° left turn -> straight line
– angular displacement around any axis remains
roughly the same before/after the turn
17
Mechanism I
• straight line -> 90° left turn -> straight line
– acceleration does not fluctuate much before/after
the turn, but is quite unusual during the turn
18
Mechanism I
• angular displacement around Z-axis
– α, β, γ are the angular displacements around x, y, z
axis
– µx , µy , µz are the acceleration readings in x, y, z
direction
– the average acceleration during a straight walk should
approximate gravity
– Z-axis vector (the gravity) is decomposed into
three components
???????
19
Mechanism I
• The angular displacement is 91.56◦ in this case
– But the error (1.56◦) is inevitable
20
Mechanism II
• Map-Matching algorithm
21
Mechanism II
• Map-Matching algorithm
 Use GPS here
 trial-and-error
22
Mechanism II
• Map-Matching algorithm
– Two position fixes can determine a matching
– Basic idea : Trial-and-error
• Starting from one position fix, find out all possible
routes
• use subsequent points in the walk to test and extend
these routes
23
Mechanism II
• Map-Matching algorithm
– Assume “perfect information”
• First assume that accelerometer, gyroscope, GPS
readings are 100% accurate
– Update when
• New step
• New turn
• New GPS reading
24
Mechanism II
 Use GPS here
 Use MAP here
 Reversely check
↑ Use GPS here
↑ Use GPS here if multiple routes
to reduce ambiguity
25
Mechanism II
• Dealing with errors
– Initial routes
• We enumerate all possible locations of the user on the
map by considering GPS error
– A new step
• An adjacent route segment is possible if walking to it
only requires a shallow turn within angular error
tolerance
26
Mechanism II
• Dealing with errors
– A new turn
• Find out all route segments that are reachable by a turn
within the range: the reported angular displacement
plus/minus angular error tolerance
– A new GPS coordinate
• When a new GPS coordinate is available, check each
possible route by verifying whether the new GPS
coordinate is within a certain distance: (distance error
tolerance plus GPS error)
27
Mechanism II
• Map-Matching algorithm
– If no possible route exists
• the system will restart by requesting a new GPS
coordinate
– When a step and a turn arrive simultaneously
• ignoring the steps during a turn
– When the number of possible routes becomes
intolerable
• request a GPS coordinate
28
Evaluation
• Experiment
– In each second
• 50 accelerometer readings
• 50 gyroscope readings
• 1 GPS reading ???? Energy ????
• Tolerance setting
– Distance error tolerance : 20 m
– Angular error tolerance : 30°
– Based on experience and haven’t been optimized
29
Evaluation
30
Evaluation
• Compare APT algorithm to:
– Raw GPS coordinates tracking system
– Combine the raw GPS coordinates with the map
information
• In all three routes, our algorithm have
consistently less error
– The most complicated route, contains more turns, the
error is 0 at most anchor points
– The error at non-turn anchor points is at most 5m
31
Evaluation
32
Conclusion
• This paper present APT, a system targeting at
accurate pedestrian localization
• Uses the accelerometer, gyroscope and GPS
component of modern smartphones, and
integrates them with map information
• Can tolerate GPS error and the different ways
to hold the smartphone
• Achieve better performance than GPS only
33