Adaptive Radio Interference Positioning System
Hao-ji Wu, Henry Chang, Bing You, Hao Chu, Polly
Huang National Taiwan University
Modeling and Optimizing Positional Accuracy based on Hyperbolic Geometry
The Short Story
• Improve accuracy of the RIP system (Vanderbilt U.)– RIP: Radio interference positioning
• Improvement comes from our adaptive mechanism– Model the error in the original RIP system
– From the error model, identify the controllable parameter (i.e., beacon node selection)
– Adapt based on the controllable parameter
• Implemented & evaluated in a real testbed– Verified our error model is accurate
– Showed our accuracy improvement is real.
Now the long story
4
Motivation
• Accuracy and precision are important performance goals in localization systems.– WiFi localization systems > 1 meter
– Look for a localization system with sub-meter positional accuracy (ideally with long range and low cost) similar to WiFi.
– Duplicate it and improve upon it.
(about the same time last year.)
5
Radio Interference Positioning (RIP) System• Proposed by Vanderbilt University (Kusy et al)
– Found it in Sensys 2005– Recent papers: Sensys 2007, Mobisys 2007, IPSN 2006, etc.
• High accuracy– Average accuracy 3 cm (Sensys’05)
• Long sensing range– 160 meters (Sensys’05), few nodes can cover wide area
• Low cost– Using standard sensor network nodes (Mica2), no additional
specialized hardware
• Amazing accuracy?– Mica2 node dimension (5.8 x 3.2 x 0.7) cm
6
Two starting questions
• Are the RIP system results reproducible?
• Can the RIP system be improved further?
7
Our Findings
• Are the RIP results reproducible?– Reproduced results
• Average: 75 cm; 90%: 1.41 meters
– Experimental setup is slightly different form Vanderbilt U.• 900 Mhz versus 433 Mhz (Vanderbilt U.)• 3 cm is very ideal situation
• Can the RIP system be improved further?– More room for improvement in our adaptation mechanism
8
Overview of RIP system
• Two phases• In the ranging phase: two ranging rounds• In each ranging round, we need to select
two beacon nodes (called sender-pair)– Target is a receiver
– Another infrastructure node is a receiver
• Why two beacon nodes?
Ranging Positioning
Anchor node(known location)
• A, B are senders– Transmit at two nearby
frequency (900 MHz)
– Produce an interference wave with low-beat frequency envelop (350 Hz)
• C (target), D are receivers– Each detects phase, and
jointly the phase difference
– Map the phase difference to a distance difference | AC – BC |
From Sensys 2005 (Kusy et al)
10
RIP ranging & positioning
• Ranging_result
= distance(T,sender1) – distance(T,sender2)
= d1 – d2
• Each ranging result in a
hyperbolic curve• Two ranging needed for positioning• Two sender-pairs selected in two
ranging round are called
Sender Pair Combination (SPC) T
sender1
sender2
sender1 sender2
d1
d2
d1
d2
11
Modeling Position Error in RIP
• Ranging phase– Imperfect measurements lead to ranging error, i.e., error in (d1 –
d2) ~ 26 cm, more or less independent of distance to the beacon
– Ranging error leads to shift in the hyperbolic curve
• Positioning phase– Compute intersection of two shifted hyperbolic curves (due to
ranging errors)
– Ranging error may amplify the positioning error to different amounts
• Depend on the geometry layout between target and SPCs• [an example]
12
How geometry layout affects positional error?• Two geometry layouts from two different SPC selections• With the same amount of range error, Layout (A) has a
larger error than Layout (B), why?– Can you identify two geometric properties leading to the larger
error in (A)?
sender1
sender2
sender1 sender2
sender1
sender2sender1
sender2
Errorrange
T
T’
T
T’
(A) (B)
13
Geometric property #1
• The position of T on the curve– Closer to the foci, smaller the error displacement from the curve,
and smaller the error amplification
sender1
sender2 sender1
sender2
Errorrange
T T
14
Geometric property #2
• Intersectional angle of hyperbolic curves• Shaper the intersectional angle, larger the error amplification
T’
T T
T’
T
T’
θ θ θ
θ= 90° θ= 60° θ= 30°
15
Geometric property #2 (cont.)
• Intersectional angle of hyperbolic curves• (A) has sharper intersectional angle than (B)
sender1
sender2
sender1 sender2
sender1
sender2sender1
sender2
Errorrange
T
T’
T
T’
θ θ
(A) (B)
16
Map Geometry properties to estimation error model
2
yx2range
error2
2
yxrange2
error2yxrange2
error2
yxrange1
error1
)T,(Tq
q tan
)T,(Tq
q
sin
sec)T,(Tq
q
)T,(Tq
q
PT
2
yx2range
error2
2
yxrange2
error2yxrange2
error2
yxrange1
error1
)T,(Tq
q tan
)T,(Tq
q
sin
sec)T,(Tq
q
)T,(Tq
q
PT
17
Validation of Estimation Error Model
Static RIP (6 anchor nodes, 10 meters radius)
Blue line : target pathBlue point : ground truthRed point : estimation position
18
Validation of Estimation Error Model
• Estimation error falls within 6 cm 90% of the time.
19
Adaptive RIP system
• We have an accurate Estimation Error Model– Predict Errorpositional using specific SPC
• We run Estimation Error Model for all SPC (exhaustive search), and find the SPC with minimum error
20
Evaluation of adaptive RIP system
• Single-target positioning experiment• Multi-target positioning experiment
A~F are anchor nodesEach grid is 1m2
21
Single-target positioning experiment
Static RIP Adaptive RIPBlue line : target pathBlue point : ground truthRed point : estimation position
22
Average error (meter)
90-th percentile (meter)
Static RIP 0.93 1.66
Adaptive RIP 0.49 0.75
Improvement 47% 55%
Single-target positioning experiment
walking repeatedly 5 times, around 50 samples
23
Multi-target positioning experiment
• 6 targets– 1 moving (blue point)
– 5 static (1~5 green points)
24
Average error (meter)
90-th percentile (meter)
Static RIP 0.75 1.41
Adaptive RIP 0.30 0.54
Improvement 60% 61%
Multi-target positioning experimentStatic
25
Conclusion
• Error in RIP system is determined by geometric location of the target and beacons.
• We created an Error Estimation Model– Estimate error given relative location of the target and SPCs.
• Adaptive mechanism is built upon the error estimation model.– Find SPC with minimum error
• Showed that Error Estimation Model is accurate.• Showed that Adaptive mechanism improves accuracy.
26
Q & A
27
Flow of RIP system
1. Start RIP system
2. Select 2 anchor nodes as senders
3. Other anchor nodes become receiver
4. Ranging
5. If 1st ranging => goto 2 and do 2nd ranging
else => goto 6 do positioning
6. Positioning
7. End
T
Sender1
Sender2 Receiver2
Receiver1Sender1
Sender2Receiver2
Receiver1Ranging Positioning
Anchor node(known location)
1st ranging2nd ranging
28
Geometrical Derivation of Estimation Error Model• Steps:
1.Find TN (TM)
2.Find θ
3.Find PT geometrically
If we know TN, TM, θ, we can derivative PT geometrically
Known Variables:• Target location• SPC• Errorrange
MN
MN1-
mm1
m mtan
mN
mM
29
Original RIP system• Anchor nodes are placed in known locations• The positional error of RIP system is highly affected by target
locations and the selection of beacon nodes• Original RIP system is “static”
– The selection of beacon nodes is fixed, doesn’t change depend on target locations
T
Beacon
Beacon
30
Our Contribution• Design, implement and evaluate an adaptive RIP system
– Dynamically select beacon nodes based on target locations
31
How SPC selection affect positional error (cont.)• Displacement of a hyperbolic curve
– Shortest distance from hyperbolic curve with error to target
sender1
sender2 sender1
sender2
Errorrange
T T
• PT is the estimation Errorpositional
• N is the projection point of T on H12
– TN is the displacement of hyperbolic curves H12
• θ is the intersectional angle of hyperbolic curves H12 and H34
• Find TM (TN) and θ geometrically
– Find PT geometrically
32
Map Geometry properties to estimation error model
Errorrange