1 Localization Technologies for Sensor Networks Craig Gotsman, Technion/Harvard Collaboration with: Yehuda Koren, AT&T Labs
Dec 21, 2015
1
Localization Technologies for Sensor Networks
Craig Gotsman, Technion/Harvard
Collaboration with:
Yehuda Koren, AT&T Labs
2
Sensor networks
• Set of nodes with ability to:
– Measure parameters related to environment
– Process information– Communicate / route– Estimate location– Communicate information to central
processorCreate a smart environment
3
Hardware architecture
sensors CPU/memory
radio
battery
Acoustic, seismic, image, magnetic, etc.
interface
Electro-magnetic interface
Event detectionWireless communication with neighboring nodes
In-node processing
Limited battery supply
Simple, small, cheap
4
Potential applications
• Warning
Tornado Fire
5
Potential applications
• Transportation
– Monitor traffic conditions
– Plan routes– Parking
allocation
6
Limitations
• Power is the bottleneck– Long distance communication
impossible• No pre-configuration or global knowledge
– Achieve global goals through local interaction and self organization
• Limited computational power• Price
Use a very large number of sensors in a wide region
7
Location-aware sensors
• Data should be location-stamped• Geographic routing• Region-targeted querying
(123,456)
(134,778)
(234,466)
(294,666)
(372,862)
(362,423)
(432,553)
(519,450)
(589,703)
8
Find a fully distributed algorithm for sensor localization
Why not simply use GPS ???
The problem we address:
9
Limitations of GPS
• Power• Price• Line of sight
conditions• Accuracy ?
What else can we use ?
10
Distance to neighboring sensors
• Received Signal Strength Indicator (RSSI)• Time of Arrival (ToA)
Technologies:
3
6
75
5
3
5
6
8
3
74
45
6
4
May be noisy
Local distances coordinates (??)
11
Previous Solutions• Anchor-based:
– Some beacon nodes know their exact location
– Other sensors estimate their location from nearby beacons
12
Previous Solutions• Incremental approaches:
– Assign coordinates to a small core of sensors– Repeatedly assign coordinates to more sensors
based on local calculations– Prone to error accumulation
13
Previous SolutionsAnchor Free Localization (AFL) [Priyantha et al., 2004]
A two-stage, distributed approach:
• Based on connectivity, elect central, north, south, east and west sensors
• Estimate coordinates for rest of sensors
1. Initial coordinate assignment
• Optimization using gradient-descent to approximate measured distances
2. Accurate distributed layout
14
Graph layout interpretation • Given a graph with edge
lengths• A layout that realizes all edge
lengths exist• Only close nodes are
connected (“disk graph”) • Goal:
Find this layout !
2
3
6
1 4
2
4
Two issues:
1. Layout existence - measured lengths are noisy !
2. Layout uniqueness – graph’s rigidity
15
Graph rigidity
• In 2D:
– Global rigidity 3-connectivity
– 6-connectivity Global rigidity • Computing layout of rigid graph is NP-hard [Eren et al.,
2004]
1 2
43
• A graph is globally rigid if it has a unique embedding (up to distance preserving transformations)
2,31,4
1
42,3Non rigid
Globally rigid
16
Beyond classical rigidity• In disk-graphs - close nodes must be connected (up
to noise)• Non-adjacent nodes should be placed further apart
• Optimal layout 1, , , d
n ip p p R
Prunes redundant embeddings
,
,
i j ij
i j
p p l i j Ei j
p p r i j E
,edge length- max,ij ij
i j El r l
17
Graph drawing algorithms• Energy-minimization algorithm using localized stress
energy:
2
,i j ij
i j E
p p l
• Known problem: foldovers
Why ???
18
Localized stress and foldovers
• Graph not rigid – the energy does not address nonadjacent nodes
• Local minima – global optimization infeasible
2
,i j ij
i j E
p p l
We must treat nonadjacent nodes explicitly
19
Algorithm overview
1. “Spectral” initialization– Convex optimization – insensitive to
initialization– Tends to generate fold-free layouts– Uses given distances inaccurately
2. Local stress optimization– Sensitive to initialization– Accurate use of distances – produces
optimal layout when initialized properly
– Global optimization – overcomes local noise
A two-phase approach:
20
Spectral layout - Laplacian Based on [Hall, 1970]
Given a weighted graph with n nodes, wij being the edge weights (wij=0 for non-adjacent nodes)
The Laplacian of the graph is the matrix L, where:n n
deg
ij
ij
i
w i jL
i j
5 3 2 0 0
3 10 1 6 0
2 1 9 4 2
0 6 4 14 4
0 0 2 4 6
2
3
6
1 4
2
4
21
Spectral layout - Goal
2 2
,
2 2
,
Mini jij i j
i jn
i ji ji jx y
x y
w x x y y
x x y y
R
Solve:
- coordinates of node i,i iyx
Locate related nodes closely, while spreading nodes wellLocate related nodes closely, while spreading nodes well
Weighted squared distances between
nodes
Squared distances between nodes
• Edge weights express similarity/proximity• Solution is Laplacian eigenvector(s) - scale invariant• Does not use distances directly…
22
Local stress minimization• Relocate the nodes to minimize:
2
,i j ij
i j E
p p l
• Accurate optimization process, addressing measured distances directly
• Effective only when initialized smartly• Generally, the spectral initialization is good
23
Example
Original layout
1000 sensors on 10x10 square, R=0.8
Stress with spectral initialization
24
Example
Original placement
715 sensors on 10-3 ring, R=0.8
Stress with spectral initialization
25
Conclusions• A fully distributed algorithm for sensor
network layout• Based on graph drawing methods• Main challenge: layout computation with only
local communication• Still need to represent distances better in
spectral layout
26
Future Work
• Higher dimensions (3D ?)• Improve spectral embedding using LLE• Implement on real systems
• Incorporate more geometric info (e.g. angles)• Multi-camera calibration
• Dynamic systems