This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Ecolocation Probabilistic localization methods Camera Based Localization Rigidity Other topics mentioned in discussion
• Robust Quatrilaterals• Robustness and secure localization• Radio Interferometric localization
Radio Signal Strength: Ecolocation (Yetvali et. al USC & Bosch)
Initiation: Node with unknown location (Unknown Node) initiates localization process by broadcasting a localization packet. Nodes at known reference locations (Reference Nodes) collect RSS readings and forward them to a single point.Procedure: Determine the ordered sequence of reference nodes
by ranking them on collected RSS readings. The ordering imposes constraints on the location of the unknown node.
For each location grid-point in the location space determine relative ordering of reference nodes and compare it with RSS ordering to determine how many of the ordering constraints are satisfied.
Pick the location that maximizes the number of satisfied constraints. If there is more than one such location, take their centroid.
“Constraints” & “Sequences”
1
2
34
AB
C
D
E E
D
C
B
A
1 2
34
Reference nodes (B,C,D,E) ranked into ordered sequence by RSS readings.
The sequence of reference nodes changes with the location of the unknown node (A).
Ideal Scenario: DAB < DAC => RB > RC
Constraint on the location of the unknown node. RSS relationships between all reference nodes forms the constraint set.
R4<R1
R4<R3R3<R1
R4<R3R3<R2R2<R1R1
E:4D:3C:2B:1
Error Controlling Localization
The inherent redundancy in the constraint set helps withstand errors due to multi-path effects. Analogous to error control coding. Error Controlling Localization:
Ecolocation
Constraint construction inherently holds true for random variations in RSS measurements up to a tolerance level of |Ri - Rj|.
Real World Scenario: Multipath fading introduces errors in RSS readings which in turn introduce errors in the constraint set. Location estimate accuracy depends on the percentage of erroneous constraints.
Ecolocation Examples
0 1 2 3 4 5 6 7 8 9 10 11 12
1
2
3
4
5
6
7
8
9
10
11
12
X-AXIS (length units)
Y-A
XIS
(le
ng
th u
nits
)
Location estimate for 124739586
E
P
A1
A2
A4
A7
A3 A9
A5
A8
A6
0 1 2 3 4 5 6 7 8 9 10 11 12
1
2
3
4
5
6
7
8
9
10
11
12
X-AXIS (length units)
Y-A
XIS
(le
ng
th u
nits
)Location estimate for 913276584
P
E
A9
A1
A3
A2
A7 A6
A5
A8
A4
0 1 2 3 4 5 6 7 8 9 10 11 12
1
2
3
4
5
6
7
8
9
10
11
12
X-AXIS (length units)
Y-A
XIS
(le
ng
th u
nits
)
Location estimate for 123456789
E
P
A1
A2
A3
A4
A5 A6
A7
A8
A9
No Erroneous Constraints
0 1 2 3 4 5 6 7 8 9 10 11 12
1
2
3
4
5
6
7
8
9
10
11
12
X-AXIS (length units)Y
-AX
IS (
len
gth
un
its)
Location esitmate for 123745968
P
EA1
A2
A3
A7
A4 A5
A9
A6
A8
14% Erroneous Constraints
22% Erroneous Constraints
47% Erroneous Constraints
A: Reference Node
P: True Location of unknown node
E: Ecolocation Estimated Location
Simulations
Simulation Parameters
RF Channel Parameters Path loss exponent (η) Standard deviation of log-normal shadowing model (σ)
Node Deployment Parameters Number of reference nodes (α) Reference node density (β) Scanning resolution (γ) Random placement of nodes
Compared with four other localization techniques – Proximity Localization, Centroid, MLE, Approximate Point in Triangle (APIT).
Averaged over 100 random trials with 10 random seeds.
Simulation Results
1 3 5 70
10
20
30
40
50
60
70
80
90
100
110
= 7, = 25, = 0.11, = 0.1
Path loss exponent ()
Ave
rag
e lo
catio
n e
rro
r (%
of
Da)
EcolocationCentroidAPITMLEProximity
2 4 6 8 10 12 140
10
20
30
40
50
60
70
80
90
100
110
= 4, = 25, = 0.11, = 0.1
Standard deviation ()
Ave
rag
e lo
ca
tio
n e
rro
r (%
of
Da)
EcolocationCentroidAPITMLEProximity
3 5 7 9 11 13 15 17 19 21 23 250
10
20
30
40
50
60
70
80
90
100
110
= 0.11, = 0.1, = 4, = 7
Number of reference nodes ()
Ave
rag
e lo
catio
n e
rro
r (%
of
D a)
EcolocationCentroidAPITMLEProximity
0.01 0.04 0.11 1 0
10
20
30
40
50
60
70
80
90
100
110
= 25, = 0.1, = 4, = 7
Reference node density () (log scale)
Ave
rag
e lo
catio
n e
rro
r (%
of
Da)
EcolocationCentroidAPITMLEProximity
Da: Average inter reference node distance
Systems Implementation
Outdoors: Represents a class of obstruction free RF channels. Eleven MICA 2 motes placed randomly on the ground in a 144 sq. m area in a parking lot. Locations of all motes are estimated and compared with true locations. All motes in radio range and line of sight of each other.
Indoors: Represents a class of obstructive RF channels. Twelve MICA 2 motes (Reference nodes) are placed randomly on the ground in a 120 sq. m area in an office building. A MICA 2 mote (Unknown node) is placed in five different locations to be estimated. All motes in radio range but only a subset in line of sight of each other.
Systems Implementation Results
0 2 4 6 8 10 120
2
4
6
8
10
12
X (meters)
Y (
mete
rs)
Outdoor Experiment
True locationEstimated location
1
2
34
5
6
7
8
9
10
11
1 2 3 4 5 6 7 8 9 10 110
10
20
30
40
50
60
70
80
90
Node ID
Ave
rage lo
catio
n e
rror
(% o
f D a)
EcolocationMLEProximity
0 1 2 3 4 5 6 7 80
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
X (meters)
Y (
mete
rs)
Indoor Experiment
Ref. nodesTrue pathEstimated path
1
2
3
4 5
Conference room with furniture
Office room
Office room
Door
1 2 3 4 50
20
40
60
80
100
120
140
160
180
200
Unknown node location
Ave
rage
loca
tion
erro
r (%
of
Da)
EcolocationMLEProximity
Locations estimated using Ecolocation, MLE and Proximity methods.
Results suggest a hybrid localization technique.
Conclusion and Future Work
Future Work: Exploring Hybrid Localization technique further. Making Ecolocation more efficient using greedy search, multi-resolution search algorithms. Analytical background for Ecolocation. Measuring localization costs (Time, Throughput, Energy) for various realistic system designs and protocols.
Conclusion: Ecolocation performs better than other RF based localization techniques over a range of RF channel conditions and node deployment parameters. Simulation and experimental results suggest that a Hybrid Localization technique may provide the best accuracy.
Bayesian Filtering For Location Estimation(Fox et. al [Fox02])
State estimators - probabilistically estimate a dynamic system’s state from noisy observations• In system theory, the information that dynamic system
model gives us about the system is called system state
• A system model is a set of equations that describe the system state.
• The variables in the system model are the state variables
Bayesian Filters
In localization, state is the location of an entity• State representation is based on noisy sensor
measurements• In simple cases, state can be just position in 2D• In more complex cases, state can be a complex
vector including position in 3D, linear and rotational velocities, pitch, roll, yaw, etc.
Bayesian Filters
State (location) at time t is represented by random variable(s) x(t). At each time moment, Bayesian filter represents a probability
distribution over x(t) called belief If we assume a sequence of time indexed sensor observations
the belief becomes
This is the probability distribution over all possible locations (states) x at time moment t, based on all possible sensor data available at time moment t (earlier and present measurements).
1, , tz z
1( ) ( | , , )t t tBel x p x z z
Bayesian Filters: Markov Assumption
Complexity of probability function grows exponentially with sensor measurements
Bayesian systems assume that the dynamic system is a Markov System• State at time t only depends on state at time t-1
Implementing Bayesian Filter
Under Markov assumption, the implementation of Bayesian filter requires following specifications:
1. Representation of the belief2. Perceptual model
– Probability that state x(t) produces observation z(t)
3. System dynamics– Probability that state x(t) follows state x(t-1)
4. Initialization of the belief– Initialized based on prior knowledge, if available– Typically uniform distribution, if no prior knowledge exists
( )tBel x( | )t tp z x
1( | )t tp x x
( )oBel x
Implementing Bayesian Filter
Based on given specifications, Bayesian filter acts in two steps:
1.1. PredictionPrediction. Based on system state at time = t-1, filter computes prediction (a priori estimate) to system state at time moment t
2.2. CorrectionCorrection. When new sensor information corresponding time moment t is received, filter uses it to compute corrected belief (a posteriori estimate) to system state at time moment t
1 1 1( ) ( | ) ( )t t t t tBel x p x x Bel x dx
( ) ( | ) ( )t t t t tBel x p z x Bel x (In the above equation, alpha is simply a normalizing constant ensuring that the posterior over the entire state space sums up to 1.)
Bayesian Filter Example
(a) A person carries a camera that can observe doors, but cannot distinguish different doors. Initialization is uniform distribution.
(b) Sensor sends ”door found” signal. Resulting belief places high probability at locations next to doors and low probability elshewere. Because of sensor uncertainty (noise), also nondoor locations possess small but nonzero probability.
(c) Motion’s effect to belief. Bayes filter shifts the belief (a priori estimate) in the direction of sensed motion, but also smoothens it because of the uncertainty in motion estimates.
(d) Sensor sends ”door found” signal. Based on that observation, filter corrects previous a priori belief estimate to a posteriori belief estimate.
(e) Motion’s effect to belief. Bayes filter shifts the belief (a priori estimate) in the direction of sensed motion, but also smoothens it because of the uncertainty in motion estimates. Compared to case c, belief estimate is converging to one peak that is clearly higher than other ones. One can say that filter is converging or learning.Picture and example from Fox, D., Hightower, J., Liao, L., Schulz, Picture and example from Fox, D., Hightower, J., Liao, L., Schulz,
D., Borriello, G., ”D., Borriello, G., ”Bayesian Filtering for Location EstimationBayesian Filtering for Location Estimation”, ”, IEEE Pervasive Computing 2003.IEEE Pervasive Computing 2003.
Different Types of Bayesian Filters
Kalman Filter• Most widely used variant on Bayesian filters• Optimal estimator assuming that the initial uncertainty is Gaussian
and the observation model and system dynamics are linear functions of the state
←In nonlinear systems, Extended Kalman Filters which linearize the system using first order Taylor series are typically applied
− Best if the uncertainty of the state is not too high, which limits them to location tracking using either accurate sensors or sesors with high update rates
Multihypotesis tracking• MHT overcomes Kalman Filter’s limitation to unimodal
distributions by representing the belief as mixtures of Gaussians• Each Gaussian hypothesis is tracked by using Kalman Filter• Still rely on the linearity assumptions of Kalman Filters
Other Types of Bayesian Filters
Grid-based approaches• Discrete, piecewise constant representations of the belief• Update equations othervise identical to the general Bayesian filter update
equations, but summation replaces integration• Can represent arbitrary distributions over the discrete state space• Disadvantage computational and space complexity
Topological approaches• Topological implementations of Bayesian filters, where a graph represents
the environment• The motion model can be trained to represent typical motion patterns of
individual persons moving through the environment• Main disadvantage
o Location estimates are not fine-grained
Different Bayesian Filters
Particle Filters• Bayesian filter updates are performed according to a sampling
procedure often called sequential importance sampling with resampling
• Ability to represent arbitrary probability densities, can converge to true position even in non-Gaussian, non-linear dynamic systems
• Efficient because they automaticly focus their resources (particles) on the regions in state space with high probability
• One must be careful when applying Particle Filters to high dimensional estimation problems, because worst-case complexity grows exponentially in the dimensions of the state space
Other Types of Bayesian Filters
Particle Filters• Beliefs are represented by sets of samples called particles:
In the equation, each x is a state and w:s are nonnegative weights called importance factors, which sum up to one.
For more detailed treatment of Particle Filters see [Schultz03]
( ) ( )( ) , | 1, ,i it t t tBel x S x w i n
Particle Filter Example
(a) A person carries a camera that can observe doors, but cannot distinguish different doors. A uniformly distributed sample set represents initially unknown position.
(b) Sensor sends ”door found” signal. The particle filter incorporates the measurement by adjusting and normalizing each sample’s importance factor leading to a new sample set, where importance factors are proportional to the observation likelihood p(z|x).
(c) When a person moves, particle filter randomly draw samples from the current sample set with probability given by importance factors. Then the filter use the model to guess (predict) the location for each new particle.
(d) Sensor detects door. By weighting the importance factors in proportion to this probability p(z|x), updated sample set w(x) is obtained.
(e) After the prediction, most of the probability mass is now consistent with person’s true location.
Picture and example from Fox, D., Hightower, J., Liao, L., Schulz, Picture and example from Fox, D., Hightower, J., Liao, L., Schulz, D., Borriello, G., ”D., Borriello, G., ”Bayesian Filtering for Location EstimationBayesian Filtering for Location Estimation”, ”, IEEE Pervasive Computing 2003.IEEE Pervasive Computing 2003.
Bayesian Filters - Conclusions
Dealing with uncertainty Starting for initial estimate, system converges over time to
give more accurate estimates Possibility to exploit several type of sensor measurements
and other available quantitative knowledge of sensing environment (initial estimates, digital maps...)
Suitable Bayesian Filter type depends on sensor type (what informtion is available), sensing environment (indoor, outdoor, noise level...), system model (linear, nonlinear, continuous time, discrete time..)
In addition to localization, several other application fields exists in pervasive computing• Movement recognition, data prosessing
Camera Assisted Localization
What can cameras measure?• Assuming they can identify an object in a scene, they
can measure, the relative angle between two objects
With known rotation and translation of a camera, you also have directional information
Still need to bypass the correspondence problem between camera views
Some Camera Background
World Coordinates Origin
Camera Coordinates Origin
vu
Image Coordinates
Z
X
Y
X
Y
Z
Each camera is characterized by a 3 x 3 rotation matrix R and a 3 x 1 TranslationMatrix T
w(x,yz)
Background: Camera Attributes
Each camera is characterized by:1. Its 3-D coordinates (x,y,z)2. A 3 x 3 rotation matrix R3. A 3 x 1 translation matrix TWorld to camera coordinates are related by
This also applies to transformation between camera coordinate systems.
TRww
World coordinatesCamera coordinates
Background: Camera Errors and Constraints
vuO
Y
X
Zf
z
y
z
x
w
w
f
v ,
w
w
f
u
zw
yx ww ,
TRww
)()(
)()(
yTyz
Tz
xTxz
Tz
TwRfTRv
TwRfTwRu
Basic World to Image
Equations:
Background: Errors and Constraints
vu
Image Coordinates
O
Y
X
Zf
Camera measurement precision is a function of pixel resolution and viewing angle
Error= viewing angle/pixels Each node observation is a vector Each pair of vectors forms a constraint
Problem Statement
Camera Node
Ultrasound Distance Node
Radioless Tag
Given N sensor nodes: t1,t2,t3,…,tN
A subset of the nodes: m<N, t1,t2,…,tm have cameras
A subset of inter-node distances are known Goal:
• Compute 3-D coordinates for all nodes• Compute rotation and translation matrices, R and T for all
camera nodes
Camera as a Sensing Modality
vu
Image Coordinates
O
Y
X
Zf
The 3-D location w of each node is mapped to a 2-D location (u,v) on the image plane.
Each node observation is a unit vector originating at camera’s location and pointing towards node’s 3-D location w.
Each pair of unit vectors forms a constraint. Camera measurement precision is a function of pixel resolution and
viewing angle Error= viewing angle/pixels
w(x,y,z)
Camera Basics
World Coordinates Origin
Camera Coordinates Origin
vu
Image Coordinates
Z
X
YX
Y
Z
w(x,y,z)
Each camera is characterized by: Its 3-D coordinates (x,y,z) A 3 x 3 rotation matrix R A 3 x 1 translation matrix T
World to camera coordinates:
TRww
Camera coordinates World coordinates
Need something lightweight with two cameras
If you could localize nodes using a pair of overlapping camera views then you could use that to create a 3-D coordinate system
If relative R and T are known• Can transform among coordinate systems
o Can form a chain of cameras and consider multihop measurements
So what can you really do with two cameras?• Measured Epipoles (ME)
• Estimated Epipoles (EE)
Camera Epipoles
Epipoles: the points that intersect the image plane on a straight line between two camera centers
C C’
x
epipoles
e e
Epipolar plane
Camera Background
BB
CC
VVab ab
VVbaba
nnaa
nnbb
VVbcbc
VVacac
llab ab
llbc bc
llacac
AA
)vv(R vv
vRv
bcbaabacab
baabab
Tbaab
baba
bbabbab
aabaaba
baba
bcbaabacab
baabab
)(RRR
RRR
)]n(v n v[R
)]n(v n [vR
nRn
)v(vR vv
vRv
From C. Taylor
The points where the unit vectors Vab and Vba intersect with the image planes of cameras A and B respectively are called epipoles
Camera Background (Taylor’s Algorithm)
BB
CC
VVab ab
VVbaba
nnaa
nnbb
VVbcbc
VVacac
llab ab
llbc bc
llacac
AA
0vl)v(Rlvl accabcabbcabab
Given Rab, all the distances can be computed up to a scale Given a single Euclidean
distance, all Euclidean distances can be computed
Estimating the Epipoles
0Fxx'T
What if the two cameras cannot see each other? Assuming that there are at least 8 points in the common field of view of the two cameras, the epipoles
of both cameras can be estimated using the Fundamental matrix (8-point algorithm) The fundamental matrix F relates camera’s A image coordinates x to camera’s B image coordinates
x’ as follows:
This produces an over-constrained linear system of equations. The epipoles e’ and e for the two cameras satisfy the following equations:
Knowing F we can compute estimations for e’ and e. Using the estimated epipoles and the previous formulation proposed by Taylor we can compute the
rotation matrix between the two cameras and all the node-to-camera distances up to a scale 0Fe'T 0Fe
How good are the estimations of the epipoles?
Experimental Results (Indoors)
Estimated epipoles produce inaccurate results Note that the overestimated distances by camera A are underestimated by
camera B and vice versa! When the two cameras can view each other the results are extremely accurate.
Camera as a measurement modality is very accurate!
Refining Estimated Epipoles
Stratified reconstruction (traditional approach in Vision) – too complex for small devices
Alternative formulation
Camera Node
Ultrasound Distance Node
Radioless Tag
Given N sensor nodes: t1,t2,t3,…,tN
A subset of the nodes: m<N, t1,t2,…,tm have cameras
A subset of inter-node distances is known Goal:
Compute 3-D coordinates for all nodes Compute rotation and translation
matrices, R and T for all camera nodes
Refining the Estimated Epipoles
2ajajaiaiij)l,(l ||) vl - vl || - (lmin L ajai
knownlij
Taylor’s algorithm can be applied exactly in the same way The computed distances can be refined by minimizing the following set of
equations:
Can we always minimize this set of equations? NO! Minimization is possible only when there are n known edges among n different nodes and
each one of the n nodes appears in at least 2 different known edges. What is the minimum number of known edges for which L can be minimized?
3. In this case the nodes form a triangle 3 nodes 3 known edges (the edges of the triangle formed by the nodes) Each node appears in at least 2 different edges.
All the distances from the camera nodes to the nodes forming the triangle can now be refined!
Experimental Results
Indoors
Outdoors
Some Rigidity Issues(Slides contributed by Brian Goldenberg)
Physically: • Network of n regular nodes, m beacon nodes existing in space at
locations: {x1…xm,xm+1,…,xn}• Set of some pairwise distance measurements
o Usually between proximal nodes (d < r ) Abstraction
• Given: Graph Gn, {x1,...,xm}, edge weight function δ• Find: Realization of the graph
Remove one edge…and the problem becomes unsolvable
When can we solve the problem?
Given: Set of n points in the plane, Distances between m pairs of points.Find: Positions of all n points………subject to rotation and translationssubject to rotation and translations
1:a
d
b
c
2a: 3a:
ad
a
c
2b: 3:
Discontinuous deformation
a
b
ef
c
d
a
b
e
f
c
d
Discontinuous non-uniqueness: - Can’t move points from one configuration to others while respecting constraints
flip
somethingelse
Continuous deformation
Continuous non-uniqueness: -Can move points from one configuration to another while respecting constraints-Excess degrees of freedom present in configuration
Partial Intuition, Laman’s Condition
Total degrees of freedom: 2n
How many distance constraints are necessary to limit a formation toonly trivial deformations?
==How many edges are necessary for a graph to be rigid?
Each edge can remove a single degree of freedom
How many edges necessary?
Rotations and translations will always be possible, so at least 2n-3edges are necessary for a graph to be rigid.
Is 2n-3 edges sufficient?
n = 3, 2n-3 = 3
yes
n = 4, 2n-3 = 5
yes
n = 5, 2n-3 = 7
no
Need at least 2n-3 “well-distributed” edges.
If a subgraph has more edges than necessary, some edges are redundant.
Condition for rigidity
Purely combinatorial characterization of generic minimal rigidity in the plane:
2n-3 edges necessary for rigidity, and:Laman’s condition:Laman’s condition:
A graph G with 2n-3 edges is rigid in two dimensions A graph G with 2n-3 edges is rigid in two dimensions if and only if no subgraph G’ has more than 2n’-3 if and only if no subgraph G’ has more than 2n’-3 edges.edges.
Laman’s condition is a statement that any rigid graph with n vertices must have a set of 2n-3 well-distributed edges. Analogs are necessary for rigidity in any dimension
G. Laman ‘70 * n’ is the number of nodes in the subgraph G’
Illustration of Laman’s condition
n=5, m=10 > 2n-3=7
There must be at least 3 redundant edges
remove 3 edges
too many edges inred subgraph!
remove 3 edges
2n-3 well-distributed edges
Unique Graph realizability
a e
b
f
c
d
ac
b
e
d
f
G must be 3-connected
G must be redundantly rigid:It must remain rigid upon removal of any single edge
G must rigid
Solution:
Global Rigidity
B. Hendrickson ’95, A. Berg and T. Jordan ‘02
A graph has a unique realization iff it is redundantly rigid and 3-connected
Network Localization Problem (cf. Graph Realization)
Given: Set of n points in the plane, Positions of k of them, Distances between m pairs of points.Find: Positions of all n points.
node with known position (beacon)
node with unknown position
distance measurement
Illustration:
Is the problem solvable?
Problem: By looking only at the graph structure, we neglect our a priori knowledge of beacon positions
Solution: The distances between beacons are implicitly known! • By adding all edges between beacons to network graph, we get the Grounded
Graph, whose properties determine generic solvability By augmenting graph structure in this way, we fully capture all constraint
information available in the graph itself.
is this localizable?
5
4
1
2
3 1
2
3
4
5{x4, x5}
If it is, then I can use {x1,x2,x3} and δ to get
the answer
LANS
its some sort of democratization of the nodes of the graph! We capture fully the fact that the beacons have known positions solely in the logical structure of the graph. At first, I thought this was rather trivial, but upon some thought, it is seeming more and more subtle.
LANS
now lets see the graph properties that lead to unique realizability
Conditions for localizability
Instead of this graph… this one is relevant!
A network is localizable if its grounded graph is globally rigid
Degenerate cases fool abstraction
2
1
3
4
4
2
13
2
1
3
4
{x1, x2, x3}
{d14, d24, d34}
probability 1 case:
probability 0 case:
first case: {x4}second case: ???
2
1
3
?
?
In general, this network is uniquely localizable.
In degenerate case, it is not:The constraints are redundant.
Algorithms for global rigidity
Triconnectivity – well studied Rigidity Testing
• 1985: First polynomial time algorithm• 1988: Matroid Sums - O(n2)
Redundantly Rigid Component Discovery• 1995: Pebble Game – O(n2)
Discovering Localizable Nodes
Nodes in redundantly rigid triconnected components (RRTs) containing 3 beacons are uniquely localizable
To identify RRT components, first extract triconnected subgraphs On triconnected subgraphs, discover RR components using
algorithm from computational physics called “the pebble game” (details in paper)
Why is RRT important
Nonlocalizable nodes are mislocalized if included as input to most localization algorithms
Figure compares using MDS over entire network with using MDS on localizable portions and rough estimation for nonlocalizable nodes
Large mislocalization errors under MDS
Localization
5
4
1
2
3
1
2
3
4
5
Decision problem Search problem
Rigiditytheory
Does this have aunique realization?
Yes/No
Grounded
graph 1
2
3
4
5
{x1,x2,x3}
{d14, d24, d25, d35, d45}
This graph has aUnique realization.What is it?
???
{x4,x5}
This problem is in general NP-hard
Robust Quadrilaterals(D. More et. al, SenSys 2004)
More robust method for enforcing rigidity conditions
Real implementation based on MIT’s Cricket nodes
More things not covered here
More probabilistic methods for localization Localization using distance reconstruction Localization using angles Secure Localization More recent localization technologies
• Ultra-wide band systems, Radio Interferometry
Localization Conclusions
Localization is still a very challenging problem Lots of good theoretical work contributed
• Majority of the solutions rely heavily on some underlying assumption about the technology
• Most of this technology is still not in place. Majority of pressing issues are related to technology
• How do you measure distances and angles reliably in the presence of obstacles, interference and adversaries?
• How can you do that on a small energy and HW budget? Algorithmic problems and theoretical issues still exist
• Many systems will do OK with existing algorithms, physical layer would be the most pressing component missing
The process of securing localization still has a long way to go….• Better focus on a specific application domain first. A one fits all solution