Beruflich Dokumente
Kultur Dokumente
40
915MHz band. The platform also incorporates power man- FOS provides a priority-based, non-preemptive (coop-
agement circuitry to allow solar charging of batteries in the erative) threading environment with separate stacks for
case of outdoor use. The device also incorporates a real- each thread, which has the advantage of providing a sim-
time clock chip to reduce micro-controller overheads. The ple concurrent programming model which does not require
architecture relies heavily on the SPI bus where the At- semaphores. The scheduler is also responsible for CPU
mega128 acts as the SPI master and can communicate with power management and enters the lowest mode consistent
the radio, the flash memory, the real-time clock over the SPI with thread resource requirements. Time-critical operations
interface. The real-time clock and the radio can both inter- such as analog data sampling or high-speed timers are han-
rupt the Atmega128 to signal alarms, packet transmission dled by interrupt-level callbacks.
and packet reception.
An additional feature of the Fleck-platform is the ability 3. Network Topology and Protocol
to easily expand by adding additional sensors or sensor in-
terface boards, as well as coprocessor boards such as a DSP.
This section describes the network backbone for the lo-
For this particular application we have developed a custom
calisation system. The network consists of two key parts,
inertial movement unit (IMU) daughterboard which stacks
referred to as the static and the mobile networks. The static
directly on top of the Fleck-3 as shown in Figure 1.
network is comprised of the “seed” nodes — i.e. the nodes
which act as anchor points for the entire system. The mo-
bile network comprises the nodes which are to be localized.
They are designed to be worn by persons moving through
a building and are comprised of the Fleck-3 platform and
IMU daughterboard as described in Section 2.1 and shown
in Figure 1. The overall system architecture of these static
and mobile nodes is illustrated in Figure 2.
41
23345-&4359
!"#$"%
!&F*)35-&43
;('B3,5 #:;5<; =>5'&118
$*8,3-3/5 83/*() 23345-&4356
C*8D)(AE $&'()*+(,*&-
C(,(F(83 ").&/*,01 %(835-&43
;('B3,5
2*1G)(,&/5 ?(,3@(A
23345-&4357
1
until a seed node is available for upload. It should be men-
0.8
tioned here that an overflow of the buffer is not an issue in
model
our system. The buffer has the ability to store 100 walking
delivery rate
−10dB
0.6
10dB step events. Assuming a step frequency of about 1-2Hz it
6dB
0.4 −2dB allows the person to walk around for about a minute without
0.2 connection to the static network, before the buffer is full. In
0
that case old events are overwritten.
0 1 2 3 4 5 6 7 8 9
distance [m] The delivery of packets from mobile to static nodes is as-
sured using an implicit acknowledge mechanism. If a static
Figure 3. Illustration of observed radio node receives a packet from a mobile node and forwards
throughput between a static wall nodes and it back to base, the forwarded packet is also received by
mobile node. the mobile node and triggers the upload of the next packet.
In this way we guarantee a maximum transmission rate
between the mobile and the static network with minimal
packet overhead.
due to occlusion by walls and the body of the person. We
use a modified sigmoid function to model the transmission Figure 4 shows the main threads running on a mobile
probability depending on the distance: node. Every event generated from the mobile sensor data
is locally stored in an event queue. At the same time, the
1 mobile node beacons at a predefined interval, using the first
f (d) = 1 − (1) event in the queue as a beacon message. If a nearby static
1 + e−σr (d−rˆt )
node receives this message, it forwards it back to base. The
where rˆt is the nominal range and σr is an arbitrary param- forwarded message is also received by the mobile source
eter for range noise. This model is later used to represent node, which takes this as notification of packet delivery and
uncertainty of the proximity based position information in keeps on going with the next event in the queue if there is
the localisation algorithm. any.
The local buffering of event data introduces an inherent
3.2 Mobile Network Design latency between the occurrence of a relevant motion event
and the availability of that event information to the localisa-
The mobile network, comprising nodes worn by people, tion algorithm running at the network base. The amount of
has two important roles. Firstly, the number of data packets latency depends on the density of the seed nodes, the link
being transmitted from the mobile nodes needs to be min- quality between mobile and static nodes and the beaconing
imized to avoid congestion in the static network and save interval for broadcasting data messages. This is discussed
energy. This is achieved by preprocessing the motion data in more detail in Section 6.2.
onboard the device to reduce to a minimal number of rele- The localisation algorithm makes use of proximity infor-
vant events. This process is described in detail in Section 4. mation, only for events which occur within the proximity of
Secondly, these events have to be reliably transferred to the a static node. As such, the position measurement given by
static network despite the reduced connectivity, mentioned the forwarding seed node can not automatically be assigned
above. This is achieved by the use of a delay tolerant net- to all motion events in the data message — only those which
work (DTN) [33] protocol where data is buffered locally actually occur within the vicinity of the seed node. To han-
42
ure 5 illustrates the functionality of the step detection. The
!"#$%&'(%$#)# 3")%4!5)%","-)%+!.6%20"0"
Original accelerometer magnitude signal
2000
1800
200
Filtered signal
100 Step detecton
−100
43
the gyroscope data, but the magnetometers have still enough 5.1. Recursive Bayesian Estimation
influence to compensate for gyroscope drift and unknown
starting angles. A detailed analysis of the accuracy of the Bayesian estimation methods are widely used to estimate
heading estimation is beyond the scope of this paper, how- a systems state based on noisy sensor information. We sum-
ever in Figure 6 we show its functionality in a critical sit- marize the concepts here, for detailed overview with a focus
uation. In this example, a person wearing a mobile node on localisation, see [29].
In Bayesian estimation methods, the state xk of system
at a time tk is estimated by calculating the a posteriori prob-
350
ability p(xk |z0 , . . . , zk ), where z0 , . . . , zk are all measure-
300
ments about the system up to tk . The a posteriori probabil-
heading [degree]
250
ity can be calculated using Bayes’ Rule:
200
strong strong
magnetic magnetic p(z0 , . . . , zk |xk )p(xk )
150
disturbance disturbance p(xk |z0 , . . . , zk ) = , (3)
p(z0 , . . . , zk )
100
true
50 complementary filter which can be transformed into a recursive equation:
magnetometer only
0 p(zk |xk )p(xk |z0 , . . . , zk−1 )
0 5 10 15 20
time [sec]
25 30 35 40
p(xk |z0 , . . . , zk ) =
p(zk |z0 , . . . , zk−1 )
= η · p(zk |xk )p(x̃k ). (4)
Figure 6. Illustration of the heading esti-
mation based on magnetometer and gyro- where η = 1/p(zk |z0 , . . . , zk−1 ) is a normalization factor.
scopes. The term p(zk |xk ) is called the measurement model and re-
lates the state of the system to its observations, including
knowledge about uncertainties of the sensors. In our im-
plementation the system variable will be the position of a
walks around two 90 degree corners, passes a large mag- mobile node and the observation will be the position of a
netic field disturbance, then turns around and walks back nearby seed node.
the same way. The actual (ground-truth) heading is indi- The term p(x̃k ) = p(xk |z0 , . . . , zk−1 ) is called the a
cated by the wide grey line. From the Figure 6 we can ob- priori probability and is calculated as:
serve two things. Firstly, the estimator starting with a wrong Z
heading angle converges correctly towards the true heading. p(x̃k ) = p(xk |xk−1 )p(xk−1 |z0 , . . . , zk−1 )dxk−1 (5)
This takes about 10 seconds in the example case. Secondly,
the estimator is not affected by the strong magnetic field, The term p(xk |xk−1 ) is called the process model and it de-
disturbing the magnetometer readings. We show in later scribes the knowledge (including uncertainties) about the
simulations, that the accuracy of the localisation algorithm evolvement of the system over time. In our implementation
is not critically dependent on the accuracy of the heading this model will incorporate odometry data from the mobile
data. node as well as information about the indoor environment.
Starting from an initial probability p(x0 ), each recursive
update of a Bayesian filter, as described above, can be per-
5. Localization Algorithm formed in two steps:
Prediction The a priori probability is calculated from
This section describes the algorithm which is used to per- the last posterior probability using the process model
form the localisation of mobile nodes. Our system is based (Equation 5).
around the use of Monte Carlo filtering, which has been Correction The new a posteriori probability is calculated
used in some previous localisation work [13, 25]. Our algo- using the prior and the measurement model (Equation
rithm extends previous work however, by combining three 3).
key pieces of information in the filter being:
If the probability distributions are Gaussian and the models
are linear, a Kalman Filter provides an optimal and efficient
• Proximity information from static seed nodes
implementation of Bayesian filters. If the models are non-
linear, they are often linearized using the Extended Kalman
• Improved mobility information derived from onboard Filter. In the Unscented Kalman Filter the probability dis-
inertial sensors tribution is represented by a carefully chosen set of points
in the state space, which conserves the Gaussian properties
• Indoor map information of the distribution under a non-linear transformation.
44
5.2. Sequential Monte Carlo Filters with random speed. The speed is equally distributed be-
tween 0 and a maximum velocity vmax . Other models as-
In Monte Carlo sampling based techniques, a probability sume a constant velocity and model the uncertainty by in-
distribution p(x) is represented by a number of N weighted ducing a small random acceleration in every prediction step.
samples x[i] , i = 1..N, with weights w[i] as: In that case the state variable also contains the speed of the
person.
A key advantage of the system described in this paper is
X
p(x) ≈ w[i] δ(x[i] − x). (6)
i
that we determine information about the motion of a person
from local inertial sensors. Every step that is detected, com-
Sequential Monte Carlo filters are also called Particle Fil- bined with the heading measurement, provides information
ters, where samples are referred to as particles. Starting about a relative position change. This data can then be in-
from the initial probability p(x0 ), which may be repre- corporated into the prediction model as control inputs uk .
sented as equally distributed samples with equal weights, The other source of information we utilize during the
the recursive update is performed as follows: prediction step is the indoor map of the building. A map
[i] [i] provides a natural constraint on the paths a person can take
In the prediction step every sample (xk−1 , wk−1 ) of
the a posteriori distribution p(xk−1 |z0 , . . . , zk−1 ) is re- and allows the motion model to check if a certain step is
placed by a new sample according to the process model possible — i.e. a person can’t walk through walls! Our
[i]
p(xk |xk−1 ), which leads to a new set of samples (x̃k , w̃[i] ) motion model in the prediction step (Eq. 5) is then written
representing the a priori distribution. as p(xk |xk−1 , uk , m), where m is the map. In our case,
In the correction step, the weight w[i] of every sample of the control consists of the heading angle φk and the stride
the a priori distribution, is updated according to the mea- length lk . To simplify the motion model, the stride length
surement model: is currently not measured and therefore assumed to be con-
stant.
[i] In the actual implementation of the localisation algo-
X
w[i] = w̃[i] · p(zk |x̃k ), w[i] = 1. (7)
i
rithm, discussed in Section 6.2, every sample representing a
possible position of the mobile node is recalculated as soon
Together with the normalized weights, the prior set of sam- as a new radio packet, indicating a detected walking step, is
ples now approximates the a posteriori probability. In the received at the base node as:
last step (resampling), a new set of samples is drawn with [i] [i] [i]
!
px px [i] sin( φ̂ )
replacement from the prior set with the probability of a sam- = + ˆlk · k
[i] (8)
ple being drawn given by its weight factor. The final set py k py k−1 cos(φ̂k )
represents the new posterior as well, but now the samples
As mentioned earlier, it is important to introduce noise into
are equally weighted. At this point we can start with a new
the model to avoid degeneracy of the filter. This is done in
prediction step.
a natural way due to the uncertainties nφ , nl of the heading
During the resampling, unlikely samples are omitted
and stride length information as:
which leads to many duplicates in the final set. To avoid
ending up with only a single sample, noise is introduced ˆl[i] = lk + n[i] ,
[i]
nl drawn from p(nl ) (9)
k l
during the prediction step to separate samples that have the [i] [i] [i]
same values. φ̂k = φk + nφ , nφ drawn from p(nφ ) (10)
The knowledge about the map is incorporated by setting the
5.3. Monte Carlo Localization weight of a samples to zero, if it crossed a wall during the
prediction step (Equation 8).
In this section we describe the implementation of the par- [i]
ticle filter we use to estimate the position of a mobile node [i] wk−1 no wall crossed in (8)
w̃k = (11)
within a static network of seed nodes. The role of the pro- 0 wall crossed in (8)
cess model is to describe the transition from a state xk−1 at In the correction step we utilize any available informa-
the time tk−1 to state xk at the time tk . In our case the sys- tion about the proximity of the mobile node to a static seed
tem is a person wearing a mobile node and its state is given node. This can be seen as an absolute position measure-
by its position xk = (px py )T (tk ). Thus the process model ment (position of the forwarding node) with a large error
is given by a motion model, which describes our knowledge (transmission range of the mobile node). The measurement
about the movement of the person. model in (Equation 7) can be written as:
A common model, which also has been used in [13, 25],
[i] [i] [i]
is to assume that a person is moving in random directions p(zk |x̃k ) = p(Xn |x̃k ) ∝ f (d[i]
n ), d[i]
n = Xn − x̃k (12)
45
where XN is the position of the forwarding static node and
f (d) is a function describing the transmission probability 50
seed nodes
between the mobile and the static node depending on their
distance. In the simplest case this could be defined as: 40
1 d ≤ rˆt
f (d) = (13)
0 d > rˆt 30
y position [m]
irregularities discussed earlier, we chose the more appropri- 20
ate model given in Equation 1. The correction step is only
executed, if the step was detected within the range of a static
10
node, otherwise the filter continues with the next prediction
step leading to further spreading of the samples.
After updating and renormalising the weights (Equation 0
7), the resampling step follows, where a new set of samples
is drawn with replacement using the weights as a drawing
−10
probability. The weights are set to an equal value and the
position of a person is estimated using the mean of all the −10 0 10 20 30 40
samples. The spread of the samples (e.g. the standard de- x position [m]
viation in both dimensions) is then a measure for the confi-
dence of the estimation.
Figure 7. Example path generated by the sim-
ulator
6. Evaluation
tion error. mission direction, but we apply the model to irregularities over time.
46
14
for the reconstructed path. 13
seed density
7
racy of the location estimation. Since the benefit is expected 6
5
to be larger for lower seed densities, we plot the rms error 4
3
2
and the maximum estimation error against the seed distance. 1
0
This comparison is shown in Figure 8. The random motion 0 2 4 6 8 10 12
seed distance [m]
14 16 18 20
10
Figure 9. Seed density depending on the
rms estimation error [m]
10
15 seed distance 7.5m
10m
8
4
0
0 2 4 6 8 10 12 14 16 18 20
seed distance [m]
2
0
Figure 8. Estimation error for different mod- 0 100 200 300 400 500 600 700 800 900 1000
number of particles
els with varying seed distance.
Figure 10. Estimation error for different seed
model is comparable to the model used in [25] and already distances, varying the number of particles in
described in Section 5.3. The moving direction is assumed the localisation algorithm.
to be unknown and the moving distance per time step is re-
stricted by a maximum velocity. As expected, the rms error
increases with increasing seed distance. We also notice a It is observed that the estimation accuracy does not im-
saturation of the error towards higher distances. This is due prove for particle numbers bigger than 400 if the node dis-
to the fact, that the paths generated by the simulator are in a tance is up to 15m. Smaller particle numbers may give rea-
constrained area, which limits also the maximum estimation sonable results in most cases, but have an influence on the
error. robustness of the filter. This can be seen clearly in the next
The heading based motion model is shown with and simulation shown in Figure 11.
without incorporating the map information. The error still
increases, however much less than with the random direc- 8
tion model. By also incorporating the map information, the
estimation error [m]
50 particles
6 "kidnapped"
250 particles
error is nearly constant up to a seed distance of 20m. The 500 particles
4
solid line shows the error with perfect data, that means the
2
heading measurement is assumed to be known exactly.
The variation of the seed distance also changes the seed 0
0 20 40 60 80 100 120 140 160 180
step number
density, which is defined as the average number of seed 15
particle deviation [m]
47
In Figure 11, the initial position of a person (step 0) is
unknown and particles are therefore spread over the whole
map. The filter, using 500 particles, converges to the true
6
position after about 10 steps, while the cases with 250 and 9
about 10 steps to find the new position. The lower graph real path
seed nodes
of Figure 11 shows the spreading of the particles, which is reconstructed path 7
start
calculated by adding the squared standard deviations of the reconstructed path
without map
8 stop
2
seed node positions and the building map. The dotted line
1 shows the reconstructed path without using indoor map in-
0 0 % packet loss formation.
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
10 % 0.9 1
degree of irregularity
20 % Since we do not have a reference system, we manually
12
50 %
created the true path as a polygon and defined the error met-
max estimation error [m]
10
8 ric as the closest distance to the closest polygon segment.
6 This metric results in a slight underestimation of the true
4
2
error, since it only gives the deviation orthogonal to the
0 path and not along the path. However, as can be seen in
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
degree of irregularity Figure 13, this is a reasonable estimate if the reconstructed
path is not varying much from the true path. Plot 14(d)
Figure 12. Estimation error for different
packet losses, varying the transmission ir- a) 10
regularity.
latency [sec]
0
0 20 40 60 80 100 120
observe that the estimation error is nearly independent of b) 10
the mobile and the seed node. This demonstrates a key ad- 4
0
on correct radio propagation models like RSS based algo- c) 2
0 20 40 60 80 100 120
rithms. 1.5
confidence [m]
0
0 20 40 60 80 100 120
d) 4
1
seed nodes were deployed on various positions throughout 0
0 20 40 60 80 100 120
an area in the building as shown in Figure 13. The distance time [sec]
48
sition estimation, which is given by the standard deviation 7. Conclusions and Future Work
of the particles in the x and y-direction.
7.1 Conclusions
Latency is another important metric to evaluate in sen-
sor networks. For this application we define latency as the
We have presented an indoor localisation system for a
time between the occurrence of a walking step and the a
wireless sensor network, which estimates the position of
successful transfer from the mobile node to the static net-
mobile nodes worn by persons moving through a network
work. The transfer time from a seed node to the base node
of static seed nodes with known positions. Relevant infor-
is assumed to be virtually zero since the static nodes form a
mation about the motion of the person is generated by the
fully connected network and were running at a 100% duty
mobile nodes and transferred to a central PC, where they
cycle for these experiments. The latency thus shows how
are combined with knowledge about the indoor environment
long an event is stored in the buffer of the mobile node and
and the seed node positions using Monte Carlo based esti-
depends both on the time of disconnectedness of the mobile
mation algorithms. Data packets are processed as soon as
node and on the retry time after an unsuccessful upload.
they arrive at the central PC, which enables real-time track-
Plot 14(a) shows that the latency is about a second most
ing of a person while moving through a building.
of the time, with a maximum value of 8 seconds during a
We have evaluated the performance of our localisation
longer time with bad connectivity. Plot 14(b) shows which
system through both simulations and experimental valida-
mobile node has been used in the measurement model of
tion. Our simulations show that the algorithm is robust
the localisation algorithm. A node ID of zero means, that
against radio transmission irregularities and heading esti-
the step was detected out of range of any seed nodes.
mation errors which occur from sensor noise and magnetic
The indoor walking experiment was repeated over a disturbances. In a deployment in our office building we
longer time period of about half an hour, walking the same showed, that a person walking through the area could be
path multiple times. Figure 15 shows the probability density localized with a rms error of about 1m and a maximum er-
distribution for both the latency and the reconstruction error ror of about 3.5m. The network latency was about 1 second
calculated from the experiments. We define the overall error rms and 8 seconds maximum.
0.2 0.2 There are several ways to improve the system into the
0.18 0.18 with map
without map
future. One is to extend the tracking area by using multi-
0.16 0.16
hop routing techniques in the static network. Another in-
0.14 0.14
teresting avenue is to generate more information from the
0.12 0.12
0.1 0.1
IMU data than just step events and heading values. We are
0.08 0.08 working on additional research in accelerometer-based ac-
0.06 0.06 tivity classification to detect events such as sitting, lying or
0.04 0.04 falling, which allows the system to be used for health care
0.02 0.02
applications like monitoring of elderly people in care facil-
0 0
0 2 4 6
latency [sec]
8 10 0 1 2 3 4
estimation error [m]
5 6 ities. Future work will also investigate ways to estimate the
step length of a walking person to improve localisation ac-
curacy as well as detect new types of events such as stair
Figure 15. Empirical probability density dis-
climbing which then enables position estimation over mul-
tribution for latency and estimation error cal-
tiple floors.
culated from multiple experiments.
Another direction for future work is decentralization of
the system. In the current implementation the localisation
algorithm is running on a central computer, where all the
data from the network are is collected via the base node. In
and latency as the standard deviation from zero. The overall principle however, every mobile node has all the informa-
latency is 1.3 seconds with a maximum value of 8 seconds. tion available that is needed to estimate its own position. In-
Assuming an uncertainty of 1m on the true path, the overall stead of forwarding a mobile node event message to the base
estimation error with an algorithm using the map informa- node, a seed node would simply send an acknowledge mes-
tion is 1.2m with a maximum of about 2.5m, and 2m with a sage containing also its own position information. Given
maximum of 5.2m without the map. This is consistent with this scenario, other nodes of the mobile network could also
the simulation results in Figure 8. be used as seed nodes.
49
References [17] B. Krishnamachari. Networking Wireless Sensors. Cam-
bridge University Press, 2005.
[1] P. Bahl and V. N. Padmanabhan. Radar: An in-building [18] L. Lazos and R. Poovendran. Serloc: Robust localisation
rf-based user location and tracking system. In INFOCOM, for wireless sensor networks. ACM Transactions on Sensor
2000. Networks, 1(1):73–100, 2005.
[2] S. Beauregard and H. Haas. Pedestrian dead reckoning: A [19] L. Liao, D. Fox, J. Hightower, H. Kautz, and D. Scultz.
basis for personal positioning. In Proceedings of the 3rd Voronoi tracking: Location estimation using sparse and
Workshop on Positioning, Navigation and Communication, noisy sensor data. In Proc of the IEEE/RSJ International
2006. Conference on Intelligent Robots and Systems (IROS), 2003.
[3] S. Bhatti, J. Carlson, H. Dai, J. Deng, J. Rose, A. Sheth, [20] X. Nguyen, M. I. Jordan, and B. Sinopoli. A kernel-
B. Shucker, C. Gruenwald, A. Torgerson, and R. Han. based learning approach to ad-hoc sensor network localisa-
MANTIS OS: An Embedded Multithreaded Operating Sys- tion. ACM Transactions on Sensor Networks, 1(1):134–152,
tem for Wireless Micro Sensor Platforms. Mobile Networks 2005.
[21] D. Niculescu and B. Nath. Ad hoc positioning system (aps)
and Applications, 10(4):563–579, 2005.
using aoa. In INFOCOM, 2003.
[4] U. Bischoff, M. Strohbach, M. Hazas, and G. Kortuem.
[22] L. Ojeda and J. Borenstein. Personal dead-reckoning sys-
Constraint-based distance estimation in ad-hoc wireless sen-
tem for gps-denied envionments. International Workshop
sor networks. In EWSN, 2006.
on Safety, Security and Rescue Robotics, 2007.
[5] N. Bulusu, J. Heidemann, and D. Estrin. GPS-less low cost
[23] S. O’Sullivan. Map viewer - a mapping tool for mobile
outdoor localisation for very small devices. IEEE Personal
robotics. In http://www.skynet.ie/ sos/mapviewer/main.php,
Commucations Magazine, pages 28–34, 2000.
2007.
[6] P. Corke. Fos - a new operating system for sensor networks.
[24] A. Raj, A. Subramanya, D. Fox, and J. Bilmes. Rao-
In Proceedings of the 5th European Conference on Wireless
blackwellized particle filters for recognizing activities and
Sensor Networks (EWSN08), 2008.
spatial context from wearable sensors. In The 10th Interna-
[7] P. Corke, R. Peterson, and D. Rus. Localization and naviga-
tional Symposium (ISER 2006), 2006.
tion assisted by cooperating networked sensors and robots.
[25] M. Rudafshani and S. Datta. Localization in wireless sensor
The International Journal of Robotics Research, 24(9):771–
networks. In Information Processing in Sensor Networks
786, Oct. 2005.
(IPSN), pages 51–60, 2007.
[8] C. Duffy, U. Roedig, J. Herbert, and C. J. Sreenan. [26] P. Sikka, P. Corke, P. Valencia, C. Crossman, D. Swain, and
An Experimental Comparison of Event Driven and Multi- G. Bishop-Hurley. Wireless adhoc sensor and actuator net-
Threaded Sensor Node Operating Systems. In Pro- works on the farm. In IPSN, pages 492–499, 2006.
ceedings of the Third IEEE International Workshop on [27] R. Stoleru, T. He, and J. A. Stankovic. Walking gps: A prac-
Sensor Networks and Systems for Pervasive Computing tical solution for localisation in manually deployed wireless
(PERSENS2007), White Plains, USA. IEEE Computer So- sensor networks. In EmNets, 2004.
ciety Press, Mar. 2007. [28] R. Stoleru, P. Vicaire, T. He, and J. A. Stankovic. Stardust:
[9] D. Fox, W. Burgard, F. Dellaert, and S. Thrun. Monte carlo A flexible architecture for passive localization in wireless
localization: Efficient position estimation for mobile robots. sensor networks. In SenSys’06, 2006.
In Sixtenth National Conference on Artificial Intelligence [29] S. Thrun, W. Burgard, and D. Fox. Probabilistic Robotics.
(AAAI-99), 1999. MIT-Press, 2005.
[10] G. Giorgetti, S. K. Gupta, and G. Manes. Wireless localiza- [30] T. Wark, P. Corke, P. Sikka, L. Klingbeil, Y. Guo, C. Cross-
tion using self-organising maps. In IPSN, 2007. man, P. Valencia, D. Swain, and G. Bishop-Hurley. Trans-
[11] T. He, C. Huang, B. M. Bium, J. A. Stankovic, and T. Ab- forming agriculture through pervasive wireless sensor net-
delzaher. Range-free localization schemes for large scale works. IEEE Pervasive Computing, 6(2):50–57, 2007.
sensor networks. In ACM MobiCom’03, 2003. [31] T. Wark, C. Crossman, W. Hu, Y. Guo, P. Valencia, P. Sikka,
[12] T. He, S. Krishnamurthy, J. A. Stankovic, T. Abdelzaher, P. Corke, C. Lee, J. Henshall, J. O’Grady, M. Reed, and
L. Luo, R. Stoleru, T. Yan, and L. Gu. Energy-efficient A. Fisher. The design and evaluation of a mobile sen-
surveillance system using wireless sensor networks. In Mo- sor/actuator network for autonomous animal control. In
biSys, 2004. IPSN, pages 206–215, 2007.
[13] L. Hu and D. Evans. Localisation for mobile sensor net- [32] P. Wilson, D. Prashanth, and H. Aghajan. Utlizing RFID
works. In Tenth International Conference on Mobile Com- signaling scheme for localization of stationary objects and
puting and Networking (MobiCom), 2004. speed estimation of mobile objects. In IEEE International
[14] J. Karlsson, T. Wark, P. Valencia, M. Ung, and P. Corke. Conference on RFID, 2007.
Demonstration of image compression in a low-bandwidth [33] Z. Zhang. Routing in intermittently connected mobile ad
wireless camera network. In IPSN, 2007. hoc networks and delay tolerant networks: Overview and
[15] B. Karp and H. Kung. Gpsr: Greedy perimeter stateless challenges. IEEE Communications Surveys and Tutorials,
routing for wireless networks. In MobiCom, 2000. 8(1):24–37, 2006.
[16] O. Kasten and K. Römer. Beyond event handlers: program- [34] G. Zhou, T. He, S. Krishnamurthy, and J. A. Stankovic.
ming wireless sensors with attributed state machines. Pro- Models and solutions for radio irregularity in wireless sensor
ceedings of the 4th international symposium on Information networks, 2006.
processing in sensor networks, 2005.
50