IET Wireless Sensor Systems

Research Article

Robust Kalman filter-based decentralised ISSN 2043-6386

Received on 3rd May 2017
Revised 28th September 2017
target search and prediction with topology Accepted on 3rd November 2017
E-First on 4th January 2018
maps doi: 10.1049/iet-wss.2017.0039
www.ietdl.org

Ashanie Guanathillake1,2 , Andrey V. Savkin1, Anura P. Jayasumana3

1School of Electrical Engineering and Telecommunication, University of New South Wales, Sydney, NSW 2052, Australia
2Data61|CSIRO, Eveleigh, NSW 2015, Australia
3Department of Electrical and Computer Engineering, Colorado State University, Fort Collins, CO 80523, USA

E-mail: ashanie@student.unsw.edu.au

Abstract: A novel distributed approach for searching and tracking of targets is presented for sensor network environments in
which physical distance measurement using techniques such as signal strength is not feasible. The solution consists of a robust
Kalman filter combined with a non-linear least-square method, and maximum likelihood topology maps. The primary input for
estimating target location and direction of motion is provided by time stamps recorded by the sensor nodes when the target is
detected within their sensing range. An autonomous robot following the target collects this information from sensors in its
neighbourhood to determine its own path in search of the target. While the maximum likelihood topology coordinate space is a
robust alternative to physical coordinates, it contains significant non-linear distortions when compared with physical distances
between nodes. The authors overcome this using time stamps corresponding to target detection by nodes instead of relying on
distances. The performance of the proposed algorithm is compared with recently proposed pseudo gradient algorithm based on
hop count and received signal strength. Even though the proposed algorithm does not depend on distance measurements, the
results show that it is able to track the target effectively even when the target changes its moving pattern frequently.

 Nomenclature centralised computing resources [3, 4]. Our focus is on applications

that use sensor networks to monitor details of an environment,
N number of sensor nodes in the network track a target therein, and follow or rendezvous with the target with
si sensor node i a device such as an autonomous robot or even a person carrying a
[xi, yi] topology coordinates of node i mobile sensor node [5]. Vital requirements for such applications
Mn × N packet receiving binary matrix include real-time decision making related to the trajectory of a
S(di j) packet reception probability value when the target, i.e. its current location, movement and future locations, as
distance is di j well as tracking the target.
Pi(xi, yi) probability of obtaining receiving vector when Much of the work on target tracking using WSNs rely on
si is located in (xi, yi) location accurate localisation information, specifically the physical
Lt0 initial target location coordinates of the nodes as well as the ability to measure the
distance from sensor nodes to the target [6–8]. However,
Lti = [Xti, Y ti] target location at time ti
localisation of sensor nodes and distance estimation using
Lai = [Xai, Y ai] robot location at time ti strategies such as receiver signal strength measurement or time
t
Lt j(k) the calculated target location at tk with delay is not feasible in many complex and harsh environments [3].
available information at t j( > tk) Thus, novel coordinate systems have emerged that do not rely on
St set of nodes detect the target initially distance estimations in place of physical (geographical) coordinates
Nt = | St| number of nodes detects the target initially [9, 10]. Topology maps [10], which we rely on in this paper, are an
Vt speed of the target alternative to physical maps but with the distances among nodes
significantly distorted. While such coordinate systems have been
Va speed of the autonomous robot
used extensively for sensor networking protocols such as routing
Nai set of local neighbours of robot at time i and placement, they have found only limited traction for target
t set of time instances of sensor node i captured
Zsij tracking due to the fact that such systems do not provide accurate
the target until time t j physical information such as position and velocity. Although they
Gai = {G1, …, G8} grid points at time ti can be easily and accurately generated, use of such topology
Gi = [XGi, Y Gi] topology coordinates of grid point Gi where coordinates require overcoming significant non-linear distortions
i = 1, …, 8 between the physical coordinate space and topology coordinate
SVa + ε set of nodes in V a + ε neighbourhood of space. Our approach overcomes the non-linear distance distortions
in topology coordinate space compared to physical coordinates
autonomous robot at time ti
using robust Kalman filtering. However, robust Kalman filter
(RKF)-based target tracking algorithms [11, 12] require centralised
1 Introduction operation, i.e. an environment where all the information is
Decentralised target search is an important task in many wireless processed in a centralised location. As indicated above, distributed
sensor network (WSN) applications, e.g. search and rescue, solutions possess many advantages over centralised solutions in
surveillance and military operations [1, 2]. Decentralised WSN-based environments, thus we develop a distributed approach.
algorithms possess many advantages in ad hoc WSN environments, This paper presents a novel algorithm, decentralised target
which are susceptible to high link/node failures, uncertainty search based on RKF and behaviour formula extraction (DeTarSK),
associated with multi-hop packet delivery, and inaccessibility to to track and predict target locations in an environment where
accurate physical distance estimation is not feasible. Instead of

IET Wirel. Sens. Syst., 2018, Vol. 8 Iss. 2, pp. 60-67 60

© The Institution of Engineering and Technology 2017
using physical coordinates and distances, DeTarSK relies on
maximum likelihood topology coordinates [10] and corresponding
distances. The algorithm uses RKF uncertain mobility model in a
distributed manner to filter the error calculated in target's previous
locations. Then, it predicts the movement, direction and the future
locations of the target, using an approach in which RKF is
combined with non-linear least-square method [13]. The paper
considers the scenario of an autonomous robot searching and
tracking the target using DeTarSK by communicating with sensors
within its communication range.
The search considered in this paper is discrete. The robot search
is performed on an n × n grid, and its next available location is a
function of its current location. Also, the target's movement is
independent of the robot's movement and can take any form. The
autonomous robot calculates the target detection points from the
time stamp sets received from sensors in its locality. These time
stamp sets describe the times at which certain node(s) detected the
target in its vicinity. A major challenge is that the gathered
information is not current, i.e. the information gathered from a
region may be the target's past locations and moving directions, but
not their current values. Also, the received information about the
target's past behavior may not be accurate. However, this is the
information that is available for predicting target's location.
Therefore, our objective is to predict the target's current location
and calculate the shortest path to catch the target using the
available information in real time. The main contributions of this
paper are: (i) a decentralised target-search algorithm, (ii) use of
RKF in a decentralised manner to predict future target locations,
and (iii) overcoming the distortion between topology coordinates
and physical coordinates (which are unknown) for effective target
detection.
This paper is organised as follows: Section 2 reviews related
background. Section 3 presents the maximum likelihood topology
map derivation. Section 4 presents the DeTarSK algorithm,
followed by results in Section 5. Finally, Section 6 presents the
conclusions.
2 Related work
The problem of searching for targets can be characterised based on
different aspects, such as a one- or two-sided search, tracking a
stationary versus a moving target, discrete or continuous time
search, and real-time or off-line search [7, 14]. Tracking can also
be categorised by the underlying the query routing structure, e.g.
tree based [15], hierarchical cluster based [16], geometrical [17]
and hash based [18].
Tracking a target using a large-scale network is challenging due
to limitations of WSNs such as limited energy, processing and
memory resources [19]. In addition, message losses are common
and nodes are prone to failures [4]. Therefore, target searching
algorithms have to meet constraints, including energy-efficiency,
distance-sensitivity, scalability and fault-tolerance [4]. Chong et al.
[19] proposed a target localisation algorithm in visual sensor
networks based on a certainty map that described an area in which
the target was occupied or non-occupied. This algorithm aims at an
optimum solution considering complexity, energy efficiency and
robustness. Cluster-based target tracking algorithm is proposed in
[20] to minimise the energy consumption in the network by the use
of novel communication protocol they have proposed, which
reduces the packet transmission in the target tracking process.
However, the decisions are made in a central node, which requires
packet transmission from sensors to a sink in timely manner.
An algorithm for sensor selection for target tracking is
presented in [21] with low estimation error and low computation
power. Njoya et al. [22] proposed a sensor placement algorithm to
cover a target with minimum number of sensors in a reasonable
time. A sensor selection algorithm is proposed in [23] to improve
the accuracy of target localisation while optimising the energy.
Sensor allocation for target tracking in heterogeneous networks is
discussed in [24].
A generalised search for the best path selection based on NP-
complete optimal search path problem is presented in [6]. It
considers a target moving within a known indoor environment
IET Wirel. Sens. Syst., 2018, Vol. 8 Iss. 2, pp. 60-67
© The Institution of Engineering and Technology 2017
partitioned into interconnected regions. The search path problem is
modelled as a discrete search that concerns both cell connectivity
and transmit time. The extended tracking approach for an ultra-
wideband indoor sensor network proposed in [25] is based on a
Bernouli filter and considers the uncertainties of data association,
target measurement rate, detection and noise. A real-time algorithm
based on a probabilistic version of a local search with estimated
global distance is presented in [7], which has as inputs the initial
target location probabilities and the transition matrix for possible
moves of target. The decision making at each step of this search
applies probabilistic distance estimation to find the path trajectory
of the agent that minimises the average number of search steps.
A learning real-time A* algorithm for a static target was
proposed in [26] and for a moving target in [27]. Information
moving target search algorithm [28] has enhanced the MTS
algorithm with informational distance measures based Rokhlin
metric and Ornstein metric, which give the necessary distance
measures.
Locating stationary target in an ultra-wideband network using
time of arrival parameter is proposed in [29]. Also, Wang et al. [8]
proposed an algorithm to locate a single static target using Kalman
filter and a least-squares algorithm. This algorithm is based on a
distance model and an angel model. Dong et al. [30] proposed a
mobility tracking algorithm for cellular network that uses received
signal strength indicator (RSSI) measurements and velocity metrix.
Moreover, Mahfouz et al. [31] propose a moving target tracking
method based on RSSI models and a Kalman filter. The clustering-
based fusion estimation target tracking approach in [32] calculates
the filter estimators at each cluster head, and it is suitable for
sensor networks with multiple sampling intervals. Mobile object
tracking in a randomly deployed binary sensor network is
addressed in [33] with a location aware algorithm in which the
output dynamically changes according to the object's movements.
Location prediction of mobile objects in WSNs is proposed in [34].
This method is based on Gauss–Markov mobility model and
maximum likelihood technique, which assumes that sensors are
capable to measure the velocity of the target. Mobility estimation
in WCDMA network using time difference of arrival and angle of
arrival is proposed in [35]. However, the distance uncertainty
calculation based with range measurements in complex and harsh
environments limits it applicability.
Our proposed approach deviates from prior work in two ways.
First, we use a WSN topology map instead of a physical map thus
overcoming disadvantages associated with physical localisation
and distance measurements. Second, we develop a decentralised
target search algorithm. As shown in Fig. 1a, in centralised
algorithms, the sensed information has to be routed to base station,
processed, and then to the robot [20]. Thus, data is routed through
multi-hops, which increase the network traffic and it is subject to
uncertainties in the network reducing the reliability of the
algorithm. Hence, we propose a decentralised algorithm in which
robot directly communicate with its neighbours and move towards
the target with available information as indicated in Fig. 1b. Thus
decentralised algorithms overcome the disadvantages associated
with multi-hop communication in centralised algorithms.
Moreover, in a situation that links with the base station are lost,
centralised algorithms are unable to search the target any further.
However, in decentralised algorithms, as robot is mobile and using
single-hop communication, it can move around and re-store the
communication links easily. Also, if the robot is unable to function,
another robot can replace it because navigation path is completely
independent and needs only robot's neighbours information.
However, in centralised algorithms, replacing a base station is not a
feasible solution in an emergency situation. However, there is a
disadvantage relates with decentralised algorithms, which is robot
may not get latest information about the target as centralised
algorithms as it communicates only with neighbours. However, we
have overcome this with the use of a prediction method based on
RKF combined with a non-linear least-square method.
61
3 Maximum likelihood topology maps (ML-TM)
In this paper, we use ML-TM proposed in [10], which is a
representative network layout map, which more closely resembles
the physical map compared to the traditional topology preserving
maps (TPMs) [9]. Unlike TPMs that start with a hop-distance-
based coordinate system, ML-TM relies on a distance-based packet
reception probability function. A binary matrix is used to represent
the packet reception of sensor nodes at different locations of the
mobile robot's trajectory. Then the ML topology coordinates of
sensors are extracted using the probability function for packet
reception. Hence, this method does not rely solely on connectivity
[9, 36] nor range-based parameters [37].
In [10], the map is obtained by a mobile robot that traverses the
sensor field. Robot is able to receive packets from the nodes in its
vicinity. The probability of receiving a packet is sensitive to the
distance, and this property is used to generate ML-TM. The
receiving probability function, S(d) used is an intermediate model
between RSSI [38] and virtual coordinates (VC) [9]
S(d) := p0 ∀d ≤ r
S(d) := 0 ∀ d ≥ R
(1)
p (R − d)
S(d) := 0 ∀r < d < R
(R − r)
where 0 < p0 ≤ 1, 0 < r < R ≤ Rc are some given constants. Rc is
the communication range of a sensor node.
The moving robot keeps track of received packets from
different nodes at different times, using a binary matrix M, called
the packet receiving matrix. M is created as follows:
M(i, j) = 1, if the robot receives a packet from the sensor i at
the time t j;
M(i, j) = 0, if the robot does not get a packet from the sensor i
at the time t j.
In the next step, they calculate the optimal ML topology
coordinates for each sensor using M matrix and S(d) function. For
that, R − neighbourhood of the robot's trajectory is divided into
small step grid and a function called Pi(xi, yi) is introduced. This
function is the probability of obtaining the packet receiving vector
mi for node i under the assumption that the sensor node i is located
at the point (xi, yi). Moreover, to find Pi(xi, yi), they defined a
function Z(di j) as
S(di j) if mi( j) = 1
Z(di j) =
1 − S(di j) if mi( j) = 0
Fig. 1  Packet flow of target search algorithms
(a) Centralised target search algorithm, (b) Decentralised target search algorithm
62
where
di j := (xi − xR(t j))2 + (yi − yR(t j))2 (2)
Then, Pi(xi, yi) is calculated for any vertex in the grid for each
sensor node as
Pi(xi, yi) = Z(di1)Z(di2), …, Z(diN ) (3)
Finally, the vertex that delivers the maximum value of Pi(xi, yi) is
taken as the ML optimal coordinates of the sensor node i.
4 RKF-based DeTarSK
DeTarSK algorithm based on ML topology maps of a WSN is
described in this section. The objective of this algorithm is to
construct a real-time path to find a target with least possible time
using the information gathered by robot's local neighbourhood. The
basic idea of the algorithm is described in Fig. 2. This algorithm
uses standard IEEE 802.15.4/Zigbee point-to-point protocol to
communicate with sensors and robot [39]. First, we define some
keywords as follows:
Target: a robot/person enters to the network and moves around
in an independent pattern.
Autonomous robot/agent: a robot/person with a wireless mote
that is searching for the target by communicating with sensor
nodes.
Target initial detection point: location the target enters the area
covered by the WSN and is detected by sensor nodes.
Target tracking: sensor nodes sense for the target and keep track
of time instances that they have detected the target.
Target prediction: autonomous robot predicting target's future
locations based on available tracking information.
Target search: autonomous robot looking for the target based on
tracking and prediction information.
Let consider, N is the number of nodes are randomly deployed
over an environment and Nt( < N) is the number of nodes are
detected a target entering to the network at time t. Then the
detested nodes, St = {s1, …, sk}, transmit an initial message to the
sink node informing the detection. Sink node calculates the initial
target location Lt0 = (Xt0, Y t0) using (4) and sends a message to
autonomous robot including the calculated target detection point.
Since, it takes time for these message transmission, we assume that
the robot starts the search after a ta time from target detection.
Thus, robot do not have any data about target's current location or
moving direction. These information are gathered by its
neighbourhood nodes in a decentralised way as described in
following subsections:
∑si ∈ St xi ∑si ∈ St yi
Xt0 = and Y t0 = (4)
Nt Nt
4.1 Information gathering
As in [40], information about target is gathered by robot's local
neighbourhood. A message is transmitted to its V a + ε-
neighbourhood (nodes located within a circle centred at the robot s
location with a radius of V a + ε.) to request their topology
t
coordinates, transmitting power and time sets Zsij that each node has
captured the target. The ε is a value within 0 and V a /2. Then, robot
calculates its topology coordinates Lai at time ti as
|Na |
∑ j = i1 w j x j
Xai = |Na |
and
∑ j = i1 w j
|Na |
(5)
∑ j = i1 w jy j
Y ai = |Na |
∑ j = i1 w j
IET Wirel. Sens. Syst., 2018, Vol. 8 Iss. 2, pp. 60-67
© The Institution of Engineering and Technology 2017
where s j ∈ Nai and w j = (Prx j /Ptx j). Ptx j is the transmitting power of predict the target trajectory up to tK time instances, i.e. the largest
node j and Prx j is the receiving power of the message sent by node t
time instance value of all Zsij time sets that robot received.
j. The uncertainty mobility model and measurement model used
in this paper are as follows:
4.2 Target trajectory prediction
xt(k + 1) = (A + BΔ(k)K)xt(k) + w(k) (7)
In this stage, the target trajectory is predicted using the information
t
gathered in previous stage. From Zsij time sets, we can find subset y(k) = Cxt(k) + v(k) (8)
of nodes that detected the target at each time instance. Then,
t t t t t t t t t t t
Lt j(k) = (Xt j(k), Y t j)(k), the target locations at time tk with available where xt = [Xt j, Y t j, Ẋ t j, Ẏ t j]T . (Xt j, Y t j) and (Ẋ t j, Ẏ t j) are the position
information at t j ( > tk) is calculated using (6). These location and velocity of the target in topology map.
information is stored in robot memory for future trajectory Δ(k) = [γ(k) − 10; 0β(k) − 1] is the uncertain matrix. γ(k) and β(k)
prediction as well satisfy the constraint of 1 − ζ ≤ γ(k), β(k) ≤ 1 + ζ, where
t
0 < ζ < 1. Thus Δ(k) satisfy the bound Δ(k)TΔ(k) ≤ I. w(k) is the
N
t ∑i = 1 cik xi + Xt j − 1(k)nt j − 1 process noise with Q covariance that denotes the driving/
Xt j(k) = N
, acceleration command of the target. v(k) is the measurement noise
∑i = 1 cik + nt j − 1
(6) with R covariances that denote the nodes’ topology coordinates
N t error component. A, B, C and K are as follows:
t ∑i = 1 cik xi + Y t j − 1(k)nt j − 1
Y t j(k) = N
∑i = 1 cik + nt j − 1 1 0 t 0 t 0
0 1 0 t 0 t
where cik = 1, if si node is in the robot neighbourhood and detected A= , B=
0 0 1 0 0 0
the target at time tk and cik = o, if it is not in the neighbourhood or
0 0 0 1 0 0
does not detected the target at time tk. nt j − 1 is the number of nodes
t 1 0 0 0 0 0 1 0
used to calculate Ltkj − 1 at time t j − 1. C= , K=
0 1 0 0 0 0 0 1
Therefore, we have considered the RKF with an uncertainty
mobility model to remove the errors in target detection points. where t is the sampling time.
Kalman filter provides a method for constructing an optimal The non-linear least-square method is used to calculate the
estimate of the state that consists of a linear dynamical system target trajectory from the output of the robust Kalman filter. The
driven by stochastic white noise processes [41]. However, it does trajectory equation varies with time is a 3D space–time equation
not address the issue of robustness against large parameter when we consider a 2D network. Hence, we divided the space–
uncertainty in the linear process model. Therefore, a robust Kalman time equation to two equations that describe how x and y
filtering-based state estimation algorithm is used for target detected coordinates of the target varies with time and we name it as t–x
point calculation [11, 12]. curve and t–y curve, respectively. If the network is 3D, three
Since the target detected points are calculated from its current equations must be calculated namely t–x, t–y and t–z.
and previous neighbourhood information, those might have The next challenging thing is to find the degree of the t–x and t–
encountered some errors. Kalman filter has attracted attention on y curves. Let consider p and q are the degree of t–x and t–y curves,
target tracking in sensor networks in recent past because it provides t t
an optimal way of extracting a signal from noise by exploiting a respectively. Then the sign changes of Xt j(k) and Y t j(k) coordinates
state-space signal model [12]. It provides a method for constructing are considered to calculate p and q as shown below:
an optimal estimate of the state that consists of a linear dynamical
j
system driven by stochastic white noise processes [41]. However, it
does not address the issue of robustness against large parameter p= ∑ signtx k
(9)
k=3
uncertainty in the linear process model. In our problem, model
dynamics are unknown but bounded. Therefore, Kalman filter may
where (see (10)) where
lead to poor performance. RKF that focuses on the uncertainties of
the system has been proposed in [41]. RKF was applied to derive
an estimate of the mobile target's location in mobile sensor network
[11] and in Delay–Tolerant sensor networks [12], nevertheless,
these algorithms are calculated target locations in a centralised
approach. In our algorithm, we use the RKF in a decentralised
manner to remove errors in calculated target positions and to

Fig. 2  DeTarSK algorithm work flow

IET Wirel. Sens. Syst., 2018, Vol. 8 Iss. 2, pp. 60-67 63

© The Institution of Engineering and Technology 2017
 t t t t
(Xt j(k) − Xt j(k − 1)) (Xt j(k − 1) − Xt j(k − 2))
1 if t t
× t t
= −1
|Xt j(k) − Xt j(k − 1)| |Xt j(k − 1) − Xt j(k − 2)|
signtxk = ⋅
t t
and | Xt j(k) − Xt j(k − 1) | > 1
(10)
0 otherwise
j
q= ∑ signty k

k=3

final_t − x = a × int_t − x + C (15)

where, a and C are constants and

signtyk =
|TD|
int_t − x = ∏ (t − TD(i)) (16)
t t t t i=1
(Y t j(k) − Y t j(k − 1)) (Y t j(k − 1) − Y t j(k − 2))
1 if t t
× t t
= −1
|Y t j(k) − Y t j(k − 1)| |Y t j(k − 1) − Y t j(k − 2)| To calculate a and C, we can substitute two points at which the
t t
target was detected to final_t − x equation and solve them. The
and | Xt j(k) − Xt j(k − 1) | > 1 final_t − y equation also can be obtained by following came steps
0 otherwise mention above. Then, the target's current location
Lt_current = (Xt_current, Y t_current) can be obtained by substituting t j
The t–x and t–y curve equations are given as follows: value to the final_t − x and final_t − y equation.

x = a0 + a1t + ⋯ + apt p (11) 4.3 Next state of the autonomous robot

In this stage, we calculate the next point for the robot in its move
y = b0 + b1t + ⋯ + bqtq (12) towards the target as in A* algorithm [42]. The next point
calculation is based on grid search as in [40] and we consider
Equations (13) and (14) can be used to calculate a0, a1, …, ap and number of grids in its neighbourhood as eight. The grid locations
b0, b1, …, bp, which is obtained by using the least-square method Gi are based on the robot's current location Lai and it can be
calculated as
[a0 a1 ⋅ ⋅ ⋅ ap]T = inv(T x)X (13)
XGk = Xai + V acos{(k − 1)π/4}
T (17)
[b0 b1 ⋅ ⋅ ⋅ bq] = inv(T y)Y (14) Y Gk = Y ai + V asin{(k − 1)π/4}

where where k = 1, …, 8. Once the grid layout is calculated, robot finds

the grids that are free with obstacles. Then it selects one grid as its
∑1 ∑ ti … ∑ tip next point to move based on following equation used in [40]:

Tx =
∑ ti ∑ ti 2
… ∑ tip +1
F(n) = G(n) + H(n) (18)
⋮ ⋮ ⋱ ⋮
where G(n) is the cost of moving from the initial step to the next
∑ tiq ∑ tip +1
… ∑ ti p 2
step on the grid and H(n) is the cost of moving from a grid point to
the target's predicted current Lt_current location.
∑ 1 ∑ ti … ∑ tiq For the H(n) calculation, we use the Euclidean distance in the
Ty =
∑ ti ∑ ti 2
… ∑ tiq +1
ML-TM. However, the actual length of the path cannot be
calculated because obstacles can be in the way of the target. Thus,
⋮ ⋮ ⋱ ⋮ H(n) is an approximate value. The robot calculates F(n) value for
∑ ∑ tiq… ∑ ti q
tiq + 1 2
all the grid locations and chooses the grid point that has minimum
F(n) value as its next state.
X T = ∑ xi ∑ ti xi… ∑ tip xi
Y T = ∑ yi ∑ tiyi… ∑ tiqyi 5 Performance evaluation
This section discusses the performance of the DeTarSK algorithm.
The t–x and t–y equations above describe the behaviour of the MATLAB-based simulation was carried out and two network
target (e.g. changing directions) upto tk ( < t j) time. However, our setups were considered in the evaluation. One is a sparse network
requirement is to predict the target location at time t j, which cannot with 1400 sensor nodes and the other one is a circular shape
be obtained by substituting the current time in the equation. network with three obstacles and 496 sensor nodes. The circular
Therefore, to incorporate future behaviours, we consider the time at shape network has three obstacles in the middle of the network. To
which the target changes its direction. Let us consider the t–x model a real communication link between nodes and robot, we
equation. First, differentiate the equation with respect to time and used a propagation model proposed in n [10, 40] and the equation
equate it to zero to find the set of time values {td1, td2, …} that is shown as follows:
changes the moving direction. Then, the average time of target
Prxi = Ptx j − 10εlogdi j − Lobi + vi (19)
moving in same direction Δt is calculated and the total set of
estimated time instances that target changes it direction is
TD = {td1, td2, …, tk + Δt, tk + 2Δt, …}. where the received signal strength at node i is Prxi, the transmitted
After that obtain the final estimated t–x equation that signal strength of the signal at node j is Ptx j, the path-loss exponent
incorporates the all behaviours as is ε, the distance between node i and node j is di j, Lobi is the loss
due to signal absorption from obstacles exist in the line of sight of
64 IET Wirel. Sens. Syst., 2018, Vol. 8 Iss. 2, pp. 60-67
© The Institution of Engineering and Technology 2017
Fig. 3  Results of DeTarSK for case 1
(a) Physical map, (b) Calculated ML-TM map, (c) Search using DeTarSK, (d) Target prediction error
Fig. 4  Result of DeTarSK for case 2
(a) Search using DeTarSK, (b) Target prediction error
Fig. 5  Result of DeTarSK for case 3
(a) Physical map, (b) Calculated ML-TM map, (c) Search using DeTarSK, (d) Target prediction error
nodes i and j, and the logarithm of shadowing component on node i
at is vi. The shadowing values are selected from a normal
distribution with zero mean parameter and standard deviation
parameter. The absorption coefficient and the thickness of the
obstacle medium, which signal traverses are α and do, respectively.
Then Lobi can be calculated as
n
Lobi = ∑ 10αdolog(e)
k=1
where e is the exponent and n is the number of obstacles exist in
between node i and node j.
To evaluate the performance of proposed target search
algorithm, three cases with different target motion patterns are
considered. In all three cases, the target and robot are moving with
a speed of 1.5 and 2 m s−1, respectively.
Case 1: Target enters to the sparse network at (1, 20), randomly
moves around and leaves the network after 82 s at (1, 10). The
autonomous robot starts to search the target from (30, 30) after 5 s
of target detection
Case 2: Target enters to the sparse network at (1, 25), randomly
moves around by changing the direction more frequently and leave
the network after 84 s at (60, 25). The autonomous robot starts to
search the target from (30, 30) after 10 s of target detection.
Case 3: Target enters to the circular shape obstacle network at
(2, 15), randomly moves around and leave the network after 44 s at
(28, 22). The autonomous robot starts to search the target from (3,
20) after 5 s of target detection.
For each case, we calculated target prediction error that
described the difference between predicted value and actual
^ ^ ^
location of the target. Let us consider Lti = [X ti, Y ti] as the predicted
target location at time ti using the proposed algorithm. Then the
target prediction error can be calculated as
IET Wirel. Sens. Syst., 2018, Vol. 8 Iss. 2, pp. 60-67
© The Institution of Engineering and Technology 2017
^ 2 ^ 2
(21)
Epr = (X ti − Xti) + (Y ti − Y ti)
Figs. 3–5 show the results for the three cases. The prediction error
plots describe how our proposed method adjusts for the sudden
movement changes of the target. For an example, in Fig. 3 we can
see the prediction error reduces with the time. However, when the
target makes a sudden direction change (at 43 s), the prediction
error increases, but it fine tunes using the proposed prediction
(20) method and brings down the error. It can be seen in Figs. 4 and 5
also. In Fig. 4, the target follows a random path, which has many
directional variations. However, our method captures those
variations in a few seconds and predicts the target location with an
error <5 m.
5.1 Performance comparison
The comparison of DeTarSK with other existing algorithm is
presented in this section. For that, we chose the recently proposed
P–G algorithm [2]. It is based on pseudo-gradient approach that
requires a communication hop count and RSS in the node
neighbourhood, which we do not consider in our proposed
algorithm. Additionally, this algorithm is not controlled by a
central node, but it is needed to calculate pseudo-gradient in each
target movement. For the comparison, we used the same three
cases described above. Since, DeTarSK algorithm is proposed for
topology maps and P–G algorithms are proposed for physical map,
we need to convert the results for one domain to perform a fair
evaluation. Therefore, DeTarSK robot trajectory in ML-TM is
mapped to physical map using procrustes analysis [43]. Let us
consider X and Y are the topological and physical coordinates
matrices. Then the Procrustes transformation factors can be
calculated as
Y = bXT + c; (22)
65
Fig. 6  Performance comparison for Case 1
(a) Target search using DeTarSK, (b) Target search using P–G algorithm, (c) Distance between target and robot
Fig. 7  Performance comparison for Case 2
(a) Target search using DeTarSK, (b) Target search using P–G algorithm, (c) Distance between target and robot
Fig. 8  Performance comparison for Case 3
(a) Target search using DeTarSK, (b) Target search using P–G algorithm, (c) Distance between target and robot
Table 1 Time required to capture the target
Case Robot speed = 2 m s−1 Robot speed = 3 m s−1
DeTarSK, s P–G DeTarSK, s P–G, s
1 70 —a 47
2 80 —a 52
3 29 18
—a
aCould not capture the target with 2 m s−1 robot speed.
where b is the scaling factor, T is the rotation angle and c is the
shift value. In this paper, transformation factors are calculated
locally. In other words, robot's each step coordinates are
transformed to the physical map using its local neighbourhood
sensor information.
Figs. 6–8 show the results of three cases. In those three figures,
sub-figures (a) show the proposed algorithm's autonomous robot
search path in the physical map and sub-figures (b) show the robot
search path using the P–G algorithm in the physical map. The
distance between target and the autonomous robot in each time
value is shown in sub-figures (c). The autonomous robot's speed,
and capturing time of the target in each case is stated in Table 1. In
all three cases, the target is moving with a 1.5 m s−1 speed. Also,
the P–G algorithm was unable to capture the target with 2 m s−1
robot speed as the target leaves the network as described in the
three cases. The reason that P–G algorithm was unable to capture
the target with a lower robot speed is, P–G algorithm does not have
a prediction method to calculate the future behaviour of the target,
thus, it takes time to adjust the moving direction of robot according
to the target moving pattern. Therefore, we increase the robot speed
upto 3 m s−1 and find out the time required to capture the target. In
66
case I, described by Fig. 6, proposed DeTarSK algorithm detects
the target in less time and lower speed than the P–G algorithm. In
case II also our proposed algorithm captures the target much faster
than the P–G algorithm with a lower speed. The reason, as can be
79
seen in Fig. 7, is that DeTarSK autonomous robot path does not
76 follow all the steps in the target trajectory, but it predicts the
35 target's future positions from past knowledge and makes a decision
on its next movement. Hence, it is faster than the P–G algorithm.
Case III results in Fig. 8 and shows that our algorithm can follow a
target and capture it in a network filled with obstacles. Even in this
case DeTarSK was able to catch the target with lower speed and in
less time compared to P–G algorithm. Figs. 6c–8c show that when
target changes its direction suddenly the distance increases.
However, compared to P–G algorithm, our proposed algorithm
detected it and made required changes to reduce the distance within
a few seconds. These results demonstrate that the DeTarSK
algorithm can catch the target within less time and lower robot
speed. Moreover, the DeTarSK consumes less energy in the WSN
compared to the P–G algorithm. The reason is pseudo-gradient
need to be calculated in each step of the robot movement. This
required to obtain the hop count from target to each and every node
in the network, which consumes more energy for packet
transmission and reception. Also, it increases the traffic in the
network. However, in the proposed method, the robot
communicates with its local neighbourhood only which requires
shorter length or one hop communication links. Thus, the proposed
scheme gives real-time result with less energy consumption.
6 Conclusion
We presented DeTarSK, a decentralised target search and
prediction algorithm for sensor network-based environments where
IET Wirel. Sens. Syst., 2018, Vol. 8 Iss. 2, pp. 60-67
© The Institution of Engineering and Technology 2017
it is not feasible to measure physical distances accurately using
techniques such as RSSI. The algorithm uses the maximum
likelihood topology coordinates instead of physical coordinates.
Knowledge of the target location or direction is estimated from
time stamps of observations of the target by nodes and conveyed to
the autonomous robot that is searching for the target. DeTarSK is
based on decentralised RKF and a non-linear least-square method.
It also removes the errors due to distortion of topology coordinate
domain compared to physical domain.
Even though we have considered this method in topology maps,
the proposed algorithm can also be used in geographical (physical)
coordinate system. Other advantages include the real-time decision
making with less traffic and energy consumption in the network. To
solve the drawbacks in decentralised algorithms, the future target
locations are predicted using past behaviours of the target.
DeTarSK shows better performance compared to the recently
proposed P–G algorithm. Additionally, the proposed algorithm can
be used in unknown environments containing obstacles in the
search path.
7 Acknowledgment
This research was supported in part by Data61|CSIRO Australia.
8 References
[1] Grundel, D.A.: ‘Searching for a moving target: optimal path planning’. Proc.
IEEE Networking, Sensing and Control, 2005, pp. 867–872
[2] Deshpande, N., Grant, E., Henderson, T.C.: ‘Target localization and
autonomous navigation using wireless sensor networks: a pseudogradient
algorithm approach’, IEEE Syst. J., 2014, 8, (1), pp. 93–103
[3] Gayan, S., Weeraddana, D.M., Gunathillake, A.: ‘Sensor network based
adaptable system architecture for emergency situations’, Lect. Notes Inf.
Theory, 2014, 2, (1), pp. 85–91
[4] Can, Z., Demirbas, M.: ‘A survey on in-network querying and tracking
services for wireless sensor networks’, Ad Hoc Netw., 2013, 11, (1), pp. 596–
610
[5] Beyme, S., Leung, C.: ‘Rollout algorithms for wireless sensor network-
assisted target search’, IEEE Sens. J., 2015, 15, (7), pp. 3835–3845
[6] Lau, H., Huang, S., Dissanayake, G.: ‘Probabilistic search for a moving target
in an indoor environment’. IEEE/RSJ Int. Conf. Intelligent Robots and
Systems, 2006, pp. 3393–3398
[7] Kagan, E., Goren, G., Ben-Gal, I.: ‘Probabilistic double-distance algorithm of
search after static or moving target by autonomous mobile agent’. IEEE 26th
Convention of Electrical and Electronics Engineers in Israel, 2010
[8] Wang, W., Ma, H., Wang, Y., et al.: ‘Performance analysis based on least
squares and extended Kalman filter for localization of static target in wireless
sensor networks’, Ad Hoc Netw., 2015, 25, pp. 1–15
[9] Dhanapala, D.C., Jayasumana, A.P.: ‘Topology preserving maps: extracting
layout maps of wireless sensor networks from virtual coordinates’,
IEEE/ACM Trans. Netw., 2014, 22, (3), pp. 784–797
[10] Gunathillake, A., Savkin, A., Jayasumana, A.P.: ‘Maximum likelihood
topology maps for wireless sensor networks using an automated robot’. 41st
IEEE Conf. Local Computer Networks, 2016
[11] Pathirana, P.N., Savkin, A.V., Jha, S.: ‘Location estimation and trajectory
prediction for cellular networks with mobile base stations’, IEEE Trans. Veh.
Technol., 2004, 53, (6), pp. 1903–1913
[12] Pathirana, P.N., Bulusu, N., Savkin, A.V., et al.: ‘Node localization using
mobile robots in delay-tolerant sensor networks’, IEEE Trans. Mobile
Comput., 2005, 4, (3), pp. 285–296
[13] Madsen, K., Nielsen, H.B., Tingleff, O.: ‘Methods for non-linear least
squares problem’ (Technical University of Denmark, Lyngby, Denmark,
2004)
[14] Lo, N., Berger, J., Noel, M.: ‘Toward optimizing static target search path
planning’. IEEE Symp. Computational Intelligence for Security and Defence
Applications, 2012, pp. 1–7
[15] Yoon, S., Soysal, O., Demirbas, M., et al.: ‘Coordinated locomotion of mobile
sensor networks’. 5th Annual IEEE Communications Society Conf. Sensor,
Mesh and Ad Hoc Communications and Networks, 2008, pp. 126–134
[16] Lu, X., Demirbas, M.: ‘Writing on water, a lightweight soft-state tracking
framework for dense mobile ad hoc networks’. 5th IEEE Int. Conf. Mobile
Ad Hoc and Sensor Systems, 2008, pp. 359–364
IET Wirel. Sens. Syst., 2018, Vol. 8 Iss. 2, pp. 60-67
© The Institution of Engineering and Technology 2017
[17] Yoon, S., Qiao, C.: ‘A new search algorithm using autonomous and
cooperative multiple sensor nodes’. 26th IEEE Int. Conf. Computer
Communications, 2007, pp. 937–945
[18] Li, J., Jannotti, J., De-Couto, D.S.J., et al.: ‘A scalable location service for
geographic ad hoc routing’. 6th Annual Int. Conf. Mobile Computing and
Networking, 2000
[19] Chong, E.K.P., Brewington, B.E.: ‘Decentralized rate control for tracking and
surveillance networks’, Ad Hoc Netw., 2007, 5, (6), pp. 910–928
[20] Bhatti, S., Xu, J., Memon, M.: ‘Clustering and fault tolerance for target
tracking using WSNs’, IET Wirel. Sens. Syst., 2011, 1, (2), pp. 66–73
[21] Pino-Povedano, S., Gonzalez-Serrano, F.J.: ‘Comparison of optimization
algorithms in the sensor selection for predictive target tracking’, Ad Hoc
Netw., 2014, 20, pp. 182–192
[22] Njoya, A.N., Thron, C., Barry, J., et al.: ‘Efficient scalable sensor node
placement algorithm for fixed target coverage applications of WSNs’, IET
Wirel. Sens. Syst., 2017, 7, (2), pp. 44–54
[23] Li, W., Zhang, W.: ‘Sensor selection for improving accuracy of target
localisation in wireless visual sensor networks’, IET Wirel. Sens. Syst., 2012,
2, (4), pp. 293–301
[24] Misra, S., Singh, A., Chatterjee, S., et al.: ‘QoS-aware sensor allocation for
target tracking in sensor-cloud’, Ad Hoc Netw., 2015, 33, pp. 140–153
[25] Eryildirim, A., Guldogan, M.B.: ‘A Bernoulli filter for extended target
tracking using random matrices in a UWB sensor network’, IEEE Sens. J.,
2016, 16, (11), pp. 4362–4373
[26] Sun, X., Koenig, S., Yeoh, W.: ‘Real-time adaptive A*’. Int. Joint Conf.
Autonomous Agents and Multiagent Systems, 2008, pp. 281–288
[27] Ishida, T., Korf, R.E.: ‘Moving-target search: a real-time search for changing
goals’, IEEE Trans. Pattern Anal. Mach. Intell., 1995, 17, (6), pp. 609–619
[28] Kagan, E., , Ben-Gal, I.: ‘Moving target search algorithm with informational
distance measures’, Open Appl. Inf. J., 2013, 6, pp. 1–10
[29] Hazim, T., Karagiannidis, G.K., Tsiftsis, T.A.: ‘Probability of early detection
of ultra-wideband positioning sensor networks’, IET Wirel. Sens. Syst., 2011,
1, (3), pp. 123–128
[30] Dong, B., Wang, X.: ‘Adaptive mobile positioning in WCDMA networks’,
EURASIP J. Wirel. Commun. Netw., 2005, 2005, (3), pp. 3877–3881
[31] Mahfouz, S., Mourad-Chehade, F., Honeine, P., et al.: ‘Nonparametric and
semi-parametric RSSI/distance modeling for target tracking in wireless sensor
networks’, IEEE Sens. J., 2016, 16, (7), pp. 2115–2126
[32] Yang, X., Zhang, W.A., Yu, L., et al.: ‘Multi-rate distributed fusion estimation
for sensor network-based target tracking’, IEEE Sens. J., 2016, 16, (5), pp.
1233–1242
[33] Marian, T., Mokryn, O.O., Shavitt, Y.: ‘Sensing clouds: a distributed
cooperative target tracking with tiny binary noisy sensors’, Ad Hoc Netw.,
2013, 11, (8), pp. 2356–2366
[34] Liu, B.H., Chen, M.L., Tsai, M.J.: ‘Message-efficient location prediction for
mobile objects in wireless sensor networks using a maximum likelihood
technique’, IEEE Trans. Comput., 2010, 60, pp. 865–878
[35] Zaidi, Z.R., Mark, B.L.: ‘Real-time mobility tracking algorithms for cellular
networks based on Kalman filtering’, IEEE Trans. Mob. Comput., 2005, 4,
(2), pp. 195–208
[36] Xiao, Q., Xiao, B., Cao, J., et al.: ‘Multihop range-free localization in
anisotropic wireless sensor networks: a pattern-driven scheme’, IEEE Trans.
Mob. Comput., 2010, 9, (11), pp. 1592–1607
[37] Chengdong, W., Shifeng, C., Yunzhou, Z., et al.: ‘A RSSI-based probabilistic
distribution localization algorithm for wireless sensor network’. 6th IEEE
Joint Int. Information Technology and Artificial Intelligence Conf., 2011, vol.
1, pp. 333–337
[38] Turner, D., Savage, S., Snoeren, A.C.: ‘On the empirical performance of self-
calibrating Wifi location systems’. IEEE 36th Conf. Local Computer
Networks, 2011, pp. 76–84
[39] Pagano, S., Peirani, S., Valle, M.: ‘Indoor ranging and localisation algorithm
based on received signal strength indicator using statistic parameters for
wireless sensor networks’, IET Wirel. Sens. Syst., 2015, 5, (5), pp. 243–249
[40] Gunathillake, A., Savkin, A., Jayasumana, A.P.: ‘Decentralized time-based
target searching algorithm using sensor network topology maps’. 12th IEEE
Int. Workshop on Performance and Management of Wireless and Mobile
Networks, 2016
[41] Petersen, I.R., Savkin, A.V.: ‘Robust Kalman filtering for signals and systems
with large uncertainties’, 1999
[42] ElHalawany, B.M., Abdel-Kader, H.M., TagEldeen, A., et al.: ‘Modified a*
algorithm for safer mobile robot navigation’. Int. Conf. Modelling,
Identification Control, 2013, pp. 74–78
[43] Nhat, V.D.M., Vo, N., Challa, S., et al.: ‘Nonmetric MDS for sensor
localization’. 3rd Int. Symp. Wireless Pervasive Computing, 2008, pp. 396–
400
67