Communication cost minimization in wireless sensor and actor networks for road surveillance

618 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 60, NO. 2, FEBRUARY 2011

Communication Cost Minimization in Wireless
Sensor and Actor Networks for Road Surveillance
Wenzhong Li, Member, IEEE, Edward Chan, Member, IEEE, Mounir Hamdi, Senior Member, IEEE,
Sanglu Lu, Member, IEEE, and Daoxu Chen, Member, IEEE

Abstract—In recent years, wireless sensor and actor networks
(WSANs) have been extensively deployed to monitor physical
environment and facilitate decision making based on data col-
lected. Emerging applications such as road surveillance highlight
some interesting research issues in WSANs, including coordina-
tion problems in sensor–actor or actor–actor communications.
In this paper, the issue of choosing a set of working actors for
coordinating data transmission in a road sensor and actor network
with minimum communication cost is studied. A theoretical model
is introduced to analyze the communication cost of data trans-
mission in WSANs, and the sensor–actor coordination problem
is formulated as an optimization problem. It is demonstrated
that the problem can be divided into subproblems, and optimal
solutions can be obtained by using a dynamic programming al-
gorithm. A novel graph-based algorithm is also proposed with a
communication-cost graph used to depict the cost of data trans-
mission and a modified Dijkstra’s algorithm to find optimal solu-
tions in reduced time complexity. The efficiency of the proposed
algorithms is confirmed using extensive simulations.
Index Terms—Communication optimization, road surveillance, Fig. 1. Example of road surveillance.
sensor-actor coordination problem, sensor and actor network.

I. I NTRODUCTION battery life. In WSANs, sensors and actors work together in
data-centric applications, with sensors gathering information

W IRELESS sensor and actor networks (WSANs), which
are composed of a set of sensors and actors linked
by wireless medium to perform distributed sensing and acting
about the physical world and actors taking appropriate actions
on the environment [4].
Recently, researchers have proposed the use of WSANs for
tasks, have been widely used in battlefield surveillance, chem- road surveillance; these are sometimes called road sensor and
ical attack detection, home automation, and environmental actor networks (RSANs) [5], [6]. Fig. 1 shows an example of
monitoring [1]–[3]. Sensors are low-cost, low-power devices an RSAN. A large number of sensors and actors are randomly
with limited sensing, computation, and wireless communication deployed on public roads to collect real-time data on traffic and
capacities. Actors are assumed to be equipped with better road conditions. When sensors detect an event (i.e., a passing
processing capabilities, higher transmission power, and longer car), they transmit sensing data (i.e., position and speed of the
car) to nearby actors, which transmit data to the sink through
long-range communication. The sink processes and analyzes
Manuscript received September 1, 2009; revised February 3, 2010, May 10, the sensor data on board and then issues action commands to
2010, and September 20, 2010; accepted November 2, 2010. Date of publication
November 29, 2010; date of current version February 18, 2011. This work was
actors. Typical applications in RSANs include vehicle tracking,
supported in part by the National Natural Science Foundation of China under driver warning, and incident detection. For example, if a driver
Grant 60803111, Grant 90718031, and Grant 60721002, by the 973 Program of is going in the wrong direction, the sink will detect this and
China under Grant 2009CB320705, by the JiangSu Natural Science Foundation
under Grant BK2009100, and by a grant from the City University of Hong
activate the appropriate actor to send a warning message to the
Kong under Project 7002494. The review of this paper was coordinated by driver. RSANs can also provide more sophisticated information
Prof. S. Ci. services such as road-surface condition reporting and traffic
W. Li, S. Lu, and D. Chen are with the Department of Computer Sci-
ence and Technology, Nanjing University, Nanjing 210093, China (e-mail:
information queries. There are a number of studies on RSANs,
lwz@nju.edu.cn; sanglu@nju.edu.cn; cdx@nju.edu.cn). such as the use of sensor networks for vehicle tracking in mil-
E. Chan is with the Department of Computer Science, City University of itary applications [7], [8]. Karpinski et al. proposed a software
Hong Kong, Kowloon, Hong Kong (e-mail: csedchan@cityu.edu.hk).
M. Hamdi is with the Department of Computer Science and Engineering, architecture for sensor networks in smart roads [6]. Jeong et al.
Hong Kong University of Science and Technology, Kowloon, Hong Kong presented an autonomous passive localization scheme for road
(e-mail: hamdi@cse.ust.hk). sensor networks [5]. However, not much has been done to
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. investigate how to achieve effective coordination in sensors and
Digital Object Identifier 10.1109/TVT.2010.2094631 actors in RSANs.
0018-9545/$26.00 © 2011 IEEE

LI et al.: COMMUNICATION COST MINIMIZATION IN WSANs FOR ROAD SURVEILLANCE 619

While a large number of sensors and actors are typically
deployed in a road network, such a dense deployment is usually
not necessary for actor nodes since actors have higher capac-
ities, and a single actor can cover a larger area. To reduce
network overhead, some actors could be put to sleep to save
energy and prolong network lifetime. The process of choosing
a set of working sensors and actors and establishing suitable
data paths between them is known as sensor–actor coordina-
tion [3], [4], [9]. Previous research has studied efficient data
collecting, clustering, and sleeping schedule strategies in sensor
networks [3], [10]–[17]. However, these strategies cannot be
applied directly to RSANs because of the following: First, they
work with different coordination patterns. The sensor–actor
network is like a heterogeneous network combining powerful
and resource-limited devices performing both long-range and
short-range communications. Second, most existing cluster-
ing algorithms in sensor networks are topology dependent. In Fig. 2. RSAN.
RSANs, the deployment of sensors and actors are restricted in
a road network. This restriction in topology yields a different The rest of this paper is organized as follows: Section II
communication optimization problem, which will be discussed presents the system model of RSANs and formulates the com-
in this paper. To the best of our knowledge, the sensor–actor munication cost minimization problem. Section III outlines
coordination issue with the explicit objective of minimizing both a dynamic programming solution and a graph-based solu-
network communication cost in road networks has not been tion to the problem and proves their correctness and optimality.
studied in the literature. The performance of the proposed algorithms is evaluated in
In this paper, we address the following challenging research Section IV. Section V introduces related work, and finally,
problem in the design of RSANs: how to choose a set of work- Section VI concludes this paper.
ing actors and route sensing data between sensors and actors
to minimize the total communication cost for road surveillance.
II. S YSTEM M ODEL
Specifically, we make the following contributions.
1) We present a theoretical model to analyze the communi- Fig. 2 shows the road sensor–actor network that corresponds
cation cost in an RSAN. By introducing the assumption to Fig. 1. In this example, sensors detect events and send their
of virtual actors deployed in road intersections, we formu- data to nearby actors. Unlike other research assuming actors
late the problem of choosing the working actor set (WAS) to be resource-rich nodes with unlimited power supply, we
for coordinating data transmission as an optimization make much weaker assumptions about actors in our model: We
problem. only assume that actors are capable of sensing and performing
2) We demonstrate that the optimization problem can be long-range communications. That is, in our model, actors are
divided into subproblems; thus, a dynamic programming not necessary to be powerful nodes; they could be resource-
algorithm is proposed to obtain the optimal solution to limited nodes operating on batteries, or they could be just
achieve minimal communication cost. The time complex- normal sensors that are chosen to collect data and send them
ity of the dynamic programming algorithm is proven to be to the sink (like cluster heads). The assumption of coexistence
O(nm2 ), where n is the number of nodes, and m is the of powerful and resource-limited actors coincides with our real
number of actors in RSANs. life experience. For example, people tend to deploy powerful
3) We also introduce an efficient graph-based algorithm to nodes in the busy roads to monitor their traffic condition. On
solve the problem in lower time complexity. A commu- the other hand, the rural area may not be covered by powerful
nication cost graph (CC-Graph) is used to depict the actors, where resource-limited nodes will be used for long-
cost of data transmission in RSANs, and a modified range communication. In such a case, minimizing communi-
Dijkstra’s algorithm is applied to find the shortest path, cation cost in the system is important to save energy and extend
with an optimal solution to the problem obtained as the network lifetime.
end result. The graph-based algorithm is proven to run Each actor has two states: working or inactive. If an actor
in O(nm + m2 lgm), which is more efficient than the is in the working state, it can sense events, collect data from
dynamic programming solution. nearby sensors, and establish long-range communication with
4) We further prove that, without the deployment of actors the sink. If it is in the inactive state, it acts like a normal sensor.
in virtual nodes, the proposed strategies produce a near- Consider the example in Fig. 2; if sensor A has detected an
optimal solution, which approaches the optimal solution event, the data collected by the sensor need to travel three hops
with asymptotic order O(1/m). to send to the nearest actor B. However, if actor C is in the
5) We produce extensive simulations to study the system working state, it only needs one hop for A to communicate
performance, which shows the efficiency of the proposed with C. More working actors clearly improves communication
approaches. efficiency as the sensor data can be more quickly processed but


Fig. 3. Data communication in a road segment.

Fig. 4. Data communication between two working actors.
at the expense of increased communication cost and greater
power consumption in the actors. Hence, the selection of a wi ∈ A (i = 1, . . . , K). The set WK = {w1 , w2 , . . . , wK } is a
suitable set of working actors to minimize the total network subset of A. We call WK a WAS of A.
communication cost is a key design decision in RSANs.
In most cases, RSANs are complicated and irregular net-
works, which are hard to analyze. To simplify the system B. Network Communication Cost
model, we introduce virtual actors. We assume that there are It is essential to minimize network communication in a
virtual actors deployed at intersections of the road networks, sensor network since sensors and some actors are resource-
and they are always in the working state. As shown in Fig. 2, constrained nodes with limited battery lifetime and communi-
the actors with dotted circle are virtual actors. The reason cation capabilities, and network communication is a significant
behind virtual nodes is that traffic is busier at the intersections component of energy consumption in sensors. Hence, we will
and requires careful surveillance; hence, people tend to deploy use energy cost as the key metric in our analysis.
actors in these places. In a real road, if no actors are deployed In RSANs, there are three main components in network com-
in such areas, the virtual node will redirect the workload to a munication cost: sensor–sensor, sensor–actor, and actor–sink
nearby working actor. For example, in Fig. 2, node D is a virtual communication. To simplify the analysis, we assume that
actor; if there is no real actor deployed in its area, the virtual the energy cost for unit data transmission in each hop of
node will direct its workload to a nearby working actor E or B. sensor–sensor and sensor–actor communication is the same,
The virtual nodes divide the road networks into a number which is denoted by ε. (Note that our model can be extended
of segments. Since virtual actors are assumed to always stay to the condition where ε is different.) Consequently, the cost
in the working state, we only need to optimize sensor–actor for an actor collecting data from a sensor is proportional to the
coordination in each road segment. In the rest of this paper, sensor’s data generation rate and the number of hops between
we formulate the communication cost minimization problem in them. For example, in Fig. 3, the energy cost between actor 1
a road segment and propose a dynamic programming solution and sensors 2 and 3 can be calculated by λ2 ε + 2λ3 ε. However,
and a graph-based algorithm to obtain the optimal solutions. energy cost for actor–sink communication is different for each
Since a virtual node may not exist in the actual network, the actor since the distance of each actor to the sink is not the same.
solutions we obtained may be not optimal. However, we will We use εai to denote the energy cost for actor ai sending a
show in Theorem 5 that our approach produces a near-optimal unit of data to the sink. We also assume that there is a constant
solution, with an asymptotic order of O(1/m), where m is the energy cost mai for keeping an actor in the working state,
number of actors in a road segment. meaning that there is background energy consumption, even in
Consider data communication in a road segment between the absence of data transmission. If an actor is a virtual node, we
two virtual actors, as shown in Fig. 3. Assume that there are n set its εai and mai to the minimum value of all working actors.
nodes in the road numbered by 1, 2, . . . , n, of which there are m In our analysis, we use energy consumption to describe
actors, which are denoted by a1 , a2 , . . . , am . Notice that a1 = 1 communication cost. We should point out that our model could
and am = n are virtual actors. Sensor i (1 ≤ i ≤ n) generates be used in more general situations by applying different cost
sensing data and query data with mean arrival rate λi . Actors metrics. For example, if we want to extend the lifetime of
collect data from nearby sensors and send them to the sink. the network, we need to reduce the energy consumption of
The sink stores the data, performs data-processing operations, resource-constrained nodes as much as possible. To achieve
handles user queries, sends back the requested data, broadcasts that, we could assign ε and εai relative cost values: the trans-
traffic information, and triggers the appropriate actor to perform mission energy cost to the energy it left. If a node is running
some actions such as accident alarm and driver warning. out of energy, its relative cost is high; hence, it is less likely to
be chosen to perform long-range communications. On the other
hand, if an actor is powerful, its εai will be assigned a small
A. WAS
value, and as a result, more sensors will transmit data through it.
Our objective is to find a set of working actors to minimize In the rest of this paper, we only use an absolute value of energy
the total network communication cost. We first provide the consumption as the cost metric to formulate the problem.
following formal definition of a WAS. Given a WAS WK , we now formulate its communication
Definition 1 (WAS): Assume that n sensor and actor nodes cost. We first consider the scenario shown in Fig. 4. In this case,
are deployed in a road, which are numbered by S = {1, 2, . . . , x and y are two working actors, and there is no working actor
n}. Among them are m (m ≤ n) actors denoted by A = {a1 , between them; thus, sensing data in the path are sent to either
a2 , . . . , am } (1 = a1 < a2 < · · · < am = n). Suppose that x or y. Consider a node i (x < i < y); its communication cost
there are K working actors, which are denoted by w1 , w2 , . . . , to x and y is |i − x|λi ε + λi εx and |i − y|λi ε + λi εy accord-
wK , where 0 ≤ K ≤ m, 1 = w1 < w2 < · · · < wK = n, and ingly. To minimize the cost, if i satisfies |i − x|λi ε + λi εx ≤


|i − y|λi ε + λi εy , i.e., i ≤ (1/2)(x + y − (εx − εy /ε)), data III. S OLUTIONS
from i should be sent to x; otherwise, it should be sent to
In this section, we present two solutions to the (a, b) −
y. We denote t(x, y) = (1/2)(x + y − (εx − εy /ε)) as the
W orkingSet problem: a dynamic programming solution and
split point of x and y. As shown in Fig. 4, data from node
a graph-based solution.
x + 1, . . . , t(x, y) are sent to actor x, and data from node
t(x, y) + 1, . . . , y − 1 are sent to actor y. If εx = εy , t(x, y) is
exactly the midpoint of x and y: t(x, y) = (x + y)/2 . We use A. Dynamic Programming Solution
the function φ(x, y) to indicate the total communication cost of
The following theorem shows that a solution to the (a, b) −
the nodes between x and y, which can be calculated as
W orkingSet problem can be obtained by solving sub-
t(x,y) y−1 problems.
φ(x, y) = λi · (i − x)ε + λj · (y − j)ε Theorem 1: Assume that S and A constitute a node set and
i=x+1 j=t(x,y)+1 an actor set, respectively, as defined in Definition 2. Suppose
t(x,y) y−1 that WK = {w1 , w2 , . . . , wK } is an optimal solution to the
+ λ i · εx + λ j · εy . (1) (a, b) − W orkingSet problem, and Qμ = {q1 , q2 , . . . , qμ } is
i=x+1 j=t(x,y)+1 an optimal solution to the (a, wK−1 ) − W orkingSet problem;
then, W = Qμ ∪ {wK } is also an optimal solution to the
In (1), the first two terms represent the cost for the collection (a, b) − W orkingSet problem.
of data from sensor nodes by x and y, and the last two terms Proof: As Qμ is an optimal solution to the (a, wK−1 ) −
represent the cost for sending the data from the actors to the W orkingSet problem, we have
sink. In addition, nodes x and y also act as sensors, which will
incur actor–sink communication cost λx εx and λy εy . Hence, C(a, wK−1 , Qμ ) ≤ C (a, wK−1 , WK − {wK }) . (5)
the total communication cost in this scenario is
The communication cost of W is
φ(x, y) + λx εx + λy εy + mx + my . (2) μ−1
C(a, b, W) = φ(qi , qi+1 ) + φ(qμ , wK )
Without loss of generality, assume that WK = {w1 , w2 ,
i=1
. . . , wK } (1 = w1 < w2 < · · · < wK = n) is a WAS. The set ⎛ ⎞
of nodes 1, 2, . . . , n can be divided into a number of sub-
sets: {w1 , . . . , w2 }, {w2 , . . . , w3 }, . . . , {wK−1 , . . . , wK }. The +⎝ (λj εj + mj ) + λwK εwK + mwK ⎠
communication cost of each subset can be calculated as in j∈Qµ

the preceding scenario. Thus, the total communication cost of ⎛ ⎞
μ−1
WK is =⎝ φ(qi , qi+1 ) + (λj εj + mj )⎠
K−1 i=1 j∈Qµ
φ(wi , wi+1 ) + (λj εj + mj ). (3)
i=1 j∈WK
+ φ(qμ , wK ) + λwK εwK + mwK
= C(a, qμ , Qμ ) + φ(qμ , wK ) + λwK εwK + mwK .

C. (a, b)-WorkingSet Problem On the other hand, the communication cost of WK is
K−1
We now study the problem of choosing the optimal work-
C(a, b, WK ) = φ(wi , wi+1 ) + (λj εj + mj )
ing actors to minimize the total network communication cost,
i=1 j∈WK
which we call (a, b)-WorkingSet problem. We provide a formal ⎛ ⎞
definition of the problem as follows: K−2
Definition 2 [(a, b)-WorkingSet Problem]: Given a set of =⎝ φ(wi , wi+1 ) + (λj εj + mj )⎠
sensor and actor nodes S = {a, a + 1, . . . , b} (a < b), a set of i=1 j∈WK −{wK }
actor nodes A = {a1 , a2 , . . . , am } satisfies a = a1 < a2 < · · ·
+ φ(wK−1 , wK ) + λwK εwK + mwK
< am = b (m ≤ b − a). Suppose that WK = {w1 , w2 , . . . ,
wK } is a WAS of A. Define an objective function as = C (a, wK−1 , WK − {wK }) + φ(wK−1 , wK )
K−1 + λwK εwK + mwK .
C(a, b, WK ) = φ(wi , wi+1 ) + (λj εj + mj ). (4)
i=1 j∈WK Notice that qμ = wK−1 and that according to (5)

The (a, b)-WorkingSet problem is given as follows: Find K C(a, b, W) − C(a, b, WK )
and WK so that C(a, b, WK ) is minimized.
= C(a, wK−1 , Qμ ) − C (a, wK−1 , WK − {wK }) ≤ 0.
According to this definition, the sensor–actor coordination
problem in road network is simply a (1, n) − W orkingSet That is
problem. Two solutions to the problem are presented in the next
section. C(a, b, W) ≤ C(a, b, WK ). (6)


On the other hand, since WK is an optimal solution to the
(a, b) − W orkingSet problem

C(a, b, WK ) ≤ C(a, b, W). (7)

Combining (6) and (7), we have

C(a, b, W) = C(a, b, WK ).

Hence, W is also an optimal solution to the problem.
Theorem 1 shows that the problem can be divided into sub-
problems; thus, dynamic programming can be applied to obtain
an optimal solution. By trying WK−1 on different actors, the
problem can be divided into different subproblems, which form
different WASs. The WAS with minimal cost is the optimal
solution to the problem. Let OP T (a, b) be the minimum cost
obtained in the (a, b) − W orkingSet problem. A dynamic
programming solution is given as follows:

OP T (a, b) = min {τ1 , OP T (a, i) + τi }
(i = a2 , a3 , . . . , am−1 ) (8)

where τ1 = φ(a, b) + λa εa + ma + λb εb + mb (which is the
cost when no working actors are chosen between nodes a and Fig. 5. Example of CC-Graph.
b), and τi = φ(i, b) + λb εb + mb .
According to the dynamic programming equation, the mini-
access cost. We call this graph a CC-Graph. Fig. 5 gives the
mal cost of the (1, n) − W orkingSet problem can be obtained
corresponding CC-Graph of the road in Fig. 3. The construction
by calculating OP T (1, n) recursively. At the same time, the
of the CC-Graph is described next.
optimal WAS WK can also be obtained when the process is
finished. Nodes: As shown in Fig. 5, starting from an entrance node
Theorem 2: The time complexity of the dynamic program- V (0, 0), the CC-Graph is composed of m rows and m
ming algorithm in solving a (1, n) − W orkingSet problem is columns, where m is the number of actors. The ith (1 ≤
O(nm2 ), where n is the number of nodes and m is the number j ≤ m) row of the graph has m + 1 − i nodes, and the
of actors. jth (1 ≤ j ≤ m) column has j nodes, forming an inverted
Proof: According to (1), the time complexity of φ(a, b) triangle. The nodes in row i (1 ≤ i ≤ m) are denoted as
or φ(i, b) is O(n). By applying dynamic programming to V(i,i) , V(i,i+1) , . . . , V(i,m) from left to right. We call the last
calculate (8), when we want to get the result of OP T (a, b), node in each column a critical node (marked in black). It
the result of OP T (a, i) (i = a2 , a3 , . . . , am−1 ) already exists can be seen from Fig. 5 that there are m critical nodes in
in the system memory. Hence, the value of OP T (a, b) can the CC-Graph: V(1,1) , V(2,2) , . . . , V(m,m) . Node V(m,m) is
be obtained by checking m − 1 values, each of which takes also known as the exit node.
O(n) time in calculating a function φ. This process runs Edges: There are horizontal and vertical edges in the CC-
in time complexity of O(mn). To calculate OP T (1, n), the Graph. For the ith (1 ≤ i ≤ m) row, there are horizontal
result of OP T (1, a2 ), OP T (1, a3 ), . . . , OP T (1, am−1 ) should directed edges from V(i,j) to V(i,j+1) (i ≤ j ≤ m − 1),
be sequentially calculated. Thus, the overall time complexity of and for the jth (1 ≤ j ≤ m) column, there are vertical
this problem is O(n) + O(2n) + · · · + O((m − 1)n), which is edges from V(i,j) to V(i+1,j) (1 ≤ i < j). There is also
O(nm2 ). a vertical edge from the entrance node V(0,0) to node
V(1,1) . We call the vertical edges from noncritical nodes
to critical nodes solid edges and the other vertical edges
B. Graph-Based Solution
virtual edges. We use dashed lines to indicate virtual edges
In this section, we will introduce a graph-based solution, and solid lines to indicate solid edges in the CC-Graph
which is more efficient than the dynamic programming algo- (see Fig. 5).
rithm. First, we construct a CC-Graph to represent the com- Weights: We assign a weight for each edge in the CC-Graph.
munication cost of the (1, n) − W orkingSet problem. Then, Let W (V(i,j) , V(i,j+1) ) denote the weight from node V(i,j)
we apply a modified Dijkstra’s algorithm to find a shortest path to node V(i,j+1) . The weights of horizontal edges are
in the graph. The shortest path is mapped to a WAS, which is calculated by
output as an optimal solution to the problem. Finally, we prove
the correctness of the graph-based algorithm. φ(ai , aj ), j = i+1
W V(i,j) , V(i,j+1) =
1) Construction of CC-Graph: For a (1, n) − W orkingSet φ(ai , aj+1 )−φ(ai , aj ), i+1 < j < m.
problem, we can use a weighted directed graph to describe its (9)


To reduce computational complexity, we calculate By applying the graph traveling rule, a CC-Path can be
W (V(i,j) , V(i,j+1) ) as follows: denoted by a sequence of critical nodes that it passes through.
For example, the aforementioned CC-Path passes through four
φ(ai , aj ), j =i+1
W V(i,j) , V(i,j+1) = critical nodes, which can be denoted by (V(1,1) , V(3,3) , V(5,5) ,
W1 + W2 + W3 , i + 1 < j < m
V(6,6) ). Obviously, each CC-Path must contain the entrance
(10)
node and exit node. On the other hand, each critical node se-
where
⎧ quence containing V(1,1) and V(m,m) denotes a unique CC-Path,
⎪ t(ai ,aj+1 )
⎪ i.e., the sequence (V(1,1) , V(m−1,m−1) , V(m,m) ) denotes the CC-
⎪
⎪ k=t(a ,a )+1[(2k−ai −aj )λk ε
⎪
⎪ Path: V(0,0) → V(1,1) → · · · → V(1,m−1) · · · → V(m−1,m−1) →
⎪
⎪ i j
V(m−1,m) → V(m,m) .
⎨
+(εai −εaj )λk , t(ai , aj+1 ) < aj
W1 = 2) Graph-Based Algorithm: The graph-based algorithm is
⎪
⎪
aj −1
⎪
⎪ [(2k−ai −aj )λk ε very simple. The basic idea is to construct a CC-Graph and find
⎪
⎪ k=t(ai ,aj )+1
⎪
⎪ the shortest CC-Path in the graph. The intuition is that the sum
⎩
+(εai −εaj )λk , aj ≤ t(ai , aj+1 ) of weights in a CC-Graph represents the communication cost of
⎧ a WAS. Thus, the shortest path represents the lowest cost WAS.
⎪ (aj+1 −aj )ε+(εaj+1 −εaj )
⎪
⎪
⎪ It is easy to modify the Dijkstra’s algorithm to follow our graph
⎪
⎪
aj −1
⎨ × λk , t(ai , aj+1 ) < aj traveling rule in finding a shortest CC-Graph. The detailed steps
W2 = k=t(ai ,aj+1 )+1 are described as follows.
⎪ t(a ,a )
⎪ i j+1
⎪
⎪
⎪
⎪ [(k−ai )λi ε+λi εai ] , aj ≤ t(ai , aj+1 ) Step 1) Construct a CC-Graph, and assign weights
⎩ to its edges using the method introduced in
k=aj
⎧ aj+1−1 Section III-B1.
⎪
⎪
⎪ k=a (aj+1 −k)λk ε+λk εaj+1 ,
⎪ t(ai , aj+1 ) < aj Step 2) Apply a modified Dijkstra’s algorithm on the CC-
⎪
⎨ j Graph with its weight assignment to get a shortest
W3 = aj+1 −1
path.
⎪
⎪ [(a −k)λk ε
⎪ k=t(a ,a )+1 j+1
⎪ Step 3) Assuming that the shortest cost path finding in
⎪
⎩ i j+1

+λk εaj+1 , aj ≤ t(ai , aj+1 ). step 2 is represented by a sequence of critical
nodes (V(x1 ,x1 ) , V(x2 ,x2 ) , . . . , V(xk ,xk ) ) (1 = x1 <
The equality of (9) and (10) can be proven by using the x2 < · · · < xk = m), map the sequence to a set of
definition of φ (1). The detailed derivation is shown in the working actor nodes W = {x1 , x2 , . . . , xk }. Output
Appendix. W as the optimal WAS.
For vertical edges, the weight of a virtual edge is
The correctness of the graph-based algorithm will be proven
assigned 0, and the weight of a solid edge is assigned the
in the next section.
communication cost of the corresponding actor
3) Correctness of the Graph-Based Algorithm: We will now
λaj εaj + maj , i = j − 1 prove the correctness of the graph-based algorithm. Several
W V(i,j) , V(i+1,j) =
0, otherwise lemmas are needed to get the desired result represented by
Theorem 3.
(1 ≤ i < j, 2 ≤ j ≤ m). (11) Lemma 1: Each CC-Path in the graph corresponds to a WAS
In addition, the weight of edge V (0, 0) → V (1, 1) is as- of actor nodes; each WAS of the actor node corresponds to a
signed by CC-Path in the graph.
Proof: We will show that a CC-path can be mapped to
W V(0,0) , V(1,1) = λ1 ε1 + m1 . a WAS of actor nodes. It has been shown in Section III-B1
that a CC-Path can be denoted by a sequence of critical nodes.
CC-Path and graph traveling rule: We define a communi- Assume that (V(x1 ,x1 ) , V(x2 ,x2 ) , . . . , V(xk ,xk ) ) (1 = x1 < x2 <
cation cost path (CC-Path) as a path from the entrance · · · < xk = m) is such a sequence. We now map it to a set
node V(0,0) to the exit node V(m,m) in the CC-Graph. For of working actors: W = {ax1 , ax2 , . . . , axk }. W is a subset
example, in Fig. 5, a CC-Path is the path from V(0,0) to of actors {a1 , a2 , . . . , am }. According to Definition 1, W is a
V(1,1) , traversing to V(1,3) , then down to V(3,3) , then to WAS of actor nodes.
V(3,5) , then down to V(5,5) , then to V(5,6) , finally reaching Now, we show that a WAS can be mapped to a CC-Path.
V(6,6) . Assume that W = {ax1 , ax2 , . . . , axk } (1 = x1 < x2 < · · · <
Traveling on a CC-Graph is like walking through a CC-Path xk = m) is a WAS of actor nodes. We can map it to a critical
from the entrance node to the exit node. We make the following node sequence (V(x1 ,x1 ) , V(x2 ,x2 ) , . . . , V(xk ,xk ) ). According to
graph traversal rule: each time a virtual edge is traversed, the the definition of CC-Path, this sequence corresponds to a CC-
traversal vertically continues until a solid edge is reached. In Path in the graph. The lemma is proven.
other words, both horizontal and vertical traversals are possible. Lemma 2: The sum of the weights in a CC-Path is equal to
Direction can be changed any time in horizontal traversals, the communication cost of the corresponding WAS.
whereas in vertical traversals, a critical node must be reached Proof: According to Lemma 1, a CC-Path can be de-
before direction can be changed. noted by a critical node sequence (V(x1 ,x1 ) , V(x2 ,x2 ) , . . . ,


V(xk ,xk ) ) (1 = x1 < x2 < · · · < xk = m), and its correspond- Proof: According to Lemma 1, there exists a one-to-one
ing WAS is W = {x1 , x2 , . . . , xk }. The CC-Path contains k − 1 mapping between the CC-Paths and the possible WAS of m
horizontal subpaths and k vertical subpaths. actor nodes. Lemma 2 shows that the sum of weights of a CC-
The horizontal subpaths and their weights are given as Path is equal to the communication cost of the corresponding
follows: WAS. By applying the modified Dijkstra’s algorithm, a shortest
path can be found. The sum of weights of the shortest path
1 V(x ,x ) → · · · → V(x ,x ) corresponds to the WAS with minimal communication cost,
1 1 1 2
which is the optimal solution to the (1, n) − W orkingSet
whose sum of weights is
problem.
x2 −1
Theorem 4: The time complexity of the graph-based algo-
W V(x1 ,j) , V(x1 ,j+1)
rithm in solving a (1, n) − W orkingSet problem is O(mn +
j=x1
x2 −1
m2 lgm).
= φ(x1 , x1 + 1) + (φ(x1 , j + 1) − φ(x1 , j)) Proof: The graph-based algorithm contains three steps.
j=x1 +1
We demonstrate the running time of each step as follows:
For step 1, a CC-Graph has O(m2 ) nodes and O(m2 )
= φ(x1 , x2 );
edges. According to (10), the running time of calculating
2 V(x ,x ) → · · · → V(x ,x ) W (V(i,j) , V(i,j+1) ) is O(t(ai , aj+1 ) − t(ai , aj )) + O(aj+1 −
2 2 2 3 aj ). Thus, the running time for calculating the weights
whose sum of weights is of edges in each row is O( m j=i+1 W (V(i,j) , V(i,j+1) )) =
x3 −1
O(t(ai , am )) + O(am ) = O(n). Thus, the time complexity of
W V(x2 ,j) , V(x2 ,j+1) CC-Graph construction in step 1 is O(mn). According to [18],
j=x2
the running time of Dijkstra’s algorithm in a graph with V
x3 −1
nodes and E edges is O(V lgV + E) by using a Fibonacci
= φ(x2 , x2 + 1) + (φ(x2 , j + 1) − φ(x2 , j))
heap implementation. In our case, the running time of applying
j=x2 +1
the modified Dijkstra’s algorithm in an CC-Graph (step 2) is
= φ(x2 , x3 )
O(m2 lg(m2 ) + m2 ) = O(m2 lgm). The running time of step 3
.
. is clearly O(m). Thus, the total running time of the graph-based
.
algorithm is O(mn + m2 lgm).
k−1 V Compared with Theorem 2, it can be seen that the graph-
(xk−1 ,xk−1 ) → · · · → V(xk−1 ,xk )
based algorithm is more efficient than the dynamic program-
whose sum of weights is
xk −1
ming algorithm.
W V(xk−1 ,j) , V(xk−1 ,j+1) = φ(xk−1 , xk ).
j=xk−1 C. Discussions

The k vertical subpaths and their weights are given as As we have mentioned, the dynamic programming algorithm
follows: and the graph-based algorithm produce an optimal solution
based on the assumption of virtual working actors. In real
RSANs, working actors may be not deployed in each intersec-
1 V(0,0) → V(x ,x ) tion; thus, the solution may be not optimal in the real case. The
1 1

whose sum of weights is λx1 εx1 + mx1 ; following theorem shows that the proposed algorithms produce
a near-optimal solution when there are no working actors in the
2 V(x ,x ) → · · · → V(x ,x ) intersections.
1 2 2 2

whose sum of weights is λx2 εx2 + mx2 ; Theorem 5: Given n sensors and actors deployed in a road
. segment, with m actors among them, assume that C is the
.
. minimum communication cost obtained by the dynamic pro-
gramming algorithm or graph-based algorithm, and C is the
k V(x ,x ) → · · · → V(x ,x )
k−1 k k k actual communication cost. Let ΔC = C − C, and consider
whose sum of weights is λxk εxk + mxk . the expectation of ΔC/C. We have
Thus, the sum of the weights of the CC-Path is ΔC 1
E =O . (13)
k−1 C m
φ(xi , xi+1 ) + (λj εj + mj ). (12)
Proof: Assume that S = {1, 2, . . . , n} is the set of sensors
i=1 j∈W
and actors. We now consider the first node to be a virtual node.
Compared with (3), the sum of the weights in a CC-Path Due to symmetry, the situation is similar when the last node is a
is equal to the communication cost of the W . The lemma is virtual node. If there is no actual working actor deployed in the
proven. first node, the solution obtained by the dynamic programming
Theorem 3: The graph-based algorithm gives an optimal algorithm and graph-based algorithm is not optimal since the
solution to the (1, n) − W orkingSet problem. traffic to the first node needs to be redirected to a nearby


Fig. 6. Communication cost when the first node is not a working actor.

working actor. Assume that y (1 < y < n) is the first working
actor from the first node. As shown in Fig. 6, the traffic in the
path from x to t(x, y) (x = 1 in this case) is directed to y, but
the traffic of the rest of the nodes remains unchanged.
According to (1), if the first node is a working actor, the
communication cost is
t(1,y) y−1
φ(1, y) = λi · (i − 1)ε + λj · (y − j)ε
i=1 j=t(1,y)+1 Fig. 7. Simulation area.
t(1,y) y−1
+ λ i · ε1 + λ j · εy . According to (3), the optimal cost C satisfies
i=1 j=t(1,y)+1 m
n n
C≥ φ(wi , wi+1 ) ≥ mλmin ε 1 +
If the first node is not a working actor, as shown in Fig. 6, the i=1
2m m
communication cost becomes
n
= λmin ε 1 + n.
t(1,y) y−1 y 2m
φ (1, y) = λi ·(y−i)ε+ λj ·(y−j)ε+ λi ·εy .
i=1 j=t(1,y)+1 i=1
Thus
2
Thus ΔC E(y 2 ) 1 + 3 n−m + 2 n−m
m m
E =O =O n
t(1,y) t(1,y)
C C λmin ε 1 + 2m n
ΔC = φ (1, y)−φ(1, y)= λi (y+1−2i)+ λi (εy −ε1 ). 1
i=1 i=1 =O .
m
Let λmax = max{λi } and λmin = min{λi }, (i = 1, 2, . . . , n).
We have The theorem is proven.
Theorem 5 shows that, without deploying virtual nodes,
ΔC ≤ y 2 λmax + yλmax (εy − ε1 ) = O(y 2 ). the proposed algorithms will give a near-optimal solution,
which approaches the optimal solution with asymptotic order
Assume that m actors are randomly deployed in the n nodes, O(1/m). If the number of actors m in a road segment is large
i.e., each node has a probability of m/n being an actor. Thus, enough, our solution will be very close to the optimal solution.
the probability of the first working actor being the ith node is
m i−1 m
P r{y = i} = 1 − · . IV. P ERFORMANCE E VALUATION
n n
In this section, we conduct simulation experiments to evalu-
Thus, the expectation of y 2 is ate the performance of the proposed algorithms. Our simulation
n
i−1 2 scenario is based on the application of road condition surveil-
m m n−m n−m
E(y 2 ) = i2 1− ≈ 1+3 +2 . lance in a city block of Beijing, China. Experiment results show
i=1
n n m m the efficiency of the proposed strategies.
On the other hand, by (1)
t(x,y) y−1 A. Simulation Environment
φ(x, y) ≥ λi · (i − x)ε + λj · (y − j)ε We simulate the application of road surveillance for the
i=x+1 j=t(x,y)+1 Beijing Olympic Stadium. Fig. 7 shows a rectangular area
⎛ ⎞ (2980 m × 2030 m), where the Beijing Olympic Stadium
t(x,y) y−1
≥ λmin ε ⎝ (i − x) + (y − j)⎠ is located. Sensors and actors are randomly deployed in the
i=x+1 j=t(x,y)+1
three main roads from east to west and the two main roads
from north to south (yellow lines in the figure) around the
y−x Olympic Stadium to monitor road conditions. When sensors
≥ λmin ε 1 + (y − x).
2 detect events, the sensing data are sent to the nearby actors and


Fig. 8. Performance under various node density (λ = 0.01). (a) Energy cost. (b) Number of working actors.

then onward to the sink, which is located inside the Olympic function of node density ρ. When ρ varies from 200 to 400,
Sports Center, as shown in the figure. the energy cost of Random, Top K, and ALL increases, but
We list a number of tunable parameters used in the experi- the cost of OPT remains constant. As can be seen, OPT has the
ments as follows: We use ρ to indicate node density, which is lowest energy cost and is invariant to node density. As expected,
calculated by the average number of nodes per kilometer in the ALL has the highest energy cost, which is about dozens of
WSAN. We vary ρ from 200 to 400 to check the scalability times higher than that of the OPT strategy. Fig. 8(b) shows
of the algorithm. ξ represents the actor density (the average the average number of working actors selected by the four
number of actors per kilometer), varying from 30 to 150 in the different algorithms on each road. The OPT strategy shows a
experiments. near constant number of working actors under all conditions.
We consider three traffic conditions: 1) busy traffic (repre- There are many more working actors in the other algorithms,
senting rush hours), where each sensor generates hundreds of which linearly increase with the node density.
sensing data per second; 2) normal, where each sensor gener- The reason OPT consumes much less energy than the other
ates one or two units of data per second; 3) idle (representing algorithms can be explained as follows: For the condition
late evening or early morning), where only a few sensing data λ = 0.01, there is only a few data items to be transferred
are generated in a long interval. We use λ to denote the average in the network. Actors are idle most of the time. The OPT
data generation rate of a sensor. λ = 100, 1, 0.01 corresponds to strategy turns off as many actors as possible to save energy.
the preceding three conditions, respectively. Each sensing data As shown in Fig. 8(b), OPT selects about four working actors
unit is assumed to be a single byte. The transmission range of and is insensitive to node density. Thus, the energy cost remains
sensors is set to 10 m. constant. The Random, Top K, and ALL strategies choose more
We employ an energy model similar to [3] and [10]. The working actors while the node density increases. As a result,
energy consumption per bit is calculated as E = 2Eelec + βdα , more energy is consumed, even if there is no data transmission.
where α is the exponent of the path loss (2 ≤ α ≤ 5), β is a Fig. 9 shows the performance when the node density is fixed
constant [J/(bit · mα )], and Eelec is the energy needed by the to 300 and the actor density varies from 30 to 150. It is shown
transceiver circuitry to transmit or receive one bit [in joules per that the energy cost of all algorithms, with the exception of
bit]. The communication energy parameters are set as Eelec = OPT, increases with the node density. Again, OPT demonstrates
50 nJ/bit, β = 10 pJ/bit/m2 , and α = 2. a constant number of working actors and the lowest energy cost.
As both the dynamic programming algorithm and graph- Figs. 10 and 11 compare performance for λ = 1. In this case,
based algorithm give near-optimal solutions to the problem, we the energy cost and working actor number of all algorithms
only implement the graph-based algorithm in our simulation for increase when the node density increases. Still, OPT has the
its complexity is lower. The graph-based algorithm is referred lowest energy cost. The performances of Random and Top K
to the OPT algorithm in the rest of this paper. We compare our are similar: they both consume more energy than OPT. ALL
algorithm with several other heuristic solutions: 1) Random: has the highest energy cost, which is about 10%–20% higher
Select working actors at random. 2) Top K: Select the K largest than OPT. Fig. 11 also shows a favorable property of the OPT
aggregate data transmission rate actors as the working set. The algorithm: the number of working actors is insensitive to the
top 30% actors are chosen in our simulation. 3) ALL: All actors actor density ξ. For the OPT algorithm, the number of working
are assumed to be in the working state. actors remains constant when ξ varies from 30 to 150. Its
corresponding energy cost also remains the same most of the
time. Other algorithms use more working actors and consume
B. Simulation Results
more energy than OPT algorithm. In this figure, the energy
Fig. 8 shows performance under the traffic condition λ = cost of ALL varies from 675 to 850 mJ when the actor density
0.01. Fig. 8(a) compares system energy consumption as a increased. However, the cost of OPT remains below 650 mJ,


Fig. 9. Performance under various actor density (ρ = 300, λ = 0.01). (a) Energy cost. (b) Number of working actors.

Fig. 10. Performance under various node density (λ = 1). (a) Energy cost. (b) Number of working actors.

Fig. 11. Performance under various actor density (ρ = 300, λ = 1). (a) Energy cost. (b) Number of working actors.

which represents an energy savings of about 30%, compared transmission dominates energy consumption. Fig. 13(a) shows
with ALL. the percentage of energy consumed in long-range actor–sink
Fig. 12 compares performance for the different schemes communication. As can be seen, more than 90% of the energy
when λ = 100. As illustrated in the figure, even though the is consumed in actor–sink communication. When there is a
number of working actors selected by these algorithms is quite sufficient number of working actors, adding more working
different, the energy costs of the four strategies are roughly the actors only affects data communication between sensors, which
same. The reason is given as follows: Unlike the other two cannot reduce the cost of actor–sink communication. Thus,
conditions, data traffic is quite busy in this case; thus, data energy cost is insensitive to actor-selecting algorithms when


Fig. 12. Performance under various node density (λ = 100). (a) Energy cost. (b) Number of working actors.

Fig. 13. Energy and workload under various node density (λ = 100). (a) Proportion of actor–sink energy cost. (b) Average workload.

traffic is busy. However, the number of actors does affect data collection, coordination, and responses to events. However,
the average workload of each node. Fig. 13(b) compares the the heterogeneity of WSANs also introduces many research
average data transmission rate of the algorithms. It shows that challenges including node deployment, real-time requirement,
OPT has the lowest average workload, as all actors are working. coordination, mobility, etc. [2]–[4], [19], [20].
The number of working actors in OPT is very close to ALL [see The use of WSANs for vehicular tracking has been proposed
Fig. 12(b)]; thus, it also has a reasonable workload. TOP K has in a number of recent projects [5]–[8]. Karpinski et al. proposed
the heaviest workload, which is more than twice that of the OPT a software architecture for sensor networks in smart roads
algorithm. [6]. Jeong et al. presented an autonomous passive localization
In conclusion, according to the simulation, the advantages of scheme for road sensor networks [5]. However, little attention
our algorithm can be summarized as follows: When the system has been paid to effective coordination in sensors and actors in
is in idle condition, the OPT algorithm turns off as many actors RSANs.
as possible to save energy. For normal road surveillance, it Some recent papers have considered the issue of real-time
yields the lowest energy cost, and its performance is insensitive communication and coordination in sensor networks [2], [3],
to actor density. In a busy traffic environment, it keeps the [21]. The SPEED protocol [21] provides real-time commu-
workload of working actors at a low level. nication services and is designed to be a stateless localized
algorithm with low control overhead. Melodia et al. proposed
a sensor–actor coordination model based on an event-driven
V. R ELATED W ORK
partitioning paradigm to achieve global network objectives
Unlike traditional wireless sensor networks, which are com- such as real-time constraints and minimum energy consumption
posed of homogeneous nodes and single-sink performing cen- [3]. Wu et al. studied sensor–actor connectivity strategies to
tralized operations [1], WSANs enable heterogeneous nodes put as many sensors as possible to sleep for energy-saving
to cooperatively perform distributed sensing and active tasks purposes [2].
[2]–[4]. Equipped with more energy and computing resources, Energy-efficient routing and data-collecting protocols have
the actor nodes can perform more complex functions such as been extensively studied, and a wide range of special network


structures, such as shortest path tree, minimum spanning tree, If t(ai , aj+1 ) < aj , i.e., node aj is in the right part of the
minimum Steiner tree, clustering, grid, and chain, are utilized midpoint t(ai , aj+1 ), according to (1), we have
to achieve the goal of reducing energy consumption [3], [10]–
t(ai ,aj ) aj −1
[17], [22]–[26]. Sleep management has been shown to be an
effective method for saving energy [27]–[29]. The protocols φ(ai , aj ) = λk ·(k−ai )ε+ λk ·(aj −k)ε
proposed in [30] and [31] can improve network performance k=ai +1 k=t(ai ,aj )+1

by maintaining a backbone composed of a small number of t(ai ,aj ) aj −1
active nodes while scheduling other nodes to operate in low + λk ·εai + λk ·εaj
duty cycles to conserve energy. However, as we mentioned k=ai +1 k=t(ai ,aj )+1
before, these strategies cannot be directly used in RSANs. To
t(ai ,aj )
our knowledge, the sensor–actor coordination issues with the
explicit objective of minimizing network communication cost = [λk ·(k−ai )ε+λk ·εai ]
in road networks has not been extensively studied in the past. k=ai +1

t(ai ,aj+1 )

VI. C ONCLUSION + λk ·(aj −k)ε+λk ·εaj
k=t(ai ,aj )+1
In this paper, we have addressed the question of choosing
aj −1
a set of working actors for coordinating data transmission
in WSANs to minimize communication cost. We have first + λk ·(aj −k)ε+λk ·εaj
k=t(ai ,aj+1 )+1
presented a theoretical model to analyze the communication
cost of data transmission in RSANs. Based on the assump- t(ai ,aj+1 )
tion of virtual nodes, we have formulated the sensor–actor φ(ai , aj+1 ) = λk ·(k−ai )ε
coordinating communication problem as an optimization prob- k=ai +1
lem and proven that it can be divided into subproblems, thus aj+1 −1
allowing optimal solutions to be obtained by using dynamic
+ λk ·(aj+1 −k)ε
programming. The time complexity of the dynamic program-
k=t(ai ,aj+1 )+1
ming algorithm has been proven to be O(nm2 ). We have also
proposed a lower complexity graph-based algorithm, which t(ai ,aj+1 ) aj+1 −1
uses a CC-Graph to depict the cost of data communication and + λk ·εai + λk ·εaj+1
a modified Dijkstra’s algorithm to find an optimal solution to k=ai +1 k=t(ai ,aj+1 )+1
the problem. The running time of the graph-based algorithm
t(ai ,aj )
is O(mn + m2 lgm), which is more efficient than the dynamic
programming algorithm and more applicable in large-scale = [λk ·(k−ai )ε+λk ·εai ]
k=ai +1
distributed systems. We have further proven that, without the
deployment of actors in virtual nodes, the proposed strategies t(ai ,aj+1 )
produce a near-optimal solution, which approaches the optimal + [λk ·(k−ai )ε+λk ·εai ]
solution with asymptotic order O(1/m). Performance of the k=t(ai ,aj )+1
proposed algorithms has been evaluated by simulation, which aj −1
confirms the efficiency of the proposed algorithms.
+ λk ·(aj+1 −k)ε+λk ·εaj+1
The strategies discussed in this paper are local optimization
k=t(ai ,aj+1 )+1
based on road segments divided by virtual nodes. In the future,
we will address the following issues to improve it. We intend aj+1 −1

to consider how to minimize communication cost in the entire + λk ·(aj+1 −k)ε+λk ·εaj+1 .
network, instead of individual segments, which is a substan- k=aj
tially more complicated optimization problem. We will also
investigate the feasibility of exploiting the mobility pattern of Thus
the vehicles to carry the sensing data to the nearby actors as a
means to decrease the cost of sensor–actor communications. φ(ai , aj+1 ) − φ(ai , aj )
t(ai ,aj+1 )

A PPENDIX = (2k − ai − aj )λk ε + (εai − εaj )λk
k=t(ai ,aj )+1
P ROOF OF THE E QUALITY OF (9) AND (10)
aj −1
The equality of (9) and (10) can be verified by
showing that φ(ai , aj+1 ) − φ(ai , aj ) = W1 + W2 + W3 when + (aj+1 − aj )ε + (εaj+1 − εaj ) λk
k=t(ai ,aj+1 )+1
i + 1 < j < m.
There are two possible relative positions of node aj and aj+1 −1
the split point of ai and aj+1 : t(ai , aj+1 ) < aj , and aj ≤ + (aj+1 − k)λk ε + λk εaj+1 .
t(ai , aj+1 ). k=aj


Compared with the expression of W1 , W2 , and W3 , when [22] S. Doshi, S. Bhandare, and T. X. Brown, “An on-demand minimum energy
t(ai , aj+1 ) < aj , it is easy to see that routing protocol for a wireless ad hoc network,” SIGMOBILE Mobile
Comput. Commun. Rev., vol. 6, no. 3, pp. 50–66, Jul. 2002.
[23] C. Intanagonwiwat, R. Govindan, D. Estrin, J. Heidemann, and F. Silva,
φ(ai , aj+1 ) − φ(ai , aj ) = W1 + W2 + W3 . “Directed diffusion for wireless sensor networking,” IEEE/ACM Trans.
Netw., vol. 11, no. 1, pp. 2–16, Feb. 2003.
[24] M. Èagalj, J.-P. Hubaux, and C. Enz, “Minimum-energy broadcast in all-
If aj ≤ t(ai , aj+1 ), i.e., node aj is in the left of the midpoint wireless networks: Np-completeness and distribution issues,” in Proc. 8th
t(ai , aj+1 ), we can obtain the same conclusion similarly. Thus, Annu. Int. Conf. Mobile Comput. Netw., MobiCom, 2002, pp. 172–182.
the equality of (9) and (10) is proven. [25] A. Manjeshwar and D. Agrawal, “APTEEN: A hybrid protocol for effi-
cient routing and comprehensive information retrieval in wireless sensor
networks,” in Proc. IPDPS, 2002, pp. 195–202.
R EFERENCES [26] J. E. Wieselthier, G. D. Nguyen, and A. Ephremides, “Energy-efficient
broadcast and multicast trees in wireless networks,” Mobile Netw. Appl.,
[1] I. F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci, “Wireless vol. 7, no. 6, pp. 481–492, Dec. 2002.
sensor networks: A survey,” Comput. Netw., vol. 38, no. 4, pp. 393–422, [27] W. Ye, J. Heidemann, and D. Estrin, “An energy-efficient MAC protocol
Mar. 2002. for wireless sensor networks,” in Proc. IEEE INFOCOM, Jun. 2002,
[2] J. Wu, S. Yang, and M. Cardei, “On maintaining sensor–actor connec- pp. 1567–1576.
tivity in wireless sensor and actor networks,” in Proc. IEEE INFOCOM, [28] T. van Dam and K. Langendoen, “An adaptive energy-efficient MAC
Apr. 2008, pp. 286–290. protocol for wireless sensor networks,” in Proc. 1st Int. Conf. Embedded
[3] T. Melodia, D. Pompili, V. Gungor, and I. Akyildiz, “Communication and Netw. Sens. Syst., SenSys, 2003, pp. 171–180.
coordination in wireless sensor and actor networks,” IEEE Trans. Mobile [29] G. Lu, B. Krishnamachari, and C. S. Raghavendra, “An adaptive energy-
Comput., vol. 6, no. 10, pp. 1116–1129, Oct. 2007. efficient and low-latency MAC for tree-based data gathering in sensor
[4] I. F. Akyildiz and I. H. Kasimoglu, “Wireless sensor and actor networks: networks: Research articles,” Wireless Commun. Mobile Comput., vol. 7,
Research challenges,” Ad Hoc Netw., vol. 2, no. 4, pp. 351–367, Oct. 2004. no. 7, pp. 863–875, Sep. 2007.
[5] J. Jeong, S. Guo, T. He, and D. Du, “APL: Autonomous passive local- [30] B. Chen, K. Jamieson, H. Balakrishnan, and R. Morris, “Span: An energy-
ization for wireless sensors deployed in road networks,” in Proc. IEEE efficient coordination algorithm for topology maintenance in ad hoc wire-
INFOCOM, Apr. 2008, pp. 583–591. less networks,” Wireless Netw., vol. 8, no. 5, pp. 481–494, Sep. 2002.
[6] M. Karpiriski, A. Senart, and V. Cahill, “Sensor networks for smart [31] G. Xing, C. Lu, Y. Zhang, Q. Huang, and R. Pless, “Minimum power
roads,” in Proc. 4th Annu. IEEE Int. Conf. Pervasive Comput. Commun. configuration in wireless sensor networks,” in Proc. 6th ACM Int. Symp.
Workshops, Mar. 2006, pp. 306–310. Mobile Ad Hoc Netw. Comput., MobiHoc, 2005, pp. 390–401.
[7] T. He, S. Krishnamurthy, J. A. Stankovic, T. Abdelzaher, L. Luo,
R. Stoleru, T. Yan, L. Gu, J. Hui, and B. Krogh, “Energy-efficient sur-
veillance system using wireless sensor networks,” in Proc. 2nd Int. Conf.
Mobile Syst., Appl., Serv., MobiSys, 2004, pp. 270–283.
[8] G. Simon, M. Maróti, Á. Lédeczi, G. Balogh, B. Kusy, A. Nádas,
G. Pap, J. Sallai, and K. Frampton, “Sensor network-based countersniper
system,” in Proc. 2nd Int. Conf. Embedded Netw. Sens. Syst., SenSys,
2004, pp. 1–12. Wenzhong Li (S’05–M’08) received the B.S. and
[9] J. Stankovic, T. Abdelzaher, C. Lu, L. Sha, and J. Hou, “Real-time com- Ph.D. degrees in computer science from Nanjing
munication and coordination in embedded sensor networks,” Proc. IEEE, University, Nanjing, China, in 2002 and 2007,
vol. 91, no. 7, pp. 1002–1022, Jul. 2003. respectively.
[10] W. Heinzelman, A. Chandrakasan, and H. Balakrishnan, “An application- He is currently an Assistant Professor with the
specific protocol architecture for wireless microsensor networks,” IEEE Department of Computer Science and Technology,
Trans. Wireless Commun., vol. 1, no. 4, pp. 660–670, Oct. 2002. Nanjing University. He has published more than
[11] J. Polastre, J. Hill, and D. Culler, “Versatile low power media access for 30 research papers at international conferences and
wireless sensor networks,” in Proc. 2nd Int. Conf. Embedded Netw. Sens. journals, which include the International Conference
Syst., SenSys, 2004, pp. 95–107. on Distributed Computing Systems, the International
[12] V. P. Mhatre, C. Rosenberg, D. Kofman, R. Mazumdar, and N. Shroff, “A Workshop on Quality of Service, the International
minimum cost heterogeneous sensor network with a lifetime constraint,” Conference on Parallel Processing, the International Conference on Commu-
IEEE Trans. Mobile Comput., vol. 4, no. 1, pp. 4–15, Jan./Feb. 2005. nications, the IEEE Wireless Communications and Networking Conference,
[13] O. Younis and S. Fahmy, “HEED: A hybrid, energy-efficient, distributed Computer Communications, and the Journal of Supercomputing. His research
clustering approach for ad hoc sensor networks,” IEEE Trans. Mobile interests include wireless networks, sensor networks, pervasive computing, and
Comput., vol. 3, no. 4, pp. 366–379, Oct.–Dec. 2004. peer-to-peer computing.
[14] F. Kuhn, T. Moscibroda, and R. Wattenhofer, “Initializing newly deployed Dr. Li is a member of the Association for Computing Machinery. He has
ad hoc and sensor networks,” in Proc. 10th Annu. Int. Conf. Mobile published papers in the IEEE T RANSACTIONS ON V EHICULAR T ECHNOL -
Comput. Netw., MobiCom, 2004, pp. 260–274. OGY. He was the recipient of the Best Paper Award at the 2009 ICC.
[15] W. Yu, J. Cao, X. Zhou, X. Wang, K. C. C. Chan, A. T. S. Chan, and H. V.
Leong, “A high-throughput mac protocol for wireless ad hoc networks,”
IEEE Trans. Wireless Commun., vol. 7, no. 1, pp. 135–145, Jan. 2008.
[16] Q. Han, S. Mehrotra, and N. Venkatasubramanian, “Application-aware
integration of data collection and power management in wireless sensor
networks,” J. Parallel Distrib. Comput., vol. 67, no. 9, pp. 992–1006,
Sep. 2007.
[17] L. Paradis and Q. Han, “TIGRA: Timely sensor data collection using Edward Chan (M’88) received the B.Sc. and M.Sc.
distributed graph coloring,” in Proc. 6th Annu. IEEE Int. Conf. PERCOM, degrees in electrical engineering from Stanford
2008, pp. 264–268. University, Stanford, CA, and the Ph.D. degree
[18] T. H. Cormen, C. E. Leiserson, R. L. Rivest, and C. Stein, Introduction to in computer science from Sunderland University,
Algorithms, 2nd ed. Cambridge, MA: MIT Press, 2006. Sunderland, U.K.
[19] K. Akkaya and M. Younis, “C2ap: Coverage-aware and connectivity- He worked in industry for a number of years in
constrained actor positioning in wireless sensor and actor networks,” in the design and implementation of computer networks
Proc. IPCCC, Apr. 2007, pp. 281–288. and real-time control systems before joining City
[20] B. McLaughlan and K. Akkaya, “Coverage-based clustering of wireless University of Hong Kong, Kowloon, Hong Kong,
sensor and actor networks,” in Proc. IEEE ICPS, Jul. 2007, pp. 45–54. where he is currently an Associate Professor with
[21] T. He, J. A. Stankovic, C. Lu, and T. Abdelzaher, “SPEED: A stateless the Department of Computer Science. His current
protocol for real-time communication in sensor networks,” in Proc. 23rd research interests include performance evaluation of high-speed networks,
ICDCS, 2003, pp. 46–55. mobile data management, power-aware computing, and network management.


Mounir Hamdi (S’90–M’91–SM’06) received the Sanglu Lu (M’04) received the B.S., M.S., and
B.S. degree in electrical engineering (computer en- Ph.D. degrees in computer science from Nanjing
gineering option with distinction) from the Univer- University, Nanjing, China, in 1992, 1995, and 1997,
sity of Louisiana, Lafayette, in 1985 and the M.S. respectively.
and Ph.D. degrees in electrical engineering from the She is currently a Professor with the Depart-
University of Pittsburgh, Pittsburgh, PA, in 1987 and ment of Computer Science and Technology and the
1991, respectively. State Key Laboratory for Novel Software Technol-
From 1985 to 1991, he was a Teaching/Research ogy, Nanjing University. Her research interests in-
Fellow with the Department of Electrical Engineer- clude distributed computing, pervasive computing,
ing, University of Pittsburgh, where he was involved and wireless networks.
in major research projects, as well as teaching un-
dergraduate courses. Since 1991, he has been a faculty member with the
Department of Computer Science and Engineering, Hong Kong University of
Science and Technology, Kowloon, Hong Kong, where he is currently Head and
Professor of Computer Science and Engineering. He has been the Director of
the Computer Engineering Program that has about 350 undergraduate students Daoxu Chen (M’04) received the B.S. degree in
for nine years and the Director of the Master of Science in Information Tech- computer science from Nanjing University, Nanjing,
nology that has about 150 postgraduate students for three years. He is also the China, in 1982.
Director of the Computer Engineering Laboratory. From 1999 to 2000, he held He has been a Visiting Researcher with Purdue
Visiting Professor positions at Stanford University, Stanford, CA, and the Swiss University, West Lafayette, IN, and City University
Federal Institute of Technology, Lausanne, Switzerland. His general areas of of Hong Kong, Kowloon, Hong Kong. He is cur-
research are high-speed packet switches/routers, optical networks, and wireless rently a Professor with the Department of Com-
local area networks, in which he has published more than 260 research publica- puter Science and Technology, Nanjing University,
tions, received numerous research grants, and supervised some 30 postgraduate Nanjing, China. His research interests include dis-
students. His current research interests are high-speed networks, including the tributed computing, parallel processing, and com-
design, analysis, scheduling, and management of high-speed switches/routers, puter networks.
wavelength division multiplexing networks/switches, and wireless networks. Prof. Chen is a member of the Association for Computing Machinery and a
Dr. Hamdi is a Senior Member of the Association for Computing Machinery. Senior Member of the China Computer Federation.

Communication cost minimization in wireless sensor and actor networks for road surveillance

More Related Content

What's hot (19)

Viewers also liked (11)

Similar to Communication cost minimization in wireless sensor and actor networks for road surveillance (20)

More from ambitlick (20)

Recently uploaded (20)

Communication cost minimization in wireless sensor and actor networks for road surveillance