CCF Transactions on Networking

, Volume 1, Issue 1–4, pp 65–77 | Cite as

Data fusion-oriented cluster routing protocol for multimedia sensor networks based on the degree of image difference

  • Fan He
  • Haiping HuangEmail author
  • Ruchuan Wang
  • Lingyun Jiang
Regular Paper


In recent years, wireless multimedia sensor networks (WMSN) have attracted considerable attentions because of the extensive applications of providing audio-visual information conveniently. However, the traditional routing protocols are not directly applicable to WMSN due to the disadvantages on real-time or energy cost. The routing based on combination of clusters and data fusion provides an effective approach to improve the efficiency of data transmission. Consequently, this paper proposes a novel data fusion-oriented routing protocol, which takes the degree of image difference into consideration when clustering. And meanwhile, the cluster-head is rotated periodically within a cluster, for the purpose of the establishment of data fusion tree which is responsible for the data transmission and redundancy elimination. The simulation results show that the proposal achieves a better performance on energy efficiency, network delay and loading balance, and simultaneously prolongs the network lifetime, compared with other typical protocols.


Wireless multimedia sensor networks Clustering Degree of image difference Routing protocol Data fusion Network lifetime 

1 Introduction

Wireless multimedia sensor networks (WMSN) are composed of a large number of sensor nodes that can capture multimedia content, such as image, video and audio. Compared with traditional wireless sensor networks (WSN), WMSN can acquire various data types to complete more complex tasks such as wildlife tracking, elderly anti-fall monitoring and shared vision in field environment (Hasan et al. 2017; He et al. 2018a, b). However, the processing of multimedia information also brings difficulties and challenges, especially in the aspects of QoS-guarantee and energy consumption (Bhandary et al. 2016; Wang et al. 2016a). The demands on higher bandwidth, more desirable real-time performance and more powerful processing capacity are indispensable in WMSN owing to huge amounts of multimedia data. Unfortunately, resource-constrain is exactly one of the most typical characteristics of sensor networks due to limited computation, communication and storage capacity. If every node transmits data to the sink without data processing, the data redundancy will cause a great waste of bandwidth and energy.

To solve these issues, researchers find the combination of data fusion algorithm and clustering based routing protocol is an effective method to reduce the amount of data transmission, improve network reliability and prolong the network lifetime (Alanazi and Elleithy 2015; Wang et al. 2016b).

However, existing clustering methods of WSN can not be directly applied to WMSN due to the uniqueness of WMSN. Some current studies on clustering consider the field of view of multimedia sensors but are with heavy computation overheads or require accurate location information. Moreover, most WMSN clustering methods haven’t comprehensively consider energy consumption and QoS of the network.

Therefore, the challenges of real-world applications based on WMSN and limitations of existing studies have motivated to propose a clustering algorithm for WMSN which considers both the unique properties of WMSN and low computation overheads. Furthermore, the selection of cluster heads and the structure of the data fusion tree have great influence on the network lifetime and QoS of WMSN, which are also worth being further studied.

We also consider that the function of data aggregation for WMSN is not only to simply discard or merge duplicate data, but also to pick the most important media information from specific sensors to further reduce data redundancy. Furthermore, clustering enables the above functions to be executed parallelly, and meanwhile the proposed routing algorithm based on clustering can select optimal paths for those specific nodes to achieve QoS-guarantee multimedia data transmission. So the data fusion-oriented routing protocol becomes the focus of this paper.

The contributions of this paper can be described as follows.
  • The definition of image difference in WMSN is introduced.

  • A clustering algorithm for WMSN based on image difference is proposed.

  • The selection of cluster heads and the construction of cluster head tree are studied to achieve the satisfactory QoS performance and balanced network lifetime.

The rest of this paper is organized as follows. Section 2 reviews the related work briefly and summarizes our contributions. In Sect. 3, it introduces the definition of the degree of image difference for clustering. Section 4 describes our proposed data fusion-oriented routing protocol in detail. Simulation evaluation and performance analysis will be presented in Sect. 5. Finally, in Sect. 6, we conclude this paper and outline our future work.

2 Related works

LEACH (Heinzelman et al. 2000) protocol is the earliest sensor network routing algorithm based on data fusion, which employs clustering to aggregate data, however LEACH cannot achieve satisfactory efficiency and availability. The directed diffusion routing protocol (ab. DD) (Intanagonwiwat et al. 2000) uses caching mechanism to achieve data transmission in path establishment phase and packet routing phase, whose disadvantage is not all the optimal path can be obtained for each selection. PEGASIS (Lindsey and Raghavendra 2002) is a chain-based routing algorithm which collects data on the constructed intermediate nodes, however this causes the fast energy consumption of these intermediate nodes. Besides these three classic routing protocols, Necchi et al. (2007) also point out that clustering network structure is suitable for data fusion; and Mehrabi et al. (2015) adopt a data fusion scheme which employs mobile sink to discover the transmission path simultaneously taking the energy cost into account. However, most research on data fusion algorithm and routing protocol is based on the traditional wireless sensor networks, such as AFST (Luo et al. 2006), GRMax (Attoungble and Okada 2012), EAR (De et al. 2012), DRINA (Villas et al. 2013), HCCR (Nayak et al. 2012), NCOF (Alaei and Barcelo-Ordinas 2010), EEGRA (Chen et al. 2016), NEECRP (Lee et al. 2014) and DGFP (Sha et al. 2016) and so on. In WMSN, each multimedia sensor node has a certain field of view (Fov), Soro et al. (2005) indicate that the classic algorithms are not directly applicable to WMSN not only because of their potential limitations such as infeasibility or expensive energy cost, but also because the Fov of multimedia sensor is different from that of the traditional sensor. The Fov of traditional sensor node is a circle that generally takes itself as its center, sensing distance as its radius, while the multimedia sensor’s Fov is a sector. Therefore, Obraczka et al. (2002) suggest that Fov is an important factor to be considered when studying clustering in WMSN.

Ma and Liu (2005) propose a simplified model of Fov for wireless multimedia sensor nodes (seen in Fig. 1). Each sensor \( S_{i} \) is associated with a quad \( (O_{i} ,R,\mathop V\limits^{ \to } ,\alpha ) \), where \( O_{i} \) represents location of \( S_{i} \), \( R \) represents the size of sensing radius, \( V \) represents the unit vector of sensing orientation \( V \) (\( V \) is the angular bisector of Fov), \( \alpha \) represents the angle between \( R \) and \( \mathop V\limits^{ \to } \).
Fig. 1

Simplified model of Fov for wireless multimedia sensor nodes

Alaei and Barcelo-Ordinas (2010) put forward a clustering algorithm based on overlapping Fovs for wireless multimedia sensor networks. It calculates overlapping Fovs by geometric method and takes the result as an important basis of clustering. Nevertheless, the calculation process is complicated and not suitable for practical applications of WMSN. Similar with Alaei and Barcelo-Ordinas (2010), Li and Chuang (2014) propose a multi-path scheme for the achievement of effective transmission for multimedia data; however, the use of accurate geographic information brings extra expense. Aiming at different service demands, Demir et al. (2013) suggest the image transmission routing algorithm based on source coding, which is only suitable for image sensor networks. Furthermore, a scheme of tasks schedule and data transmission for video sensor networks is proposed in Huang et al. (2009), where the convergence between simulated annealing algorithm and ant colony algorithm obtains the optimal solution to tasks schedule and meanwhile reduces the end to end delay. Spachos et al. (2015) propose an angle-based QoS and energy-aware dynamic routing scheme designed for wireless multimedia sensor networks which extends network lifetime by optimizing the selection of the forwarding candidate set. In Magaia et al. (2015), the routing problems in WMSN are studied as multi-objective optimization problems, and it puts forward an improved genetic algorithm which shows significant improvement on QoS metrics. Ahmed (2017) evolves a novel algorithm for accomplishing adaptive traffic shaping of multimedia streaming and utilizes the multipath forwarding with dynamic cost calculation for selecting next hop in WMSN.

In this paper, based on the model in Ma and Liu (2005), we propose a Data Fusion-oriented Routing Protocol for WMSN (ab. DFRP), whose contributions can be described as follows. First it can select those nodes with little degree of image difference into the same cluster with low computation complexity. Second, it can ensure the energy consumption balance while the selection of cluster-head occurs in the network, which indicates an energy-efficient data fusion process can be carried out in cluster-heads. Finally, a data fusion tree will be established, which can further save energy and prolong the lifetime of WMSN.

3 The degree of image difference

In this section, we assume that multimedia sensors are intentionally deployed in monitor area, providing detailed visual information from disparate viewpoints. A multimedia sensor network with n sensors is represented by \( S = \{ S_{1} ,S_{2} , \ldots ,S_{n} \} . \)

The image capturing of multimedia sensor is a process of mapping three-dimensional objects onto two-dimensional capturing planar. In order to get the degree of image difference, it needs to transform coordinates of objects from three-dimension into two-dimension. Assume A, B, C, D, E, F are points of dimensional coordinate system, their coordinates are \( O(0,0,0)^{T} ,A(1,0,0)^{T} ,B( - 1,0,0)^{T} ,C(0,1,0)^{T} ,D(0, - 1,0)^{T} ,E(0,0,1)^{T} ,F(0,0, - 1)^{T} \) respectively, which may construct six unit vectors \( \overrightarrow {OA} ,\overrightarrow {OB} ,\overrightarrow {OC} ,\overrightarrow {OD} ,\overrightarrow {OE} ,\overrightarrow {OF} \) as shown in Fig. 2.
Fig. 2

Points of dimensional coordinate system

For example, in Fig. 3, the coordinates of multimedia sensor \( S_{1} \) is \( ( - d,0)^{T} \), whose sensing orientation points to X-axis positive orientation, the coordinates of \( S_{2} \) is \( ( - d\cos \theta , - d\sin \theta )^{T} \), whose sensing orientation has an included angle \( \theta \) with X-axis positive orientation, and the coordinates of \( S_{3} \) is \( ( - d\cos \theta + r\sin \theta , - d\sin \theta - r\cos \theta )^{T} \). The sensing orientation of \( S_{1} \) and \( S_{2} \) both pass through the origin \( O \), and the sensing orientation of \( S_{3} \) is parallel with \( S_{2} \)’s. In spite of the different locations and sensing orientations, all of the three nodes have a same depth of focus \( d \) and focal length \( f \).
Fig. 3

Deployed situations of nodes

To get the coordinates on capturing planar of sensors, we define the coordinate’s transformation formula as Eq. (1):
$$ P_{i} = R \times P + O_{i} $$

In which, \( P_{i} \) is the coordinates of point \( P \) on the capturing planar of sensor \( S_{i} \). \( R \) is the rotation matrix, which represents the rotation degree relative to standard coordinate system. \( O_{i} \) is the coordinates of origin with respect to \( S_{i} \).

Take \( S_{1} \) as an example, where \( R = \left[ {\begin{array}{*{20}c} 1 & 0 & 0 \\ 0 & 1 & 0 \\ 0 & 0 & 1 \\ \end{array} } \right] \), \( O_{i} = (d,0,0)^{T} \). So, the coordinates of points in Fig. 2 on the capturing planar of sensor \( S_{1} \) are \( O_{1} = (d,0,0)^{T} ,A_{1} = (d + 1,0,0)^{T} ,B_{1} = (d - 1,0,0)^{T} ,C_{1} = (d,1,0)^{T} ,D_{1} = (d, - 1,0)^{T} \), \( E_{1} = (d,0,1)^{T} ,F_{1} = (d,0, - 1)^{T} \).

Table 1 lists the coordinates of points in Fig. 3, after mapping them onto the capturing planar of \( S_{1} ,S_{2} \) and \( S_{3} \) separately.
Table 1

Coordinates of points after mapping


Coordinates on imaging planar

S 1

S 2

S 3


\( (0,0)^{T} \)

\( (0,0)^{T} \)

\( \left( {\frac{r}{d}f,0} \right)^{T} \)


\( (0,0)^{T} \)

\( \left( {\frac{ - \sin \theta }{d + \cos \theta }f,\,0} \right)^{T} \)

\( \left( {\frac{r - \sin \theta }{d + \cos \theta }f,\,0} \right)^{T} \)


\( (0,0)^{T} \)

\( \left( {\frac{\sin \theta }{d - \cos \theta }f,\,0} \right)^{T} \)

\( \left( {\frac{r + \sin \theta }{d - \cos \theta }f,\,0} \right)^{T} \)


\( \left( {\frac{f}{d},0} \right)^{T} \)

\( \left( {\frac{\cos \theta }{d + \sin \theta }f,\,0} \right)^{T} \)

\( \left( {\frac{r + \cos \theta }{d + \sin \theta }f,\,0} \right)^{T} \)


\( \left( { - \frac{f}{d},0} \right)^{T} \)

\( \left( {\frac{ - \cos \theta }{d - \sin \theta }f,\,0} \right)^{T} \)

\( \left( {\frac{r - \cos \theta }{d - \sin \theta }f,\,0} \right)^{T} \)


\( \left( {0,\frac{f}{d}} \right)^{T} \)

\( \left( {0,\frac{f}{d}} \right)^{T} \)

\( \left( {\frac{r}{d}f,\frac{f}{d}} \right)^{T} \)


\( \left( {0, - \frac{f}{d}} \right)^{T} \)

\( \left( {0, - \frac{f}{d}} \right)^{T} \)

\( \left( {\frac{r}{d}f, - \frac{f}{d}} \right)^{T} \)

Based on the coordinates as above, the unit vectors are shown as Table 2.
Table 2

Unit vector after mapping onto imaging planar

Unit vector

Unit vector on imaging planar

S 1

S 2

S 3

\( \overrightarrow {OA} \)

\( (0,0)^{T} \)

\( \left( {\frac{ - \sin \theta }{d + \cos \theta }f,0} \right)^{T} \)

\( \left( {\frac{r - \sin \theta }{d + \cos \theta }f - r\frac{f}{d},0} \right)^{T} \)

\( \overrightarrow {OB} \)

\( (0,0)^{T} \)

\( \left( {\frac{\sin \theta }{d - \cos \theta }f,0} \right)^{T} \)

\( \left( {\frac{r + \sin \theta }{d - \cos \theta }f - r\frac{f}{d},0} \right)^{T} \)

\( \overrightarrow {OC} \)

\( \left( {\frac{f}{d},0} \right)^{T} \)

\( \left( {\frac{\cos \theta }{d + \sin \theta }f,0} \right)^{T} \)

\( \left( {\frac{r + \cos \theta }{d + \sin \theta }f - r\frac{f}{d},0} \right)^{T} \)

\( \overrightarrow {OD} \)

\( \left( { - \frac{f}{d},0} \right)^{T} \)

\( \left( {\frac{ - \cos \theta }{d - \sin \theta }f,0} \right)^{T} \)

\( \left( {\frac{r - \cos \theta }{d - \sin \theta }f - r\frac{f}{d},0} \right)^{T} \)

\( \overrightarrow {OE} \)

\( \left( {0,\frac{f}{d}} \right)^{T} \)

\( \left( {0,\frac{f}{d}} \right)^{T} \)

\( \left( {0,\frac{f}{d}} \right)^{T} \)

\( \overrightarrow {OF} \)

\( \left( {0, - \frac{f}{d}} \right)^{T} \)

\( \left( {0, - \frac{f}{d}} \right)^{T} \)

\( \left( {0, - \frac{f}{d}} \right)^{T} \)

As can be seen from Table 2, the lengths of vectors vary with the imaging planar. As O, E and F have the same depth of focus and focal length for different nodes, \( \frac{{\overrightarrow {OE} }}{{\overrightarrow {OF} }} \) is a constant. We use the projection factor \( s = \frac{f}{d} \) to represent the respective length of \( \overrightarrow {OE} \) and \( \overrightarrow {OF} \) after mapping.

Moreover, it can be learned that the length of \( \overrightarrow {OA} ,\overrightarrow {OB} ,\overrightarrow {OC} ,\overrightarrow {OD} \) is changing with the change of node’s location and sensing orientation. Based on the above conditions, we propose a method to describe the different lengths of unit vectors for different nodes. If \( S_{i} \) and \( S_{j} \) are two multimedia sensor nodes deployed on the monitor area, the degree of image difference between them can be figured out by the following steps:
Step 1

Obtain the location and Fov information of \( S_{i} \) and \( S_{j} \) via any lightweight localization method for wireless sensor networks;

Step 2

Calculate the distances for unit vectors \( \overrightarrow {OA} ,\overrightarrow {OB} ,\overrightarrow {OC} ,\overrightarrow {OD} \). For example, assume that the mapping of \( \overrightarrow {OA} \) on \( S_{i} \) is \( \overrightarrow {OA}_{i} = (u_{i} ,v_{i} )^{T} \) and \( \overrightarrow {OA}_{j} = (u_{j} ,v_{j} )^{T} \) is its mapping on \( S_{j} \), and the distance between the two mappings is \( d_{OA} = \sqrt {(u_{i} - u_{j} )^{2} + (v_{i} - v_{j} )^{2} } \);

Step 3

Take the average distances of the 4 vectors as the degree of image difference \( \delta_{i,j} \) for \( S_{i} \) and \( S_{j} \):

$$ \delta_{i,j} = \frac{1}{4}(d_{OA} + d_{OB} + d_{OC} + d_{OD} ). $$

If \( S_{i} \) and \( S_{j} \) are deployed as \( S_{1} \) and \( S_{2} \) in Fig. 3, according to Table 2, we can figure out \( \delta_{1,2} = \frac{1}{4}(\left| {\frac{d\sin \theta }{d + \cos \theta }} \right| + \left| {\frac{d\sin \theta }{d - \cos \theta }} \right| + \left| {\frac{d\cos \theta }{d + \sin \theta } - 1} \right| + \left| {\frac{ - d\cos \theta }{d - \sin \theta } + 1} \right|) \);

In generally, any two multimedia sensors \( S_{i} \) and \( S_{j} \) with respective location information \( (d_{i} ,r_{i} ,\theta_{i} ) \) and \( (d_{j} ,r_{j} ,\theta_{j} ) \) have a degree of image difference:
$$ \begin{aligned} \delta_{i,j} = \frac{1}{4}\left( \begin{aligned} \left| {\frac{{ - d_{i} \sin \theta_{i} - r_{i} \cos \theta_{i} }}{{d_{i} + \cos \theta_{i} }} - \frac{{ - d_{j} \sin \theta_{j} - r_{j} \cos \theta_{j} }}{{d_{j} + \cos \theta_{j} }}} \right| \hfill \\ \hfill \\ \end{aligned} \right. \hfill \\ + \left| {\frac{{d_{i} \sin \theta_{i} + r_{i} \cos \theta_{i} }}{{d_{i} - \cos \theta_{i} }} - \frac{{d_{j} \sin \theta_{j} + r_{j} \cos \theta_{j} }}{{d_{j} - \cos \theta_{j} }}} \right| \hfill \\ \;\;\;\;\;\;\;\;\;\; + \left| {\frac{{d_{i} \cos \theta_{i} - r_{i} \sin \theta_{i} }}{{d_{i} + \sin \theta_{i} }} - \frac{{d_{j} \cos \theta_{j} - r_{j} \sin \theta_{j} }}{{d_{j} + \sin \theta_{j} }}} \right| \hfill \\ \;\;\;\;\;\;\;\;\;\; +\left. \left| {\frac{{ - d_{i} \cos \theta_{i} + r_{i} \sin \theta_{i} }}{{d_{i} - \sin \theta_{i} }} - \frac{{ - d_{j} \cos \theta_{j} + r_{j} \sin \theta_{j} }}{{d_{j} - \sin \theta_{j} }}} \right| \right)\hfill \\ \end{aligned} $$

The larger \( \delta_{i,j} \) is, the images captured by the two nodes will have a smaller correlation, which means less redundant information. If two nodes have the same sensing orientation and their locations are close, the value of \( \delta_{i,j} \) can be 0; when their sensing orientations are mutually perpendicular, \( \delta_{i,j} \) comes to 1.

4 Data fusion-oriented routing protocol

A multimedia sensor network can be divided into two layers, where the capturing-layer is composed of member nodes, and the fusion-layer consists of cluster-heads, as shown in Fig. 4. Actually, the network also can be expanded multi-layer construction with multilevel cluster-heads.
Fig. 4

The network structure based on cluster

The capturing-layer consists of member nodes within clusters and will transmit data gathered to cluster heads in the fusion-layer. The cluster heads will then transmit data along the paths in the constructed fusion tree to the upper layer until to the sink node. The multi-layer structure illustrated in Fig. 4 can effectively reduce the data redundancy and energy consumption in data transmission.

In the rest of this section, we will explain the details of DFRP which has three major phases: (1) clustering; (2) cluster-head selection and (3) construction of cluster-head fusion tree, as shown in Fig. 5. Each phase has a specific task and uses a set of parameters or messages.
Fig. 5

Three major phases in DFRP

In the clustering phase, nodes exchange image difference information (which comes from location and Fov information) and perform clustering based on image difference.

In the cluster head election phase, a cluster head is generated according to the weighted average of its remaining energy and its average distance from other nodes in the same cluster. The election result is then broadcasted within the cluster.

In the final phase, the structure of the data fusion tree is decided according to the remaining energy of all cluster heads as well as the distances from the sink node. The link message of the data fusion tree will be broadcasted for further routing selection.

4.1 Clustering

According to Eq. (3), the degree of image difference \( \delta_{i,j} \) between \( S_{i} \) and \( S_{j} \) can be figured out, and it can finally be expressed with a n order matrix \( \delta_{n \times n} \):
$$ \delta_{n \times n} = \left[ {\begin{array}{*{20}c} 0 & {\delta_{1,2} } & {\delta_{1,3} } & \ldots & {\delta_{1,n} } \\ {\delta_{2,1} } & 0 & {\delta_{2,3} } & \ldots & {\delta_{2,n} } \\ {\delta_{3,1} } & {\delta_{3,2} } & 0 & \ldots & {\delta_{3,n} } \\ \ldots & \ldots & \ldots & \ldots & \ldots \\ {\delta_{n,1} } & {\delta_{n,2} } & {\delta_{n,3} } & \ldots & 0 \\ \end{array} } \right] $$

\( \delta_{n \times n} \) is a symmetric matrix whose main diagonal is composed of 0. If \( \delta_{i,j} \le \rho \) (\( \rho \) is a given threshold), it considers images captured by the two nodes have the high correlation, which means they can be assigned into the same cluster.

Initially there are n clusters and each cluster has only one member node \( S_{i} \). The maximum capacity of each cluster is nmax. Unlike traditional clustering algorithms which considers Euclidean distance between nodes, the clustering result is mainly influenced by the correlation of the images captured between different nodes. The algorithm tries to find all nodes \( S_{j} (j \ne i) \) that satisfies \( \delta_{i,j} \le \rho \), and merges \( cluster\_j \) with \( cluster\_i \) into one set \( cluster\_i^{'} \). The algorithm terminates when m clusters are constructed or no more nodes which satisfies \( \delta_{i,j} \le \rho \) can be found.

The pseudo-code of mentioned clustering process is listed as follows:

4.2 Cluster-head selection

Selecting the cluster-head is the next step after the clustering phase. Every node first broadcasts a HELLO_MSG with the form \( Sensor(ID_{i} ,E_{i} ,x_{i} ,y_{i} ) \) to indicate its unique ID, rest energy and located coordinate. Every node maintains and updates a member table which contains the information of the nodes in the same cluster.

\( d_{i,j} \) denotes the distance between \( S_{i} \) and \( S_{j} \), which can be calculated by \( d_{i,j} = \sqrt {(x_{i} - x_{j} )^{2} + (y_{i} - y_{j} )^{2} } \). According to the member table mentioned above, in one cluster with k nodes, any node \( S_{i} \) can calculate the average rest energy \( E_{avg} = \frac{{\sum {E_{i} } }}{k} \) and the average communication distance to other nodes \( \overline{{d_{i} }} = \frac{{\sum {d_{i,j} } }}{k - 1} \), where \( \sum {E_{i} } \) represents the total sum of rest energy of k nodes in this cluster, \( \sum {d_{i,j} } \) indicates the distance sum of \( S_{i} \) to all other nodes \( S_{j} \), \( i,\;j = 1,2, \ldots ,k \) and \( j \ne i \). After obtaining \( E_{avg} \) and \( \overline{{d_{i} }} \), any node \( S_{i} \) can deduce its own competition factor \( CB_{i} \) according to Eq. (4), and then it broadcasts a message CANDI_MSG with the format \( Candi(ID_{i} ,CB_{i} ) \), and joins into the candidate set \( V_{candi} \):
$$ CB_{i} = \alpha *\left( {\frac{{E_{i} }}{{E_{avg} }}} \right) + \beta *\left( {\frac{1}{{\overline{{d_{i} }} }}} \right) $$
where \( \alpha + \beta = 1 \), \( \alpha \) and \( \beta \) represents the weight-value of rest energy and average communication distance respectively in \( CB_{i} \). And then, the cluster-head with the largest \( CB_{i} \) can be picked out from the set \( V_{candi} \) by intercomparisons among nodes, and subsequently the other nodes become the members of the current cluster. The cluster-head broadcasts a HEAD_MSG involved its ID \( CH(ID_{i} ) \), and the members send MEM_MSG with their respective ID \( Mem(ID_{i} ) \) to the cluster-head. After that, the process of selecting cluster-head has been completed. After a round of running, if the rest energy of current cluster-head is not less than a given threshold, it would continue its duty of cluster-head. Otherwise, cluster-head selection will be conducted according to the above method.
In this process, it considers both the rest energy and the average communication distance to other nodes, which means, nodes with more energy and closer to other nodes have greater opportunity to become the cluster-head, which ensures the loading balance and prolongs lifetime of the whole network.

4.3 Construction of cluster-head fusion tree

The construction of cluster-head fusion tree can effectively avoid every cluster-head communicating with the sink which may cause more energy consumption. This construction of fusion tree will consider both the rest energy of nodes and the distance between cluster-head to the sink, which means the cluster-head with more energy and closer to the sink will connect to the sink in the first place, forming the primary trunk. Then, it finds out the next suitable node among the remainder cluster-heads, and connects it to the fusion tree to form the next trunk, and so forth, until all the cluster-heads are included in the fusion tree.

Firstly, according to Eq. (5), each cluster-head \( C_{i} \) figures out its weight \( W_{i} \) for the construction of fusion tree:
$$ W_{i} = \chi \left( {\frac{{E_{i} }}{{E_{\hbox{max} } }}} \right) + (1 - \chi )\left( {\frac{{d_{\hbox{min} } }}{{d_{i.s} }}} \right) $$
where \( E_{i} \) represents the rest energy of cluster-head \( C_{i} \), \( E_{\hbox{max} } \) represents its primary energy, \( d_{\hbox{min} } \) represents the shortest distance to the sink among those cluster-heads, \( d_{i,s} \) represents the distance from \( C_{i} \) to the sink, and χ is described as the factor of influence between rest energy and distance to the sink regarding \( W_{i} \). All the cluster-heads send a WEIGHT_MSG with the form \( Weight(ID_{i} ,W_{i} ,x_{i} ,y_{i} ) \) to the sink, which contains the identification \( ID_{i} \), the weight value \( W_{i} \) and the location data \( (x_{i} ,y_{i} ) \).

According to the WEIGHT_MSG, the sink chooses the cluster-head who has the largest \( W_{i} \) as the first trunk node. The next suitable one will be chosen as the next trunk node, and so forth. The sink sends the message LINK_MEG with the form \( Link(ID_{i} ,ID_{j} ) \), where \( ID_{i} \) and \( ID_{j} \) represent the parent node and the child one separately. After receiving the LINK_MEG, the child node will proactively send the message CHILD_MEG with the form \( Child(ID_{i} ) \) to its parent in order to request connection, and consequently a fusion tree will be established by this way.

The pseudo-code of the construction of cluster-head fusion tree is shown as follows:

5 Simulation and analysis

We evaluate the performance of our proposed algorithm (DFRP) through the degree of image difference, the mean square error of energy consumption, the network lifetime, the average rest energy, the average delay and the data delivery success rate. The simulation experiments are conducted in the video sensor networks based on NS2.31 + NSG2.1 platform and the main parameters are given in Table 3. All the data depicted in the following figures are the results of several times average and meanwhile the measurement is completed based on the video stream tools Evalvid and MyEvalvid respectively, which can process each image from distinguished video frame.
Table 3

Simulation parameters



Network size

100 m × 100 m

Sink location

(50, 50) m

Number of multimedia nodes

\( 50 \)

Number of video frames transmitted


Communication radius of multimedia nodes

10 m

Transmission speed of video stream

2 frame/s

Initial energy

2 J

Energy consumption of transmitting one data packet

0.01 J

\( \rho \)

0.26, 0.38, 0.45, 0.6, 0.8 respectively

\( \alpha \)

\( 0.33 \)

\( \beta \)

\( 0.67 \)

\( \chi \)

\( 0.4 \)

The values of α as well as χ are weighted parameters which reflects the importance of different metric when clustering (or constructing data fusion trees). Greater α value indicates that the remaining energy of sensor node is considered in first priority in the clustering process. The sensor node with more remaining energy is more likely to be selected as cluster head. Here α is set as 0.33 and β as 0.67, which means that network delay and communication energy cost (influenced by \( d_{i,s} \)) are considered to be more important than energy balance when clustering. The choice of value of χ is similar to that of α, which values the remaining energy more when generating a data fusion tree. The values of α and χ should be chosen carefully in real-world applications according to the requirements.

5.1 Clustering results

Since the parameter \( \rho \) influences whether the nodes may join a certain cluster, by varying \( \rho \), we can control the number of nodes in a certain cluster. Figure 6a–e show the clustering results when \( \rho = 0.26,\;0.38,\;0.45,\,0.6,\;0.8 \) respectively.
Fig. 6

Clustering results when ρ = 0.26 (a), 0.38 (b), 0.45 (c), 0.6 (d) and 0.8 (e)

From Fig. 6, we can find out that the average size of a cluster increases as the value of \( \rho \) increases, while the number of clusters decreases as \( \rho \) increases. The figures show when the number of video sensor nodes is 50, 0.4 < \( \rho \) < 0.6, it can achieve the reasonable clustering result. In actually, the efficiency of clustering is related with the area size, the counts of the nodes and the value of ρ. Under the given area size and the number of nodes, how to choose the appropriate ρ value is significant. The too small value of ρ may cause the overfull levels of data fusion tree while too large one may lead to the runty tree, and both of them will result in low efficiency of data fusion.

The simulation evaluations of Sect. 5.3, 5.4 and 5.5 are conducted in the scenario based on the clustering result of Fig. 6c (\( \rho \) = 0.45).

5.2 The degree of image difference

Shown as Fig. 3, it assumes that the coordinates of \( S_{1} \) is \( ( - d,0)^{T} \), whose sensing orientation points to X-axis positive orientation, the coordinates of \( S_{2} \) is \( ( - d\cos \theta , - d\sin \theta )^{T} \), whose sensing orientation has an included angle \( \theta \) with X-axis positive orientation, where \( d = 2.5m \), keeping the location of \( S_{1} \) unchanged, and setting \( \theta = \{ - 75^\circ , - 60^\circ , - 45^\circ , - 30^\circ , - 15^\circ ,0^\circ ,15^\circ ,30^\circ ,45^\circ ,60^\circ ,75^\circ \} \) separately. According to Eq. (3), we can figure out the degree of image difference, shown as Fig. 7.
Fig. 7

The degree of image difference

In Fig. 7, the degree of image difference between \( S_{1} \) and \( S_{2} \) increases along with the angle \( \theta \). The larger the degree of image difference is, the smaller the correlation is, and there will be less redundancy in the images captured by \( S_{1} \) and \( S_{2} \).

5.3 The mean square error of energy

The MSE (mean square error) of energy reflects the uniformity of node’s energy consumption, the lower the MSE is, and the better the balance of energy consumption is. Figure 8a, b show the MSE of different schemes with total number of sensors as 50 and 200 respectively. From Fig. 8, we can observe that DFRP has a lower MSE, and the growth is smaller than others, which benefits from that DFRP has a better clustering topology and the cluster-head is rotated periodically within a cluster. NCOF (Alaei and Barcelo-Ordinas 2010???) has a good MSE performance in the beginning; however as time goes by, its performance drops a lot because it only takes FOV into consideration when clustering. Instead, DFRP takes the degree of image difference as the primary metric, which ensure the images captured by nodes in a cluster have fewer difference. So the data fusion process carried out by cluster-heads will consume less energy. In AFST (Luo et al. 2006), the nodes closer to the sink may have heavier burden than others, which means more energy consumption, so the energy distribution is uneven in the network. It can also be noted that the MSE of energy decreases with the growth of node number.
Fig. 8

The mean square error of energy. a 50 nodes. b 200 nodes

Fig. 9

The network lifetime. a 50 nodes. b 200 nodes

5.4 The network lifetime and average rest energy

Both the lifetime cycle and the average rest energy are important factors to evaluate network performance. Here, the energy consumption of the network is studied with a total number of 50 and 200 sensor nodes respectively. We compare our clustering method with AFST (Luo et al. 2006) and NCOF (Alaei and Barcelo-Ordinas 2010???), and the simulation results of network lifetime and average rest energy are shown as Figs. 9 and 10 respectively.
Fig. 10

The average rest energy. a 50 nodes. b 200 nodes

From Figs. 9 and 10, we can observe that DFRP has a better performance compared with AFST and NCOF in both experiment settings. In DFRP, nodes capturing high correlation images may be classified into the same cluster based on the degree of image difference, which enables a lot of redundant data eliminated, and simultaneously enables the data fusion phase carried out by cluster-head with a higher efficiency. In addition, as the cluster-head is rotated periodically, the energy consumption of each node is approximately average, which prolongs the network lifetime.

5.5 The average delay and delivery success rate

Real-time characteristic is considered as one of the most important metrics in WMSN. Whether the multimedia data can accurately arrive at the destinations on time is to partially determine the QoS of the network. The construction of data fusion tree based on clusters has addressed the challenging problem of network delay, which can be demonstrated in Fig. 11 by contrast with the “SA-ACO” algorithm in (Huang et al. 2009). We randomly select 50 video frames to observe their respective network delays and the experiment is repeated 20 times. Overall, DFRP achieves the better performance on network delay, compared with SA-ACO, which owes to the selection of the trunk node in DFRP can be calculated quickly and timely, unlike the SA-ACO algorithm which needs the intelligent training and decisions with the expense of time. Certainly, due to network congestion, cluster classification or packet loss, delays of individual frames in DFRP are not superior to the SA-ACO algorithm; especially those frames from nodes close to the area border have to cross a longer path to reach the destinations. Meanwhile, other simulation experiments are conducted to validate the delivery success rate of the same 50 video frames. As shown in Fig. 12, after 20 times average, we observe that most of frames achieve more than 93% of the delivery success rate, however as some nodes lie on the clusters close to the area border, their delivery success rates may be relatively low.
Fig. 11

The average delay

Fig. 12

The average delivery success rate

6 Conclusions

Both energy efficiency and loading balance are two important metrics in routing protocols in wireless multimedia sensor networks. In DFRP, clustering process based on the degree of image difference makes sure that nodes with similar images will be divided into the same cluster, which enables the data fusion process will consume less energy. And meanwhile, the cluster-head selection considers both the rest energy and the average communication distance to other nodes, which means more suitable nodes will be chosen as the cluster-head and it will be rotated periodically within a cluster. In addition, the construction of cluster-head fusion tree further decreases the energy consumption. So our proposal i.e. DFRP has better applicability and efficiency in the scenarios of wireless multimedia sensor networks, which has been verified through simulation comparisons with other routing protocols. However, improved fusion method should be designed needs to more discussion in order to further enhance the data fusion efficiency by special techniques such as compressive sensing and noise recognition, and accurate selection algorithm of value ρ is another future work will be concerned.



The authors would like to thank the anonymous reviewers of this paper for his/her objective comments and helpful suggestions while at the same time helping us to improve the English spelling and grammar throughout the manuscript. And meanwhile, the subject was sponsored by the National Natural Science Foundation of People’s Republic of China (No. 61672297), the Key Research and Development Program of Jiangsu Province (Social Development Program, No. BE2017742), Jiangsu Natural Science Foundation for Excellent Young Scholar (No. BK20160089), and the Sixth Talent Peaks Project of Jiangsu Province (No. DZXX-017).

Compliance with ethical standards

Conflict of interest

The authors declare that there is no conflict of interests regarding the publication of this paper.


  1. Ahmed, A.A.: A real-time routing protocol with adaptive traffic shaping for multimedia streaming over next-generation of wireless multimedia sensor networks. Pervasive Mob. Comput. 40, 495–511 (2017)CrossRefGoogle Scholar
  2. Alaei, M., Barcelo-Ordinas, J.M.: Node clustering based on overlapping FOVs for wireless multimedia sensor networks. In: Wireless Communications and Networking Conference (WCNC), April 18–21, pp. 1-6. IEEE, Sydney (2010)Google Scholar
  3. Alanazi, A., Elleithy, K.: Real-time QoS routing protocols in wireless multimedia sensor networks: study and analysis. Sensors 15(9), 22209–22233 (2015)CrossRefGoogle Scholar
  4. Attoungble, J.M.K., Okada, K.: A novel energy efficient routing protocol for wireless sensor networks: greedy routing for maximum lifetime. IEICE Trans. Commun. 95(12), 3802–3810 (2012)CrossRefGoogle Scholar
  5. Bhandary, V., Malik, A., Kumar, S.: Routing in wireless multimedia sensor networks: a survey of existing protocols and open research issues. J. Eng. 2016, 27. doi: (Article ID 9608757)
  6. Chen, T.Y., Wei, H.W., Lee, C.R., et al.: Energy efficient geographic routing algorithms in wireless sensor networks. J. Interconnection Netw. 14(01), 135–141 (2016)Google Scholar
  7. De, D., Song, W.Z., Tang, S., et al.: EAR: an energy and activity-aware routing protocol for wireless sensor networks in smart environments. Comput. J. 55(12), 1492–1506 (2012)CrossRefGoogle Scholar
  8. Demir, A.K., Irgan, K., Baydere, S, et al.: Transmitting objects in images with service differentiation based source coding in wireless sensor networks. In: The 9th IEEE Wireless Communications and Mobile Computing conference (IWCMC), Sardinia, pp. 1488–1493. IEEE (2013)Google Scholar
  9. Hasan, M.Z., Al-Rizzo, H., Al-Turjman, F.: A survey on multipath routing protocols for QoS assurances in real-time wireless multimedia sensor networks. IEEE Commun. Surv. Tutor. 19(3), 1424–1456 (2017)CrossRefGoogle Scholar
  10. He, X., Wang, K., Huang, H., Miyazaki, T., Wang, Y., Guo, S.: Green resource allocation based on deep reinforcement learning in content-centric IoT. IEEE Trans. Emerg. Top. Comput. 1–15 (2018a).
  11. He, X., Wang, K., Huang, H., Liu, B.: Qoe-driven big data architecture for smart city. IEEE Commun. Mag. 56(2), 88–93 (2018b)CrossRefGoogle Scholar
  12. Heinzelman, W.R., Chandrakasan, A., Balakrishnan, H.: Energy-efficient communication protocol for wireless microsensor networks. In: Proceedings of the 33rd Annual Hawaii International Conference on System Sciences, Hawaii, pp. 3005–3014. IEEE (2000)Google Scholar
  13. Huang, H.P., Wang, R.C., Sun, L.J., Wang, H.Y., Xiao, F.: Research on tasks schedule and data transmission of video sensor networks based on intelligent agents and intelligent algorithms. J. China Univ. Posts Telecommun. 16(6), 84–91 (2009)CrossRefGoogle Scholar
  14. Intanagonwiwat, C., Govindan, R., Estrin, D.: Directed diffusion: a scalable and robust communication paradigm for sensor networks. In: Proceedings of the 6th Annual International Conference on Mobile Computing and Networking, Boston, pp. 56–67. ACM (2000)Google Scholar
  15. Lee, H., Jang, M., Chang, J.W.: A new energy-efficient cluster-based routing protocol using a representative path in wireless sensor networks. Int. J. Distrib. Sens. Netw. 12 (2014) (Article ID 527928) Google Scholar
  16. Li, B.Y., Chuang, P.J.: Geographic energy-aware non-interfering multipath routing for multimedia transmission in wireless sensor networks. Inf. Sci. 249, 24–37 (2014)CrossRefGoogle Scholar
  17. Lindsey, S., Raghavendra, C.S.: PEGASIS: power-efficient gathering in sensor information systems. In: Aerospace Conference Proceedings, pp. 3-1125-1130. IEEE (2002)Google Scholar
  18. Luo, H., Luo, J., Liu, Y., et al.: Adaptive data fusion for energy efficient routing in wireless sensor networks. IEEE Trans. Comput. 55(10), 1286–1299 (2006)CrossRefGoogle Scholar
  19. Ma, H., Liu, Y.: Correlation based video processing in video sensor networks. In: 2005 International Conference on Wireless Networks, Communications and Mobile Computing, Maui HI, pp. 987–992. IEEE (2005)Google Scholar
  20. Magaia, N., Horta, N., Rui, N., et al.: A multi-objective routing algorithm for wireless multimedia sensor networks. Appl. Soft Comput. 30(C), 104–112 (2015)Google Scholar
  21. Mehrabi, A., Kim, K., Senior Member: Maximizing data collection throughput on a path in energy harvesting sensor networks using a mobile sink. IEEE Trans. Mob. Comput. 14(8), 1–16 (2015)Google Scholar
  22. Nayak, B.K., Rai, S.C., Misra, B.B.: A high connectivity cluster routing for energy-balanced wireless sensor network. In: 2012 IEEE International Conference on Computing, Communication and Applications (ICCCA), Dindigul, pp. 1–6. IEEE (2012)Google Scholar
  23. Necchi, L., Bonivento, A., Lavagno, L., et al.: E2rina: an energy efficient and reliable in-network aggregation for clustered wireless sensor networks. In: Wireless Communications and Networking Conference (WCNC 2007), Kowloon, pp. 3364–3369. IEEE (2007)Google Scholar
  24. Obraczka, K., Manduchi, R., Garcia-Luna-Aveces, J.: Managing the information flow in visual sensor networks. In: The 5th International Symposium on Wireless Personal Multimedia Communications, Honolulu Hi, pp. 1177–1181. IEEE (2002)Google Scholar
  25. Sha, C., Qiu, J., Li, S., Qiang, M., Wang, R.: A type of energy-efficient data gathering method based on single sink moving along fixed points. Peer-to-Peer Netw. Appl. 12, 1–19 (2016)Google Scholar
  26. Soro, S., Heinzelman, W.B.: On the coverage problem in video-based wireless sensor networks. In: The 2nd International Conference on Broadband Networks, Broadband Networks (BroadNets 2005), Boston, pp. 932–939 . IEEE (2005)Google Scholar
  27. Spachos, P., Toumpakaris, D., Hatzinakos, D.: QoS and energy-aware dynamic routing in wireless multimedia sensor networks. In: IEEE International Conference on Communications, pp. 6935–6940. IEEE (2015)Google Scholar
  28. Villas, L.A., Boukerche, A., Ramos, H.S., et al.: DRINA: a lightweight and reliable routing approach for in-network aggregation in wireless sensor networks. IEEE Trans. Comput. 62(4), 676–689 (2013)MathSciNetCrossRefzbMATHGoogle Scholar
  29. Wang, K., Mi, J., Xu, C., Zhu, Q., Shu, L., Deng, D.J.: Real-time load reduction in multimedia big data for mobile Internet. ACM Trans. Multimed. Comput. Commun. Appl. (TOMM). 12(5s), 76 (2016)Google Scholar
  30. Wang, K., Wang, Y., Sun, Y., Guo, S., Wu, J.: Green industrial internet of things architecture: an energy-efficient perspective. IEEE Commun. Mag. 54(12), 48–54 (2016)CrossRefGoogle Scholar

Copyright information

© China Computer Federation (CCF) 2019

Authors and Affiliations

  1. 1.Nanjing University of Posts and TelecommunicationsNanjingChina

Personalised recommendations