This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution license (http://creativecommons.org/licenses/by/3.0/).
One of the important applications in Wireless Sensor Networks (WSNs) is video surveillance that includes the tasks of video data processing and transmission. Processing and transmission of image and video data in WSNs has attracted a lot of attention in recent years. This is known as Wireless Visual Sensor Networks (WVSNs). WVSNs are distributed intelligent systems for collecting image or video data with unique performance, complexity, and quality of service challenges. WVSNs consist of a large number of batterypowered and resource constrained camera nodes. Endtoend delay is a very important Quality of Service (QoS) metric for video surveillance application in WVSNs. How to meet the stringent delay QoS in resource constrained WVSNs is a challenging issue that requires novel distributed and collaborative routing strategies. This paper proposes a NearOptimal Distributed QoS Constrained (NODQC) routing algorithm to achieve an endtoend route with lower delay and higher throughput. A Lagrangian Relaxation (LR)based routing metric that considers the “system perspective” and “user perspective” is proposed to determine the nearoptimal routing paths that satisfy endtoend delay constraints with high system throughput. The empirical results show that the NODQC routing algorithm outperforms others in terms of higher system throughput with lower average endtoend delay and delay jitter. In this paper, for the first time, the algorithm shows how to meet the delay QoS and at the same time how to achieve higher system throughput in stringently resource constrained WVSNs.
In recent years, Wireless Visual Sensor Networks (WVSNs) have emerged as an interesting field. Popular applications are environmental monitoring, seismic detection, military surveillance, medical monitoring, video surveillance, or the Internet of Things (IoT),
In WVSNs for video surveillance, the delay is a very important QoS metric, and the delay is highly dependent on the routing strategies. However, delay routing in WVSN is more difficult than in WSN due to the fact visual data is much larger and complicated than scalar data. Furthermore, batterylimited powered and resource (channel, CPU) constrained camera nodes make this delay routing problem more challenging in WVSNs. [
Delay QoS routing is a challenging issue in WVSNs. Traditional approaches for wireless sensor networks are not efficient or even feasible in a WVSN due to three reasons: first of all, visual data are much bigger and complicated than scalar data, and processing and transmitting visual data will consume much more system resources than scalar data. Secondly, WVSNs need efficient collaborative image processing and coding techniques that can exploit correlation in data collected by adjacent camera sensor nodes. Video coding/compression that has low complexity, produces a low output bandwidth, tolerates losses, and consumes as little power as possible is required. Third, it needs to reliably send the relevant visual data from the camera sensor nodes to aggregation nodes or the sink node in an energyefficient way [
In wireless visual sensor networks for real time video surveillance, sensor nodes need to capture and forward the packets to the sinks within an acceptable delay under the limited resource constraints, including embedded vision processing, data communication, and battery energy issues. It is a challenging issue since it needs to meet the endtoend delay for each application and at the same time optimize the system resource utilization by minimizing the average system delay so that more realtime applications can be granted access in the near future. Existing research on wireless sensor networks address the link scheduling techniques [
The routing algorithm is determined by the shortest path and the routing decision for the OD pair. When a new route is established and all traffic still follows the preexisting path, this phenomenon is called “session routing” because a path remains in force for the entire user session. The routing metric (
QualityofService (QoS) is a major issue that can be divided into several factors including the reliability, mean delay, delay jitter, and bandwidth of packet transmissions [
Routing protocols are categorized into table driven and ondemand, based on route calculation. Researches indicate that a Mobile
Proactive (tabledriven) protocols are based on periodic exchange of control messages and maintaining routing tables. These protocols maintain complete information about the network topology locally. On the other hand, reactive (ondemand) protocols try to discover a route only ondemand, when it is necessary. These protocols usually take more time to find a route compared to proactive protocols. The following sections briefly introduce three well known protocols applied to MANET: Optimized Link State Routing (OLSR),
OLSR is an optimization of a pure link state algorithm for mobile
Being a proactive protocol, routes to all destinations within the network are known and maintained before use. Having the routes available within the standard routing table can be useful for some systems and network applications as there is no route discovery delay associated with finding a new route [
AODV is a combination of ondemand and distance vector hoptohop routing methodology [
All nodes that participate in forwarding this reply to the source node create a forward route to the destination. This route created from each node from source to destination as a hopbyhop state and not the entire route as in source routing [
The main advantage of this protocol is having routes established on demand and that destination sequence numbers are applied to find the latest route to the destination. The connection setup delay is lower. One disadvantage of this protocol is that intermediate nodes can lead to inconsistent routes if the source sequence number is very old and the intermediate nodes have a higher but not the latest destination sequence number, thereby having stale entries. Also, multiple route reply packets in response to a single route request packet can lead to heavy control overhead. Another disadvantage of AODV is unnecessary bandwidth consumption due to periodic beaconing [
DSDV is a hopbyhop distance vector routing protocol. It is a tabledriven routing scheme for
For realtime WVSN applications, the routing strategies should meet the endtoend delay requirements and at the same time optimize the system resources so that more realtime applications could be admitted in the near future. To the best of our knowledge, there is no existing literature addressing these two issues at the same time. In this paper, for the first time, we propose WSN routing strategies that meet the endtoend delay requirement and ate the same time optimize the system resources by minimizing the average delay. The rest of this paper is organized as follows: in Section 3, we propose the routing mathematical model to satisfy the endtoend delay and minimize the average delay in the WSN. In Section 4, we present the solution approaches for our model and develop a heuristic to get a primal feasible solution. In Section 5, computational experiments are performed to verify the solution quality of our approach. Finally, we conclude this paper and outline future research in Sections 6 and 7.
The environment considered here is multichannel WSNs. This paper addresses the problem of determining a good QoS with the average crossnetwork packet delay while taking “user perspective” and “system perspective” into consideration. The following are details of the problem identification:
We propose a channel assignment heuristic and routing algorithm to maximize the transmission rate for each node and enhance system performance by WVSN planning. A similar formulation that aimed to minimize the average crossnetwork packet delay subject to endtoend delay constraints for users was proposed by Yen and Lin [
The following notations list the given parameters and the decision variables of our formulation, as illustrated in
The objective function is illustrated as (IP) to minimize the mean delay on link
Objective function (IP) is actually the summation of average number of packets on each link,
Through Little's Law (
QoS constraints:
Constraint (IP 1) confines that the endtoend delay should be no larger than the maximum allowable endtoend QoS requirement.
Path constraints:
Constraint (IP 2) confines that all the traffic required by each OD pair is transmitted over exactly one candidate path.
Constraint (IP 3) confines that once path
Capacity constraints:
Constraint (IP 4) confines the boundaries of aggregate flow on link
Flow constraints:
Constraint (IP 5) confines the aggregate flow on link
Integer constraints:
Constraint (IP 6) and (IP 7) are the integer constraints of decision variables.
The Lagrangian Relaxation Method was introduced in the 1970s to solve largescale mathematical programming problems and was followed by a large amount of subsequent research [
The main idea of the Lagrangian Relaxation method is to pull apart the model by relaxing (
Through the adoption of Lagrangian Relaxation, a complicated programming problem can be viewed as a small set of easytosolve problems with side constraints. The method simplifies the original problem by decomposing it into several independent subproblems, each with their own constraints and each of which can be further solved by other wellknown algorithms.
For example, if the problem is a minimization problem, the optimal value of the relaxed constraints is always a lower bound on the optimal value of original problem under the relaxed conditions. The lower bound can be improved by adjusting the set of multipliers iteration by iteration to reduce the gap of the solution between the primal problem and the Lagrangian Relaxation. This procedure is also called the Lagrangian Dual problem.
The solution approach to the problem formulation is based on Lagrangian Relaxation. Constraints (IP 1), (IP 3) and (IP 5) are relaxed and multiplied by nonnegative Lagrangian multipliers,
The constraints are relaxed in such a way that the corresponding Lagrangian multipliers
We use the Lagrangian Relaxation Method to relax constraints in the formulation and decompose (LR) into two independent subproblems. In the first SubProblem (SP 1), the nonnegative weight is calculated as (
Besides, the link selection is based on the sum of link mean delay
On the other hand, the multiplier
A brief illustration is shown in
Finally, The LR problem can be decomposed into independent and solvable optimization subproblems which are developed in a sufficient way one by one. The physical meanings of multipliers are also developed and mentioned above to help us to determine relationships of the link mean delay and derivative of queue length. We can get these important parameters to derive the routing assignments in “system perspective” and “user perspective”. The detail solving steps are illustrated in the
In following section, the computational experiments are constructed and implemented to analyze the quality of the solution approach to verify the routing algorithms correctly. The experiments are designed for analyzing the performance of WVSN, if it can be satisfied with lower average endtoend delay and delay jitter to transmit video data by different algorithms.
Each session follows one path for routing its required traffic emulated as a real time surveillance streaming data and is not allowed to change during the holding time. The session arrivals followed a Poisson process, and the holding time of each session is set to be an exponential distribution with average 10 s. The experiment with average session arrival rates of 0.25, 0.5, 1 and 2, and the corresponding packet arrival rates are 40, 20, 10 and 5.
As
The estimate the parameters of our routing metric (shown in
If the received signal strength exceeds the reception threshold, the packet can be successfully received. If received signal strength exceeds the carriersensing threshold, the packet transmission can be sensed. However, the packet cannot be decoded unless signal strength surpasses the reception threshold. Both the transmission range and interference range are calculated by the tworay ground reflection model according to the reception threshold and the carriersensing threshold, respectively.
This paper focuses on the effectiveness of our routing metric and related parameters at the online stage, and its reduction of systemwide impact between each coming session with QoS provisioning. Thus, we use a single channel in the experimental environment to decouple the effect of the channel assignment algorithm and evaluate the performance of our routing algorithm.
In this Section, we present experimental results to demonstrate the effectiveness of our routing algorithm, NODQC. We also evaluate NODQC with other different session types of algorithms under the same average traffic loading in the network, and then compare the performance in terms of average endtoend delay, delay jitter and system throughput with QoS satisfaction. Delay jitter is defined as the variance in the following tables and figures.
Performance evaluation is defined in
If the performance evaluation is taken in the scalability scenario, Average endtoend delay (AE2ED) shows that it has no effect significant difference in smaller number of nodes obviously, but delay increases as the number of nodes increase.
As networks grow in size, the average path of each session lengthens with path selection becoming increasingly important. AE2ED taken in the scalability scenario shows that delay is much less for NODQC as compared to others. As routes break, nodes have to discover new routes which lead to longer E2ED (packets are buffered at the source during route discovery). According to the arc weight on each link is the combination of endtoend delay from the user perspective and average delay from the system perspective, NODQC can choose other less or noncongested paths for the sessions and balance networkwide the traffic loading. This makes the routing strategy more flexible for making routing decisions for new routing constructions. As far as delay is concerned, NODQC performs better than OLSR, AODV, and DSDV with large numbers of nodes.
Hence for real time traffic, NODQC is preferred over OLSR, AODV, and DSDV. This is significant for the fact that, as the variations of packet delay becomes more predictable, the routing mechanisms can factor in that delay to determine whether a packet is lost or not. Furthermore, the NODQC also takes QoS provisioning into account for the coming sessions, thereby causing system throughput with QoS satisfaction to be greater than others in larger networks. By taking into consideration both the perspective of the system as we as well as users, the NODQC has proven itself to have lower average endtoend delay and delay jitter and higher system throughput with QoS satisfaction than other routing algorithms in largescale networks.
By leveraging the Lagrangian Relaxation method, the arc weight on each link can be applied by our routing metrics. It can employ the linkstate routing protocol to construct the shortest or shorter paths within an acceptable delay. Within larger networks, routing strategies can be played to improve system performance with lower average endtoend delay and delay jitter than alternative algorithms (OLSR, AODV, DSDV). NODQC outperforms them in terms of system throughput with QoS satisfaction.
One of the most important parts of this paper is the routing metric and relevant parameters. The Lagrangian Relaxation formulations are applied to the arc weight and to infer the actual meaning of the corresponding multipliers. In addition, Lagrangian multipliers and KarushKuhnTucker Conditions can be used to succinctly describe the arc weight form. This simplifies the process of inference and has the same consequence with Lagrangian Relaxation.
Through the solution of the Lagrangian Relaxation formulation, various QoS requirements can be implemented for different purposes or services. The endtoend delay is presented as a function in our formulation to be the QoS consideration. Distinct QoS metrics (e.g., delay jitter or packet loss rate) can be applied to the Lagrangian Relaxation formulation, and the corresponding multiplier of the routing metric, link mean delay will be replaced by the newly defined QoS metric.
In wireless visual sensor networks for real time video surveillance, sensor nodes need to be forwarded the packets to the sinks within an acceptable delay under limited resource constraints, including embedded vision processing, data communication, battery energy issues. It is a challenging issue since it needs to meet the endtoend delay for each application and at the same time optimize the system resource utilization by minimizing the average system delay so that more realtime applications could be granted in the future. By leveraging the Lagrangian Relaxation method, the arc weight on each link is the combination of endtoend delay from the user perspective and average delay from the system perspective. Thereafter, our routing metric employs the linkstate routing protocol to construct shortest paths within an acceptable delay. Within larger networks, the superiority of the NearOptimal Distributed QoS Constrained (NODQC) routing algorithm is more visible as path selection plays a more important role. Experimental results show that, and especially so in largescale networks, the NODQC not only has lower average endtoend delay and delay jitter than alternative algorithms (OLSR, AODV, DSDV) but also outperforms them in terms of system throughput with QoS satisfaction. In WVSNs for video surveillance, video coding/compression that has low complexity, produces a low output bandwidth, tolerates loss, and consumes as little power as possible is required. Our routing strategies can be applied well in a large scale network with more efficiency and effectiveness.
This work was supported by the National Science Council, Taiwan (grant No. NSC 1012221E002189).
The solution approach to the problem formulation is based on Lagrangian Relaxation. Constraints (IP 1), (IP 3) and (IP 5) are relaxed and multiplied by nonnegative Lagrangian multipliers, respectively. They are added to the objective functions as follows:
The constraints are relaxed in such a way that the corresponding Lagrangian multipliers
Subproblem 1 (related to
The problem can be further decomposed into 
Subproblem 2 (related to
In objective function (SP 2), the last term
In problem (SP 2.1), the first term
Algorithm for solving problem (SP 2.1′).
Solve
Sort these break points and denote as
At each interval
Denote
By examine the  
Typical graph of problem (SP 2.1′).
Finally, this problem can be solved by the algorithm developed by Cheng and Lin in [
The Dual Problem and the Subgradient Method
According to the Weak Lagrangian Duality Theorem [
We use the Subgradient Method is used to solve that [
Getting Primal Feasible Solutions
A primal feasible solution to (IP) by definition must also be a solution to (LR) and satisfy all constraints. Otherwise, it must be modified to be feasible to (IP) by a getting primal feasible solutions heuristic. According to the computational experiments in [
Routing Metric
The cost of passing through the network can be used as a metric. By virtue of the routing algorithm, each router chooses the path with the smallest (shortest) sum of the metrics. Different routing protocols define the metric distinctly. The metric can be distance, number of hops, delay, and so on. Here, the metric is defined as a combination of the link mean delay and the derivative of queue length for each link.
Estimation of Routing Metric Parameters
In [
Other approaches can also estimate critical information of each link in a conceptually straightforward way. During packet reception, packet delay time is defined as the difference between sending time and receiving time. For each adjacent link mean delay, it the sum of each packet delay divided by the number of received packets corresponding to each neighbor node.
For the derivative of queue length, queue length and the corresponding aggregate flow over the link can be recorded at each time interval.
Estimation of the derivative of queue length.
Distributed Routing Algorithm
The routing protocol is proposed and based on the wellknown linkstate routing protocol [
The underlying idea behind this type of routing is that each node shares the state of its neighborhood to every other node in the network. By exchanging local information with all other nodes, each node will have all link states in its own database [
Simple networks with weight information of each link.
Link state packets for all nodes in simple networks.
The underlying idea behind this routing protocol is that only a few changes in present linkstate routing protocol apply to our routing algorithm. Each node must do the following steps in
Whenever a route from source to any destination in the network is required, the source node computes the arc weight for each link by the required traffic and pertinent information of each link in the link state database.
Distributed routing protocol for each node.
Discover its neighbors and learn their network address. Measure the link mean delay and the derivative of queue length to each of its neighbors. Build a packet with the above information at regular intervals. Flood this packet to all other node in the network. Compute the shortest path based on the arc weight form to every other node. Flood the path results to all other nodes included in the path. 
Admission Control Heuristic Algorithm
Through the linkstate routing protocol, the Dijkstra Algorithm can be applied to each node to calculate the routing table and can also compute the shortest path between any two nodes in the network. The algorithm denotes two states for the nodes, tentative and permanent. It is briefly described the steps of Dijkstra algorithm in
Dijkstra routing algorithm [
Start with the local node ( Assign a cost of 0 to this node and make it the first permanent node. Examine each neighboring node of the last permanent node. Assign a cumulative cost to each node and make it tentative. Among the list of tentative nodes
Find the node with the smallest cumulative cost and make it permanent. If a node can be reached from more than one direction
Select the direction with the shortest cumulative cost. Repeat Step 3 to Step 5 until every node becomes permanent. 
To keep our routing as flexible as possible, the Dijkstra Algorithm only computes shortest path for each node pair. Nevertheless, in this paper, it is also important to construct the second shortest path, the third shortest path, and so on; that is so called “K shortest paths.” A significant amount of research has centered on the K shortest paths problem as mentioned in [
The experiment environment is constructed in four grid topologies of 3 × 3, 5 × 5, 7 × 7 and 9 × 9 squares, with one node placed in each intersection point as shown in
This means that, if the Dijkstra Algorithm is the shortest path algorithm, the complexity of the K shortest paths algorithm will be
By determining both the K shortest paths and the K fastest paths, our admission control heuristic algorithm is therefore constructed to give due consideration to both the “system perspective” and “user perspective”. Unfortunately, the K shortest paths may still not satisfy the basic assumption, the QoS provisioning. Thus, other K fastest paths are provided to meet the QoS requirement, and the original arc weight corresponding to each of which cannot exceed a threshold
Admission control heuristic algorithm.
Calculate the fastest path by the predictive link mean delay as the arc weight. If the fastest path can't satisfy the QoS requirement. Reject this traffic. Calculate K shortest paths by the arc weight we proposed, and denote Compare the QoS requirement to the predictive endtoend delay of each K shortest paths. If one of the K shortest path can satisfy the QoS requirement
Choose the shortest one for routing this traffic. Calculate K fastest paths by the predictive link mean delay as the weight of each link, and set If the Kth fastest path can satisfy the QoS requirement and
Choose the fastest one ( Reject this traffic. 
The power regression function can predict the link mean delay of each link. Through admission control heuristic algorithm, any required traffic unable to satisfy the QoS requirement will be rejected with the predictive endtoend delay. This mechanism therefore reduces systemwide impact can be reduced and other flows in advance and ultimately achieve superior performance.
The authors declare no conflict of interest.
Wireless Visual Sensor Networks.
Concept of Lagrangian Relaxation Method.
Lagrangian Relaxation Procedures.
Experimental environment (5 × 5).
Evaluation with different session arrival rates.
Experiment results of routing algorithms (average endtoend delay).
Experiment results of routing algorithms (delay jitter).
Experiment results of routing algorithms (system throughput with QoS).
Routing protocol comparison.
Type  Reactive/Ondemand  Reactive/Ondemand  Proactive/Table driven  Proactive/Table driven 
Optimization with construct shortest paths within an acceptable delay.  Quick adaptation under dynamic link conditions  Optimization: MultiPoint Relays (MPRs)  
Forward broadcast messages during the flooding process  Adds two things to distancevector routing  
Short Description  Lower average endtoend delay and delay jitter  Lower transmission latency  Only partial link state is distributed  Sequence number; avoid loops 
Higher system throughput with QoS satisfaction in large networks  Consume less network bandwidth (less broadcast)  Optimal routes (in terms of number of hops)  Damping; hold advertisements for changes of short duration  
Scalable to large networks  Loopfree property  Suitable for large and dense networks  
Distributed  Yes  Yes  Yes  Yes 
Unidirectional Link Support  Yes  No  Yes  No 
Periodic Broadcast  Yes  Yes  Yes  Yes 
QoS Support  Yes  No  Yes  No 
Advantages  Arc weight of link can be determined and prioritized  
Faster decision making  Lower connection setup time than DSR  Simplicity  
Suitable for large and dense networks  Sequence numbers used for route freshness and loop prevention  Lower route request latency, but higher overhead  
Construct shortest paths within an acceptable delay.  Suitable for large and dense networks  Incremental dumps and settling time used to reduce control overhead  
Lower average endtoend delay and delay jitter  Maintains only active routes  Independent from other protocols  
Higher system throughput with QoS satisfaction in large networks  Uses sequence numbers to determine route age to prevent usage of stale routes  Perform best in network with low to moderate mobility, few nodes and many data sessions  
Low complexity with more efficiency and effectiveness  
Disadvantages  Constrains must be relaxed  Link break detection adds overhead  Lack of security  Counttoinfinity problem 
Proprietary design Applications oriented  Still have possible latency before data transmission can begin  No support for multicast  Not efficient for large Ad hoc networks  
Overhead: maintaining routes to all nodes is often unnecessary  Routing information is broadcasted periodically and incrementally 
Given parameters.
The set of OriginDestination (OD) pairs in the WSNs, where  
The set of directed paths from the origin to the destination of OD pair  
The set of communication links in the WSNs, where  
The number of interference links of link  
The indicator function which is 1 if link  
(  
(  
The mean delay on link  
The maximum allowable endtoend QoS for OD pair 
Decision variables.
1 if path  
1 if link  
( 
Arc weight of each link.

Experimental session types.
Average Holding Time  10  
Average Session Arrival Rate  0.25  0.5  1  2 
Average Number of Active Sessions  2.5  5  10  20 
Packet Arrival Rate  40  20  10  5 
Average Traffic Input  100 (packets/s) 
Parameters.
Transmit and Receive Antenna Gain  1.0 
Transmit and Receive Antenna Height  1.5 (m) 
Reception Threshold  3.625 e^{−;10} 
Carrier Sensing Threshold  1.559 e^{−;11} 
Transmission Range  250 (m) 
Interference Range  550 (m) 
Distance between Each Node  200 (m) 
UDP Packet Size  1,000 (bytes) 
Sending Interval of HELLO Message  2 (s) 
Sending Interval of TC Message  5 (s) 
Recording Interval of Fitting Data  1 (s) 
Number of Fitting Data  10 
K  5 
1.3 
TwoRay Ground Reflection Model.

Evaluation with different session rates (average endtoend delay).
Average Session Rate  0.25  0.5  1  2 
Packet Arrival Rate  40  20  10  5 
Average EndtoEnd Delay  3.325215  3.730487  3.88681  3.942485 
Evaluation with different session rates (system throughput with QoS).
Average Session Rate  0.25  0.5  1  2 
Packet Arrival Rate  40  20  10  5 
System Throughput with QoS  632.062518  521.353393  456.291533  376.553628 
Experiment results of routing algorithms (average endtoend delay).
Routing Algorithms  3 × 3  5 × 5  7 × 7  9 × 9 
NODQC  1.54599  2.550296  3.325215  4.176691 
OLSR  1.28128  2.522533  3.620345  4.792818 
AODV  1.248075  2.339906  3.848888  5.456055 
DSDV  1.220291  2.495117  3.873495  9.171553 
Experiment results of routing algorithms (delay jitter).
Routing Algorithms  3 × 3  5 × 5  7 × 7  9 × 9 
NODQC  0.623741  1.529524  2.789917  4.239667 
OLSR  0.536928  1.463894  3.332403  5.098757 
AODV  0.556566  1.720252  5.129518  13.445979 
DSDV  0.330232  1.418664  3.424205  11.586675 
Experiment results of routing algorithms (system throughput with QoS).
Routing Algorithms  3 × 3  5 × 5  7 × 7  9 × 9 
NODQC  716.849085  673.698465  632.062518  612.322159 
OLSR  753.008325  691.484223  603.130035  556.022380 
AODV  757.683145  700.203701  575.904857  516.320711 
DSDV  773.355103  701.779457  602.953338  509.221205 