A Self-Adaptive QoS Guarantee Scheme for Multimedia Sensor Networks

Abstract

In multimedia sensor networks, there may exist simultaneously two streams: end-to-end stream and event-to-sink stream. It brings a new set of challenges for QoS guarantees in multimedia transmission. In this paper, we first present a self-adaptive QoS guarantee scheme based on two-layer feedback. In transport layer, we partition network into 10 typical states and design a transport protocol, which makes the state transition to the optimal one. Moreover, we design a revenue structure of media stream to describe the relation between QoS and multimedia capturing approaches in the application layer, thus evaluating the optimal scheme of capturing multimedia according to multimedia packet generation ratio calculated by transport layer. A series of simulation experiments using NS2 are performed to demonstrate the effectiveness of our proposed scheme.

1. Introduction

With the complexity of environment, conventional sensor networks [1], which are capable of collecting and processing simple scalar data such as temperature, humidity, or light, cannot provide users with complete and accurate knowledge to the monitored environment. It is desirable to introduce multimedia data (e.g., audio, video, and image) into environment monitoring activities in order to improve efficiency [2]. Motivated by this, multimedia sensor networks (MSNs) come into being. The characteristic of rich information, powerful capabilities, and flexible services doubtlessly make MSNs benefit from a plethora of data-hungry applications such as environment monitoring, traffic surveillance, security, and emergency response [3, 4].

Compared to conventional sensor networks, MSNs can provide more various service forms thanks to their capability of sensing and processing information-immense multimedia data. It is urgent to design effective QoS guarantee scheme to support flexible and reliable service for users. In general, MSNs have two kinds of basic service modes: event driven and query driven. (1) Event Driven Service Mode (E Service). In this mode, multiple sensor nodes preprocess (e.g., compress, identify, and match) the raw environmental data then filter out and transmit valuable semantic data (called as E data) to the sink node. (2) Query-Driven Service Mode (M Service). In this mode, service request can always be initiated by users. MSNs respond to these service requests and return media stream (called as M data) to users.

The above two service modes have different transport models. Opposed to the traditional end-to-end transport model, the event-driven service mode uses event-to-sink transport model [5]. However, in query-driven service mode, media stream follows end-to-end transmission scheme. Generally, these two service modes coexist in multimedia sensor networks. To the best of our knowledge, most of the existing transport protocols [6] merely focus on one specific reliability model, which is not suitable for the case of two transport models' coexistence in MSNs. Furthermore, they pay more attention to the transport QoS of simple scalar data not complex multimedia data.

In this paper, we present a transport protocol, which mainly aims at the coexistence of event-to-sink and end-to-end transport models in MSNs. Our goal is to maximize the QoS of E and M service while guaranteeing the operating efficiency (low congestion and low-energy consumption) of MSNs. First, we define the QoS concepts of two service modes for the above two service models. Second, according to E and M packets' arrival rates on sink node and network congestion level, we categorize 10 typical states of MSNs. The protocol adjusts E and M packets' transmission rates by states of MSNs, thus it optimizes the QoS metrics.

E packets' transmission rate can be adjusted by the varied report frequency of sensor nodes, and a new report frequency can be sent to a corresponding sensor node directly. However, for M service, because of its feature of multiparameter (e.g., for video stream, key parameters include color, resolution, and frame rate), a same M packets' transmission rate may correspond to different configurations of media parameters resulting in different M service's QoS. Therefore, the objective of adjusting M packets' transmission rate is to find an optimal configuration of media parameters to maximize QoS value while ensuring that the new transmission rate equals to the adjusted one. Here, we turn the pending problem into a conditional extremum one and feed the configuration of media parameters back to the corresponding sensor node in application layer. In this way, we relate transport protocol with the optimal configuration of media parameters by two-layer (transport layer/application layer) feedback, which could maximize the QoS values of both event-to-sink and end-to-end models.

The remainder of this paper is organized as follows. Section 2 overviews the related works. Section 3 defines the QoS concept of E and M services in MSNs. Section 4 describes a self-adaptive QoS guarantee scheme based on two-layer feedback. In Section 5, we carry out extensive experiments using NS2 simulator to evaluate the performance of our transport protocol. Finally, we conclude the paper in Section 6.

2. Related Work

We overview the related works from two aspects: QoS guarantees in application layer and QoS guarantees in transport layer.

2.1. QoS Guarantees in Application Layer

From the perspective of application, QoS parameters include sampling rate, delay, packet loss rate, and the number of active nodes. In other words, the application requires the gathering rate, node deployment, and the measurement accuracy of sensor nodes related to the application-level QoS. Kogekar et al. [7] presented a dynamic reconfiguration method in wireless sensor networks to guarantee application-level QoS under energy-constrained condition. However, QoS guarantees in application layer of MSNs have not been addressed. Alwan et al. [8] proposed an adaptive mobile architecture for multimedia networks, which focused on two layers of key importance in multimedia wireless network design, namely, compression algorithms in the voice/video application layer and routing algorithms in network layer. Self-adaptation and its corresponding algorithms can be implemented in application layer and MAC layer, respectively. However, the function of self-adaptation provided by application layer is not visible to users. Foresti and Snidaro [9] proposed an application-level negotiation algorithm for multimedia networks, with two distinct advantages: (1) globalization of local parameters, which makes resource reservation flexible to let users reserve QoS parameters directly; (2) localization of global parameters, which makes resource reservation simple and makes it easy to perform the optimal delay assignment.

2.2. QoS Guarantees in Transport Layer

Many pioneering papers have addressed the issue of transport protocol [10] for conventional sensor networks. Reliable Multisegment Transport (RMST) [11] is a selective NACK-based protocol designed to run on top of the directed diffusion network layer protocol with two modes of [5] introduced the concept of “end-to-end reliable transport” and presented an event-to-sink reliable transport (ESRT) protocol. Reliable event detection at the sink is based on collective information provided by source nodes instead of on any individual report; however, the conventional end-to-end transport protocol is not applicable for event-driven traffic. Pump Slowly, Fetch Quickly (PSFQ) [12] is a transport protocol suitable for reliable data transmission, which is a simple approach with minimal requirements on the routing infrastructure using the minimum signaling, thus it reduces communication cost for data reliability and allows successful operation even under highly error-prone conditions. The delay-aware reliable transport (DART) [13] protocol is presented for timely and reliably transporting event features from the sensor field to the sink with the minimum energy consumption; the objective of DART is not QoS optimization. The log-normal shadowing model (LNSM) [14] captures the effects caused by the imperfect antennas and environmental obstructions. However, it is incapable of modeling the path loss for the continuously varying airborne height, and modeling as a function of elevation angle is more convenient.

3. Basic Concepts

Quality of Service (QoS) defines the satisfaction degree of service provided by network service provider. In this paper, we use utility [15] as an important QoS metric in MSNs, which is different from some traditional QoS metrics (e.g., average data packet delay, loss rate, resource utilization, etc.). The utility of service consists of two parts: (1) transmission efficiency, defined as the ratio of transmission rate to receive rate, is inversely related to packet loss rate; (2) user's revenue, defined as the satisfaction degree of application requirements comes from data receivers, is closely related to the number of data packets received per unit time.

As shown in Figure 1, supposing that at a given moment, there are i nodes in an event region to cooperatively gather abnormal event information to obtain better detect reliability, and j is the number of nodes providing real-time media streams for different M services. Hence, we can get that there are one E service and $j M$ services simultaneously in MSNs. To simplify the pending problem, we assume that each M data stream has the same transmission rate. $s_{m}$ and $s_{e}$ are the transmission rates of M and E packets on each sensor node, respectively, and $S (s_{m}, s_{e})$ denotes a transmission vector. $r_{m}$ and $r_{e}$ are the receive rates of M and E packets on sink node, respectively.

Figure 1

Topology of multimedia sensor network with E and M data streams.

3.1. Transmission Efficiency

For each sensor node, the transmission efficiencies of M and E services are defined as $η_{m} = r_{m} / s_{m}$ and $η_{e} = r_{e} / s_{e}$ , respectively, which are close to 1 when the network load is low. When the network is heavily loaded, $r_{m}$ and $r_{e}$ are influenced by both $s_{m}$ and $s_{e}$ ; in other words, $r_{m}$ and $r_{e}$ are determined by the transmission vector S. In the following simulations, we will discuss the relation between $r_{m}$ and S, $r_{e}$ , and S.

3.2. Revenue of E Service

For E service based on event-to-sink model, its revenue indicates the accuracy of event detection. Akan and Akyildiz [5] discussed the relation between event detection accuracy ζ and E packet receive rate $r_{e}$ , as illustrated in Figure 2. The event detection accuracy is important to provide high QoS performance. As we utilize the function $ζ = ψ_{e} (r_{e})$ , ζ will gradually increase with the increase of $r_{e}$ . Intuitively, the greater event information reaches the sink node per unit time, the lower event distortion is, and the higher the event detection accuracy will become. However, ζ increases very slowly when $r_{e}$ goes above the threshold t. $ζ = ψ_{e} (r_{e})$ denotes the upper limit of event detection accuracy, where t is the threshold of $r_{e}$ .

Figure 2

Accuracy function of event detection.

3.3. Revenue of M Service

For M service, its revenue refers to the traditional multimedia QoS metrics (e.g., color, resolution, and frame rate for video streaming). In our prior work [16], we proposed a method to describe QoS requirements for media stream in MSNs. This method categorized some principle parameters which have an influence on multimedia service. The users can assign different weights to each parameter by application requirements, and different weights reflect the importance level of parameters in different applications.

In this paper, we use the prior method to describe the relationship among the revenue of media stream, the transmission rate of M packets, and the configuration of parameters. As shown in Figure 3(a), the structure is composed of three layers; the root node is the revenue ζ of multimedia application. The second-layer nodes are media parameters $(p_{1}, p_{2}, \dots, p_{m})$ , which affect the transmission rate of multimedia data. The impact of each parameter on the application revenue is quantized by a vector of weight values $(w_{1}, w_{2}, \dots, w_{m})$ . Each parameter can have multiple possible values, for example, $v_{m, n}$ is the nth value of parameter $p_{m}$ , and $v_{m}$ denotes the current value of parameter $p_{m}$ . The revenue contribution of $v_{m}$ is denoted as $q_{m}$ . Figure 3(b) shows a case for video service.

Figure 3

The revenue structure of media stream.

There is typically a positive correlation between the value of each parameter and the revenue value. For example, the application revenue, when a node receives video stream with 30 frames/sec, is obviously greater than that with 20 frames/sec. If we assume the contribution to the revenue as 1 when the value of $p_{m}$ is the largest $v_{m, \max}$ , then revenue contribution of $v_{m, k}$ is $v_{m, k} / v_{m, \max}$ . In the above case, when a sensor node collects video frames at the maximum rate of 30 frames/sec, then its revenue contribution is 1 (we consider 10 frames as a level), but when the node collects video stream with 20 frames/sec, then its revenue contribution becomes 2/3.

Since the numbers of pixel along x-axis and y-axis determine the resolution of video stream, we express the resolution with two parameters ( $p_{x}$ and $p_{y}$ ), typically with the ratio of 4 : 3. Generally, the revenue contributions of $q_{x}$ and $q_{y}$ are equal; therefore, the revenue contribution of the resolution is $q_{x}^{2}$ or $q_{y}^{2}$ .

Definition 1.

If the revenue of a stream media service is determined by n parameters, then the parameter configuration vector $V (v_{1}, v_{2}, \dots, v_{n})$ is a set of parameter values, and $(q_{1}, q_{2}, \dots, q_{n})$ is the vector of the revenue contribution values for the parameter configuration vector.

For each V, the revenue can be calculated as

\begin{matrix} ζ_{V} = \sum_{i = 1}^{n} w_{i} q_{i} \end{matrix} .

(1)

For M service, the revenue is effected by two factors: the parameter configuration of stream media V and the receive rate of M packets $r_{m}$ . In general, the more the data packet received per unit time is, the higher the quality of media stream will become. $ψ_{m} (r_{m}, V)$ is the revenue function of M service, and assume that $ψ_{m} (r_{m}, V)$ increases linearly with $r_{m}$ , then the revenue of M service can be calculated as

\begin{matrix} ψ_{m} (r_{m}, V) = r_{m} (\sum_{i = 1}^{n} w_{i} q_{i}) . \end{matrix}

(2)

The problem of QoS guarantees in multimedia sensor networks can be described as follows. Without network congestion, how to guarantee the revenue of E service while maximizing the revenue of M service by selecting the optimal configuration of parameters of M data?

4. Self-Adaptive QoS Guarantee Scheme

First, we propose a self-adaptive QoS guarantee structure based on two-layer feedback. The QoS guarantee structure consists of two layers: transport layer and application layer. In transport layer, transport protocol adjusts the transmission rates of M and E services, in order to guarantee QoS of these two services under the condition of no network congestion. When the transmission rate of E service changes, the transport layer feeds back a new transmission rate to the corresponding sensor node; when the transmission rate of M service changes, application layer generates the optimal configuration of media parameters by the new transmission rate then feeds back the configuration to the corresponding sensor node in application layer.

The kernel of our self-adaptive QoS guarantee scheme is to adjust transmission rate of M and E packets and generate the optimal parameter configuration of M data. We use a transport protocol and configuration method of streaming media parameters to solve the above problems. Considering sensor node's limitation on both energy and computation, the main computation is executed on the sink node.

4.1. Partitioning of Value Space of S

To adjust transmission vector S, we first analyze the relation among S, $r_{m}$ , $r_{e}$ , and network congestion status. According to a series of simulation results (see Section 5.2), we get the partitioning of typical region of transmission space, as shown in Figure 4. In value space of S, a curve composed of all the corner points of contour lines $r_{m} = j 1 0^{x}$ (M data provided by j nodes, $x < a_{t}$ ) or $r_{e} = i 1 0^{y}$ (E data provided by i nodes, $y < b_{t}$ ) is denoted as the congestion boundary of S. If a point $(x, y)$ is in the shadowed area, then there is network congestion, and vice versa.

Figure 4

Partitioning of typical region of transmission space.

Assume that when $y = δ$ , $r_{e}$ approaches the upper limit of event detection accuracy, that is, $T = i 1 0^{δ}$ , we denote the region between the line $y = δ + σ$ and $y = δ - σ$ within the congestion boundary as optimal region of E service (see regions 6 and 9 in Figure 4). σ is a minimum value, the corresponding $r_{e}$ of $y = δ$ , $y = δ + σ$ , and $y = δ - σ$ are $i 1 0^{δ}$ , $i 1 0^{δ + σ}$ , and $i 1 0^{δ - σ}$ , represented as $r_{e}^{*}$ , $r_{e}^{+}$ , and $r_{e}^{-}$ . The lines $y = δ$ , $y = δ + σ$ , and $y = δ - σ$ intersect the congestion boundary at $(m_{δ}, δ)$ , $(m_{δ + σ}, δ + σ)$ , and $(m_{δ - σ}, δ - σ)$ , respectively. We denote the region between the line $x = m_{δ + σ}$ and $x = m_{δ - σ}$ within the congestion boundary as optimal region of M service (see regions 7 and 9 in Figure 4). The corresponding $r_{m}$ are $j 1 0^{m_{δ}}$ , $j 1 0^{m_{δ + σ}}$ , and $j 1 0^{m_{δ - σ}}$ , denoted as $r_{m}^{*}$ , $r_{m}^{+}$ , and $r_{m}^{-}$ . The optimal region of total network utility is the intersection of the optimal region of M service and E service (see region 9 in Figure 4).

The line $r_{e} = T$ intersects the congestion boundary at the point $(m_{δ}, δ)$ , $r_{m}$ and $r_{e}$ are equal to $r_{m}^{*}$ and $r_{e}^{*}$ . Figure 4 shows the contour lines $r_{e} = r_{e}^{*}$ and $r_{m} = r_{m}^{*}$ , the congestion boundary, and the optimal region of E service and M service.

4.2. Transport Protocol

From a series of simulation, we can find that when receive rates $r_{m}$ , $r_{e}$ are within a certain range, there exists no network congestion, and two kinds of services' QoS can be satisfied at the same time. Here, we can divide the value space of S into 10 regions, corresponding to 10 states, each state is denoted by a 3-tuple (C, E, M) (A set of simulations tesify that the states (N, H, H), (Y, O, -), (Y, -, O), (N, O, H), and (N, H, O) do not exist, so we will not discuss it here). As shown in Table 1, C denotes the congestion state of network. C = Y means that network congestion occurs; otherwise, C = N. E and M denote the revenue of E service and M service, respectively. When there is no network congestion, E and M have three kinds of values. For both E and M services, H, L, and O denote $r_{e} > r_{e}^{+} (r_{m} > r_{m}^{+})$ , $r_{e} < r_{e}^{-} (r_{m} < r_{m}^{-})$ , and $r_{e}^{-} \leq r_{e} \leq r_{e}^{+} (r_{m}^{-} \leq r_{m} \leq r_{m}^{+})$ . Once congestion occurs, the values of E and M cannot be O. H and L means $r_{e} \geq r_{e}^{*} (r_{m} \geq r_{m}^{*})$ and $r_{e} < r_{e}^{*} (r_{m} < r_{m}^{*})$ , respectively. When $s_{m}$ and $s_{e}$ are relatively large, a large number of data packets are dropped, and the receive rates of two kind of services cannot satisfy the requirements. That is, the network status is in state 10 (Y, L, L). We do not show the region corresponding to state 10 in Figure 4 due to limited range of x and y.

Table 1

Network states.

Region	C	E	M
1	N	H	L
2	N	L	L
3	N	L	H
4	Y	L	H
5	Y	H	L
6	Y	H	H
7	Y	L	L
8	N	O	L
9	N	L	O
10	N	O	O

According to network states, we design a transport control protocol, which can bring the network into the optimal state (N, O, O) by adjusting $S (s_{e}, s_{m})$ regardless of the initial state of the network. This protocol is primarily run on the sink node. In order to reduce the complexity of transport protocol, we merge state 5 and state 6 into (Y, H, -), state 4 and state 7 into (Y, L, -). Based on these states, we present a transport protocol, which can transfer the network into the optimal state (N, O, O) by adjusting $S (s_{e}, s_{m})$ regardless of the initial network state. This transport protocol is mainly run on the sink node. Our algorithm can be described in Algorithm 1.

Algorithm 1: Trust-based algorithm for fault-tolerant data aggregation.

Input: $r_{e}$ , $s_{e}$ , $r_{m}$ , $s_{m}$

/* $r_{e}^{*}$ , $r_{m}^{*}$ , $r_{e}^{+}$ , $r_{m}^{+}$ , $r_{e}^{-}$ , $r_{m}^{-}$ are constants */

Output: $s_{e}$ , $s_{m}$

(1) If (CONGESTION);

(2) If ( $r_{e} \geq r_{e}^{*}$ )

/* State(Y, H, -), $s_{e}$ is reduced by the step-by-step

approximation methods. */;

(3) $s_{e} = \frac{1}{2} s_{e} (1 + r_{e}^{-} / r_{e})$ ;

(4) ELSE

/* State $(Y, L, -)$ , $s_{m}$ is reduced by the split-half

methods.*/;

(5) $s_{m} = s_{m} / 2$ ;

(6) End;

(7) Else if (NO CONGESTION);

(8) If( $r_{e}^{-} \leq r_{e} \leq r_{e}^{+}$ and $r_{m} < r_{m}^{-}$ )

/* State (N, O, L), $s_{m}$ is increased linearly. */;

(9) $s_{m} = s_{m} (r_{m}^{*} / r_{m})$ ;

(10) Else if $(r_{e} < r_{e}^{-}$ and $r_{m}^{-} \leq r_{m} \leq r_{m}^{+})$

/* State (N, L, O), $s_{e}$ is increased linearly. */;

(11) $s_{e} = s_{e} (r_{e}^{*} / r_{e})$ ;

(12) Else if ( $r_{e}^{-} \leq r_{e} \leq r_{e}^{+}$ and $r_{m}^{-} \leq r_{m} \leq r_{m}^{+}$ )

/* State (N, O, O), the state remains unchanged. */;

(13) Else

/* State (N, L, L) or (N, H, L) or (N, L, H), $s_{m}$ and $s_{e}$

are adjusted linearly. */;

(14) $s_{m} = s_{m} (r_{m}^{*} / r_{m})$ , $s_{e} = s_{e} (r_{e}^{*} / r_{e})$ ;

(15) End;

(16) End;

4.3. Optimal Solution for Media Gathering Configuration Vector

Assume the length of M packet is T, we can get $s_{m}$ corresponding to the parameter configuration vector V as follows:

\begin{matrix} s_{m} = \frac{D (V)}{T}, \end{matrix}

(3)

where

D (V)

is the function for computing the amount of data generated from multimedia sensor nodes

\begin{matrix} D (V) = \prod_{i = 1}^{n} v_{i} . \end{matrix}

(4)

For example, consider a video stream with resolution $X * Y$ as $800 * 600$ , color as 32 bit color, and frame rate as 30 fps. Because resolution is determined by pixel amounts along x-axis and y-axis, we regard resolution as two parameters. Hence, the amount of data generated from it is $D (v_{1}, v_{2}, v_{3}, v_{4}) = κ \prod_{i = 1}^{4} v_{i} = κ \times (800 \times 600) \times 32 \times 30$ . The video stream is compressed generally, and the compression ratio is assumed as κ.

According to (2), (3), and (4), we can formalize the Problem as follows: to find a parameter configuration vector $V^{*}$ , satisfying $D (V^{*}) / T = s_{m}$ , and for any V which satisfy $D (V) / T = s_{m}$ satisfying $ζ_{V^{*}} \geq ζ_{V}$ .

If the maximum value of $p_{i}$ is $v_{i, \max}$ , then we can get $v_{i} = q_{i} v_{i, \max}$ . Solving $(q_{1}, q_{2}, \dots, q_{n}) (0 < q_{i} < 1 (i = 1,2, n))$ , satisfying $\sum_{i = 1}^{n} w_{i} q_{i}$ is maximal when $\prod_{i = 1}^{n} q_{i} v_{i, \max} = T s_{m}$ . This is a conditional extremum problem of multivariate function. We only obtain the extremum as the minimum of QoS by Langrange Multiplier method. So, we turn this problem into an unconditional extremum one. Let

\begin{matrix} ϑ = \frac{T s_{m}}{\prod_{i = 1}^{n} v_{i, \max}}, \end{matrix}

(5)

then

q_{n} = \frac{ϑ}{\prod_{i = 1}^{n - 1} q_{i}} .

(6)

QoS can be described as follows:

\begin{matrix} ζ_{V} = \sum_{i = 1}^{n - 1} w_{i} q_{i} + \frac{w_{n} ϑ}{\prod_{i = 1}^{n - 1} q_{i}} . \end{matrix}

(7)

Solving the group of equations

\begin{matrix} \frac{\partial ζ_{V}}{\partial q_{1}}, \frac{\partial ζ_{V}}{\partial q_{2}}, \dots, \frac{\partial ζ_{V}}{\partial q_{n - 1}} . \end{matrix}

(8)

We can get $(q_{1}, q_{2}, \dots, q_{n - 1})$ corresponding to an extremum point of $ζ_{V}$ .

Since $0 < q_{i} < 1 (i = 1,2, \dots, n)$ , the range of $(q_{1}, q_{2}, \dots, q_{n - 1})$ is $ϑ \leq \prod_{i = 1}^{n - 1} q_{i} \leq 1$ and $0 < q_{i} < 1 (i = 1,2, \dots, n)$ in $n - 1$ dimension space. Solving for the boundary value of D, that is, when $(q_{1} q_{2} \dots q_{k - 1} q_{k + 1} \dots q_{n}) = ϑ$ and $0 < q_{i} < 1 (i = 1,2, \dots, k - 1, k + 1, \dots, n)$ , we can calculate $(q_{1} q_{2} \dots q_{k - 1} q_{k + 1} \dots q_{n})$ corresponding to the maximum of $\sum_{i = 1}^{k - 1} w_{i} q_{i} + w_{k} + \sum_{i = k + 1}^{n} w_{i} q_{i}$ . Repeat the above method until we can reduce the dimension of this question to 1. The maximum value of $ζ_{V}$ in all extremum points is the solution. We will give a detailed solution for gathering configuration vector with 3 parameters.

5. Experimental Results and Performance Analysis

5.1. Simulation Environment

In order to perform empirical evaluation of our transport protocol, we conduct simulations using NS2. All of our simulations are based on the same scenario, where 50 nodes are randomly deployed in the target region of $(500 \times 500) m^{2}$ . Table 2 summarizes the configuration parameters of nodes. We execute 4 groups of experiments based on the topology above, and the settings are illustrated in Table 3. For each group of experiments, we obtain $\log s_{e}$ and $\log s_{m}$ values with the interval of 0.2 within $[- 1,2.4]$ , that is, for one group $(i, j)$ , we carry out $18 \times 18$ experiments corresponding to different network loads.

Table 2

Configuration parameters of nodes.

Name of Parameters	Value of Parameters
Radio propagation	Two-ray ground
Interface queue	PriQueue
Transmission range	200 m
IFQ length	50 packets
E packet length	50 bytes
M packet length	1000 bytes

Table 3

Data generation scheme.

Group id	Number of event nodes (i)	Number of stream nodes (j)
1	25	5
2	35	5
3	35	10
4	45	10

5.2. Relationship between Receive Rate and Transmission Vector

In order to partition the value space of $S (s_{m}, s_{e})$ into multiple typical regions, we need to find out the relationship among $r_{m}$ , $r_{e}$ , and transmission vector S using simulations. Here, we present a group of typical experimental results. The relationship among $r_{m}$ , $r_{e}$ , and transmission vector S is shown in Figure 5. The x-axis and y-axis are both log scaled. The plane represented by the x-axis and y-axis corresponds to the value space of S. The point $(x, y)$ is mapped into the transmission vector $(1 0^{x}, 1 0^{y})$ . The z-axis in Figure 5 denote $r_{m}$ and $r_{e}$ , respectively. We can see that $r_{m}$ and $r_{e}$ do not increase linearly with the $s_{m}$ and $s_{e}$ , since network congestion occurs once S exceeds a threshold.

Figure 5

Relationship between receive rate and transmission vector.

In order to explain the relationship among $r_{m}$ , $r_{e}$ , and S, as well as the distribution of congestion critical points in the value space of S, we discuss three points: (1) The relationships between $r_{m}$ and $s_{e}$ , $r_{e}$ , and $s_{e}$ , where $s_{m}$ is a fixed value; (2) the relationships between $r_{m}$ and $s_{m}$ , $r_{e}$ , and $s_{m}$ , where $s_{e}$ is a fixed value; (3) the relationship between $s_{e}$ and $s_{m}$ , where $r_{m}$ and $r_{e}$ are fixed values. Through the detailed analysis of the three aspects above, we divide the value space of S into 10 regions.

5.3. Case of Transport Protocol Operation

In this section, we carried out 4 groups of experiments starting with 4 different initial states, as shown in Figure 6. All experiments are based on the same topology and adopt the data generation scheme above.

Figure 6

State transition process with initial state.

For E service, we set $r_{e}^{*} = 40$ (packets) in order to satisfy the veracity of detecting event, and the optimal value of $r_{e}$ is within $[35,45]$ , we get the optimal value of $r_{m}$ belongs to $[16,20]$ . Hence, the two corresponding points on the congestion boundary are $(0.3,0)$ and $(0.2,0.2)$ , respectively.

Figures 6(a)–6(d) represent the state transition process with 4 kinds of initial states: (N, L, L), (N, H, L), (N, L, O), and (Y, H, -), respectively. The value of y-axis is $y_{m} = r_{m} / r_{m}^{*}$ for M service, and $y_{e} = r_{e} / r_{e}^{*}$ for E service. The value of x-axis is the transition amount of network states.

When $S = (1 0^{- 0.6}, 1 0^{- 0.6})$ and the initial state is (N, L, L), $r_{m} = 2.6$ , $r_{e} = 9$ , the corresponding value of y is 0.14 and 0.25. In this state, the network is in noncongestion status while $r_{m}$ and $r_{e}$ are relatively low. We can increase both $s_{m}$ and $s_{e}$ to $(1 0^{0.24}, 1 0^{0.05})$ for achieving the optimal state by one step, here, $y = (0.97,0.96)$ . Then, we improve the transmission efficiency while obtaining the optimal service revenue.

When $S = (1 0^{- 0.8}, 1 0^{0.4})$ and the initial state is (N, H, L), $y = (0.09,2.1)$ , we can get the new value $S = (1 0^{0.25}, 1 0^{0.07})$ . In this state, there is no network congestion. $r_{m}$ is relatively low, while $r_{e}$ is higher than the requirement. We can regulate to achieve the optimal state by one step by decreasing $s_{e}$ and increasing $s_{m}$ .

When $S = (1 0^{0.2}, 1 0^{- 0.8})$ and the initial state is (N, L, O), the optimal value of $r_{m}$ is achieved and no network congestion occurs. $r_{e}$ is lower than the requirement, there exists a linear growth law between transmission rate and receive rate of data, thus $s_{e}$ needs to be increased. We can regulate to achieve the optimal state by one step through increasing $s_{e}$ to 0.03, here, $y_{e} = 0.92$ .

When $S = (1 0^{2}, 1 0^{2})$ and the initial state is (Y, H, -), network congestion occurs, and $r_{e}$ is higher than the requirement. There is no linear relationship between transmission rate and receiving rate of data (see Figure 4, this region on the top of the optimal region). Consequently, we should decrease the transmission rate of event. However, to ensure event detection accuracy, we decrease $s_{e}$ using step-by-step approximation method. After decreasing $s_{e}$ , $r_{e} > e$ , namely, the state (Y, H, -) remains unchanged (from step (1) to step (2)). Meanwhile, during the process of decreasing $s_{e}$ , $r_{m}$ will increase, then the current state will transfer to (Y, L, -) (from step (2) to step (3)). At this point, we can decrease $s_{m}$ using split-half method. The rapid decrease of $s_{m}$ will cause the increase of $r_{e}$ , then the network state goes back (Y, H, -) (from step (3) to step (4)). Hence, decreasing $s_{e}$ using step-by-step approximation method and decreasing $s_{m}$ using split-half method step by step, the network state will fluctuate between state (Y, H, -) and state (Y, L, -). Once the network congestion is eliminated, we can regulate the state to the optimal one. The state transition process is shown in Figure 6(d).

5.4. Generation Case of Parameters Configuration

Assume that the revenue structure for video streaming service is as shown in Figure 3(b), there are four parameters (color, x-axis and y-axis resolutions, frame rate) in the configuration. Generally, x-axis and y-axis resolutions have a fixed ratio, and their revenue contributions are equal, so we can merge the two parameters into the resolutions. There are three types of parameters: color $(p_{1})$ , resolution $(p_{2})$ , and frame rate $(p_{3})$ .

We calculate the extremum points of the three equations above, respectively. Set $ϑ = 0.25$ and the weighted values are set as shown in Figure 3(b). We also calculate the values of $ζ_{V}$ of all the extremum points and get a group of optimal value $(q_{1}, q_{2}, q_{3}) = (1,0.25,1)$ . The parameters corresponding to this group of optimal value are as follows: color is 32 bit color, resolution is $200 \times 150$ , and frame rate is 30 fps.

Meanwhile, due to constraints of physical devices, the value range of q is not continuous but consists of discrete points within $[0,1]$ . Therefore, we can find out the best approximation value of q by means of the solution to generate the corresponding configuration scheme.

6. Conclusions

In this paper, we propose a self-adaptive QoS guarantee scheme based on two-layer feedback for multimedia sensor networks, which combines transport control protocol and the optimal configuration of media parameter in application layer. By analyzing the relation between M and E packet's transmission rate and receive rate, the network state is partitioned into 10 typical states according to the utility of two services and network congestion status. Furthermore, the transport layer maintains a network state transition map, adjusts the transmission rates, and feeds them back to the corresponding sensor nodes by the current network state. The application layer maintains a hybrid structure, which describes the relation of multimedia gathering and receipt parameters. Application layer calculates the optimal configuration of media parameter and feeds back to the sensor node according to M packet's transmission rate.

Footnotes

Acknowledgments

This work is supported by the Natural Science Foundation of China under Grant No. 61070205 and No. 61070206, the program of New Century Excellent Talents in University of China under Grant No. NCET-08-0737, the Beijing National Natural Science Foundation under Grant No. 4092030, and the Cosponsored Project of Beijing Committee of Education.

References

Akyildiz

I. F.

Sankarasubramaniam

Cayirci

Wireless sensor networks: a survey

Computer Networks 2002 38 4 393 422

2-s2.0-0037086890

10.1016/S1389-1286(01)00302-4

Akyildiz

I. F.

Melodia

Chowdury

K. R.

Wireless multimedia sensor networks: a survey

IEEE Wireless Communications 2007 14 6 32 39

2-s2.0-37249066703

10.1109/MWC.2007.4407225

Holman

Stanley

Ozkan-Haller

Applying video sensor networks to nearshore environment monitoring

IEEE Transactions on Pervasive Computing 2003 2 4 14 21

Pavlidou

Grammalidis

Dimitropoulos

Using intelligent digital cameras to monitor aerodrome surface traffic

IEEE Transactions on Intelligent Systems and Their Applicationss 2005 20 3 76 81

Akan

O. B.

Akyildiz

I. F.

Event-to-sink reliable transport in wireless sensor networks

IEEE/ACM Transactions on Networking 2005 13 5 1003 1016

2-s2.0-26644458553

10.1109/TNET.2005.857076

Iyer

Y. G.

Gandham

Venkatesan

STCP: a generic transport layer protocol for wireless sensor networks

Proceedings of the 14th International Conference on Computer Communications and Networks (ICCCN ′05)

2005

449 454

2-s2.0-33748361370

10.1109/ICCCN.2005.1523908

Kogekar

Neema

Eames

Koutsoukos

Ledeczi

Maroti

Constraint-guided dynamic reconfiguration in sensor networks

Proceedings of the 3rd International Symposium on Information Processing in Sensor Networks (IPSN ′04)

April 2004

379 387

2-s2.0-3042698232

Alwan

Bagrodia

Bambos

Gerla

Kleinrock

Short

Villasenor

Adaptive mobile multimedia networks

IEEE Personal Communications 1996 3 2 34 50

2-s2.0-0030128849

Foresti

Snidaro

A distributed sensor network for video surveillance of outdoor environments

Proceedings of IEEE International Conference On Image Processing

2002

New York, NY, USA

525 528

10.

Mao

Bushmitch

Narayanan

Panwar

S. S.

MRTP: a multiflow real-time transport protocol for ad hoc networks

IEEE Transactions on Multimedia 2006 8 2 356 369

2-s2.0-33645151441

10.1109/TMM.2005.864347

11.

Stann

Heidemann

Rmst: reliable data transport in sensor networks

Proceedings of the 1st IEEE International Workshop on Sensor Net Protocols and Applications

2003

12.

Wan

C. Y.

Campbell

A. T.

Krishnamurthy

PSFQ: a reliable transport protocol for wireless sensor networks

Proceedings of the 1st ACM International Workshop on Wireless Sensor Networks and Applications

2002

Atlanta, Ga, USA

1 11

2-s2.0-0036980191

13.

Gungor

V. C.

Akan

O. B.

Delay aware reliable transport in wireless sensor networks

International Journal of Communication Systems 2007 20 10 1155 1177

2-s2.0-34948822444

10.1002/dac.869

14.

Nikookar

Hashemi

Statistical modeling of signal amplitude fading of indoor radio propagation channels

Proceedings of International Conference on Universal Personal Communications

1993

Ottawa, Canada

84 88

15.

Konstantas

Herzog

Continuous monitoring of vital constants for mobile users: the MobiHealth approach

Proceedings of the 25th Annual International Conference of the IEEE Engineering in Medicine and Biology Society

September 2003

mex

3728 3731

2-s2.0-1542300261

16.

Sun

Auvim: a middleware for audio/video sensor network

Proceedings of the 1st International Workshop on Middleware for Sensor Networks

2006