Energy-Efficient Tracking of Continuous Objects in ... - Semantic Scholar
Energy-Efficient Tracking of Continuous Objects in Wireless Sensor Networks* Jung-Hwan Kim, Kee-Bum Kim, Chauhdary Sajjad Hussain, Min-Woo Cui, and Myong-Soon Park** Department of Computer Science and Engineering Korea University, Seoul, Korea {glorifiedjx,givme,sajjad,minwoo,myongsp}@ilab.korea.ac.kr
Abstract. The proliferation of research on target detection and tracking in wireless sensor networks has kindled development of tracking continuous objects such as fires, bio-chemical material diffusion. In this paper, we propose an energy-efficient algorithm that detects and monitors a moving event region by selecting only a subset of nodes near object boundaries. The paper also shows that we can effectively reduce report message size. It is verified with simulation results that overall size of the report message as well as the number of nodes that transmit the report message to the sink can be significantly reduced especially when the density of nodes deployed over the network field is high. Keywords: wireless sensor networks, object tracking, target tracking, boundary, edge, continuous objects, energy-efficient.
1 Introduction Large scale wireless sensor networks have been enabled by rapid technological advances in MEMS and wireless communication. They are used in a wide variety of monitoring applications, ranging from habitat/environmental monitoring to military surveillance. One of typical and famous research area in wireless sensor network is target tracking. There have been enormous research achievements on target tracking with sensor networks. The majority of them are to identify and track one or multiple number of small targets. Sensors collaboratively detect and track the objects through the emittance of energy, noise, light, or seismic waves of the objects. However, there have been relatively small research efforts on detection and tracking large phenomena or objects such as forest fires, mud flows, bio-chemical material diffusion and oil spills. The phenomena can span large geographic extents and with monitoring such continuous and randomly changeable objects, we will be able to prevent possible assaults in advance such as diffusion of hazardous gas or find safe routes in such a circumstance for instance. * **
This work was supported by the Second Brain Korea 21 Project. Corresponding author.
F.E. Sandnes et al. (Eds.): UIC 2008, LNCS 5061, pp. 323–337, 2008. © Springer-Verlag Berlin Heidelberg 2008
324
J.-H. Kim et al.
To estimate such type of objects, it usually requires inordinate amount of message exchanges to collaboratively estimate the objects’ movement and location information in real time. Furthermore, it is well known that the cost for communication between sensors is much higher than that for computation. Therefore, it is judicious to invent an efficient algorithm that minimizes the communication costs as much as possible in operating sensor networks in order to prolong the lifetime.
(a) The simplest approach
(c) COBOM in [2]
(b) boundary sensors in [1]
(d) TOCOB (our approach)
Fig. 1. Comparison of number of nodes that send report messages to sink
There would be various ways to monitor the continuous objects. The simplest way to collect data from sensors would be to let every node that is actually detecting an object transmit its reading status to the sink node. Figure 1 (a) illustrates the straightforward approach. All sensors inside a phenomenon report data to the sink. However, this approach will cause sensors inside the phenomenon to dissipate energy at breakneck pace. In [1], Xiang Ji et al. propose a mechanism that selects only a few nodes nearby object boundaries to save energy. Solid dots in Figure 1(b) depict the boundary sensors selected nearby the boundary of an object. It is clear that there would be much more energy saving compared to Figure 1 (a) since only those which are near the object boundary are chosen to send data to the sink. [1] also discusses about
Energy-Efficient Tracking of Continuous Objects
325
forming clusters among the boundary sensors. We will discuss about it in more detail in section 2. In [2], Cheng Zhong and Michael Worboys propose a new energyefficient boundary node selection algorithm described in Figure 1 (c). According to [2], although more boundary sensors are selected than previous research, actual nodes that report to the sink are representative nodes that are drawn red dots in Fig. 1. (c). In this paper, we present an energy-efficient algorithm that detects boundaries of moving phenomena so as to monitor their shapes and movement in wireless sensor networks. We will focus on reducing the number of boundary sensors that eventually results in lowering the number of representative sensors to reduce traffic to the sink node as well as communication between sensor nodes. Figure 1(d) illustrates our proposed representative node selection. In our proposed architecture, we assume that if a node detects a phenomenon in its local area and the sensing value is beyond the predetermined threshold, then the node is regarded to be inside the phenomenon. Moreover, although it is more realistic to consider three-dimensional space, we examine only two-dimensional space in our architecture for simplicity. The rest of this paper is organized as follows: Section 2 discusses about previous works and criticizes them. In section 3, we present definitions related to our proposed idea as well as assumptions required for our algorithm. Section 4 presents the proposed algorithm in detail and Section 5 analyzes simulations and finally in Section 6, we conclude our dissertation with pointing out directions of the future work.
2 Related Work There has been a lot of research on detecting and tracking single or multiple targets in sensor networks [7][8]. Since last few years, some researchers have begun to analyze continuous object detection and monitoring such as forest fires, mud flows, biochemical material diffusion and oil spills. In [4][9], authors analyze detecting some non-local events which are closely related to our topic. The main difference in the research is that they try to estimate not the boundary of continuously changing objects but non-local phenomena that is always static. Therefore, they do not examine the situation where the phenomena move randomly and unexpectedly in real time fashion. Furthermore, [4] may lead to massive quantity of energy consumption since all boundary sensors report data to the sink. Xiang Ji et al. [1] propose a dynamic cluster based algorithm that tracks the movement of continuous objects with monitoring boundary of the objects. In [1], when a sensor detects the emergence of any phenomena at current time, it immediately broadcasts a query message to its neighbors to ask for the neighbors’ readings and the neighbors reply with sending their current readings to the sensor. If any neighbor that has different detection status exists, which means if the sensor receives at least one different detection status from any neighbor, the sensor becomes a boundary sensor. In a word, only those sensors inside the phenomena that are nearby (i.e., in
326
J.-H. Kim et al.
communication range) the boundary of objects will be selected as boundary sensors. After boundary node selection, cluster formation process takes place. However, the cluster formation explanation in [1] is somewhat ambiguous and also we claim that the clustering formation itself is not a good approach when the goal of forming clusters is to save energy since the application we consider in this paper is tracking objects that randomly and unexpected diffuse and drift with gust in real time. If, for example, the objects move, expand or shrink fast, cluster reformation through all boundary sensors should then happen every moment that the object changes or moves. Further, if a cluster is wide and huge, then propagation through members and a cluster head can be attenuated rapidly which necessitate more energy. Above all, the main drawback of this paper appears where every boundary sensor is directly or indirectly involved in routing data to the sink. Since each boundary sensor will be either a member or a cluster head, the boundary sensor in the network should at least send data to the distance up to its cluster head, which means every boundary node should consume a certain amount of energy anyhow. In [2], COBOM, an energy-efficient algorithm for boundary detection and continuous monitoring is proposed. If any sensor’s detection status is changed, then the sensor broadcasts its reading and ID. A neighbor node that receives the reading and ID stores the received reading into its array (called BN-array). Any sensors that have different detection status in BN-array from itself become boundary nodes. Among those boundary sensors, a few representative nodes will be selected. The more number of different detection status a sensor has in its array, the more likely it becomes a representative node that eventually report all the gathered detection result data to the sink. This algorithm is energy-efficient in a way such that 1) only a few representative nodes, which actually send report to the sink are chosen and 2) by using the BN-array, the report message size is not increased considerably since each message contains all of its neighbors’ detection status information only rather than keeping the neighbors’ IDs also, which requires only few bits while the precision of boundary monitoring is guaranteed. Figure 2 illustrates the BN-array.
Fig. 2. BN-aray is shown where c is the start node. It is assumed that the sink knows the start node and so do the other neighbors.
In this paper, however, we claim that we can get less number of boundary nodes than [2], which will also lead to less number of representative nodes. As few
Energy-Efficient Tracking of Continuous Objects
327
representative nodes will be chosen, more energy saving will be achieved. We will discuss about this more in detail through section 4.1.2 and 4.1.3. Furthermore, we also argue that if we can truly get more precise prediction on the shape of the objects when representative nodes report all their neighbors’ detection status to the sink. We will discuss more details about this argument in section 4.2.
3 Preliminaries In this section, we make fairly general assumptions about the capabilities of sensor nodes and the framework of sensor networks. And further, we discuss definitions employed in our algorithm. 3.1 Assumptions y y y y y y
Sensor nodes and a sink are stationary. The nodes are homogeneous. I.e., every node has the same capability. Each node has the same communication range R. The sensor nodes are densely and arbitrarily deployed in the network. The sink knows all the nodes’ IDs and locations. We do not concern possible data loss or contention. I.e., all communications between sensors are error free. y Each sensor node knows its own location by possibly using the global positioning system (GPS) [5] or other techniques such as triangulation [6] or localization [3]. 3.2 Definitions Network Model: Our sensor network can be modeled by a graph G=(V, E) in 2-D plane, where vertices V = {v1, v2, …,vn} represent sensor nodes. An edge eij exists between two vertices (nodes) vi and vj when they are within each other’s communication range and E symbolizes the set of all edges in the network. Definition 1. Interior (IN) of a phenomenon We define the interior of a phenomenon (IN) to be the spatial region ℜ2 such that a sensor detects an event, that is, identify a higher value than a predefined threshold at present time (i.e., at time slot t) and thus it is supposed to be placed inside the region. Its reading function would have evaluated to 1(or true). Definition 2. Exterior (OUT) of a phenomenon The exterior (OUT) of the phenomenon can be analogously defined as definition 1. The spatial region ℜ2 where no phenomenon or the reading lower than the threshold discovered is called OUT. The reading function of the nodes in OUT would have evaluated to 0 (or false).
328
J.-H. Kim et al.
Definition 3. Neighbors (Nu) Let u and Nu represent a node and neighbors of u respectively. The neighborhoods Nu are those nodes that are within communication range R of u. Definition 4. Changed Value Nodes (CVNs) Knowing the assumption that all sensors in the network periodically activate and make local observations (sensing) for detecting target objects, we define Changed Value Nodes (CVNs), detecting emergence of the object at the current time slot t when they did not identify any phenomena at previous time(t-1) or vice versa. That is, for any sensor u, if dt-1 = 0 and dt =1 or dt-1 = 1 and dt = 0 the sensor u becomes a CVN, where dt-1 and dt denote detection result from its local area at previous and current sensing time and 0 and 1 represent the reading result, true and false respectively. Figure 3 shows the CVNs when the continuous object expands and shrinks. Definition 5. CompareOneZero message (COZ message) CVNs broadcast a COZ message to their neighbors. This message includes a CVN’s ID and detection status. Definition 6. Boundary Nodes (BNs) A boundary node u is a node that receives at least one COZ message with different detection status. For example, if u’s current detection status is 0(false) and it receives from its one of neighbors, say v, a COZ message that includes v’s reading 1(true), now u is a boundary node since u’s current detection status is different from v’s. Definition 7. Representative nodes (RNs) A representative node is a node that actually sends data to the sink. Only a few representative nodes will be selected among BNs to save energy.
(a) CVNs when expanding
(b) CVNs when shrinking Fig. 3. CVNs in definition 4
4 Tracking Continuous Objects with Boundary Detection In this section, we illustrate the TOCOB algorithm in detail. In section 4.1, we explain each step one by one carefully and in section 4.2 we will discuss about precision of expected boundary and BN-array in detail.
Energy-Efficient Tracking of Continuous Objects
329
(a) This diagram well describes which sensors become boundary nodes in TOCOB
(b) BNs selected in TOCOB when the object expands (left) and shrinks (right)
(c) BNs selected in COBOM. All nodes closely located in to the boundary will become BNs.
Fig. 4. CVNs and BNs when expanding and shrinking
330
J.-H. Kim et al.
4.1 TOCOB Algorithm 4.1.1 CVNs and COZ Message Exchange Step 1: Each sensor, say u, activates and makes local observations periodically. Step 2: CVNs appear when there are some changes of shape of phenomena in the network. When the detection status of the previous time slot t-1 and the present time slot t of u are different, i.e., when dt-1 = 0 and dt =1 or dt-1 = 1 and dt = 0, u becomes a CVN. Step 3: Each CVN broadcasts a CompareOneZero message (hereafter, COZ) to its neighbors. The COZ message includes its own ID and detection status. The more significant expansion, shrinking or change of the object occurs, the more number of sensors will be implicated in broadcasting the COZ message to its neighbors since more sensors’ readings will be changed 0(false) to 1(true) or vice versa. 4.1.2 Boundary Nodes Step 4: A sensor u may receive the COZ messages sent from its neighbor CVNs ( I.e., the CVNs are in u’s communication range) and compares with its own current detection status. If the detection result is the same, then u ignores the COZ messages. It just stays. If u receives any COZ message that includes different detection status from its own, now u is called a BN and it counts the number of receiving COZ messages during a specific given time and based on the number of COZ messages received, the sensor will set a different waiting time, which is a way to select few RNs among the BNs. The following step demonstrates more on representative node selection. Figure 4 is presented for clarification of the explanation for CVNs and BNs. Figure 4 (a) shows which sensors become BNs when the continuous object is expanding. As clearly illustrated, among the sensors which are adjacently located to the boundary of the object, only the nodes that are in OUT will get privileges to become BNs when the object expands. Conversely, only the nodes in IN will become BNs when the object shrinks according to our algorithm. This is the main difference between COBOM algorithm in [2] and our proposed algorithm. As depicted in Figure 4 (b) and (c), COBOM algorithm selects all the nodes that are proximately situated to the boundary of objects as BNs since the nodes that have any different reading in its BN-array from that of itself all become BNs whereas in our proposed architecture, only those nodes that receive different reading from its neighbors can become BNs. Therefore, the number of BNs will be surely different no matter how or how much the object changes. 4.1.3 Representative Node Selection and Back-Off Time Step5: It would be still energy inefficient if all the BNs are involved in sending their data to the sink. Since our main objective in this paper is to construct an energy efficient algorithm that monitors changeable continuous objects, we try to choose as few RNs as possible that actually transmit report to the sink while tolerable accuracy is reserved. An effective method to determine the RNs without excessive message exchanges would be to figure in the number of received COZ messages. Based on the number of received COZ messages that includes different detection status from its
Energy-Efficient Tracking of Continuous Objects
331
own, each BN sets a waiting time. The higher number of the COZ messages a BN receives, the shorter the random back-off time it will be set such that BNs more proximately situated to the boundary of an object get higher probability to become RNs. The back-off timer is set as the following equation (1): ⎧ ⎪ ⎪ ⎛ ⎪ W W W ⎞ ⎪ D = COZ + U ⎜⎜ COZ − 1 − COZ ⎟ COZ total 〉 2 total total total ⎠, ⎪ ⎝ ⎪ ⎛ W W ⎞ ⎪ − ⎜ COZ − 1 COZ ⎟ W ⎪ total total ⎜ ⎟ COZ total = 2 = + D U ⎨ COZ total 2 ⎜ ⎟ ⎪ ⎟ ⎜ ⎪ ⎠, ⎝ ⎪ ⎪ ⎛ W W ⎞ − ⎜ COZ + 1 COZ ⎪ ⎟ total total ⎪D = W ⎟ COZ total = 1 −U ⎜ COZ total 2 ⎜ ⎪ ⎟ ⎟ ⎜ ⎪ ⎠, ⎝ ⎪⎩
(1)
Where D denotes back-off time, W is the maximum waiting time, COZtotal implies total number of COZ messages received and U represents the uniform distribution in [0,W). By using this equation for setting the back-off time, there would be very low chance that two or multiple RNs in communication distance to one another will be set with the same back-off time. Step 6: The BNs that have shorter back-off time will wake up earlier and broadcast a message to suppress its neighbors to become RNs. The node that sent the broadcasting message becomes a RN and on behalf of the nodes around, only the RNs send report data to the sink. The data to the sink includes the RN’s own ID and an ID of a CVN with the most powerful signal strength, which is one of the CVNs that sent a COZ message to the RN in step 4. I.e., whenever the RN receives a new COZ message in a given time, it compares signal strength of the previous COZ message with that of the new COZ message and drops one with lower signal strength. In this way, the RN might keep the closest CVN’s COZ message that will result in better estimation of the continuous object. 4.2 Discussion on the Precision of Expected Boundary and More on BN-Array In this part, we debate on the precision of boundary tracking through algorithms used in our proposed idea and in [2]. We claim that the expected shape of a typical object derived from our algorithm is as accurate as one generated by [2] while ours is more energy efficient. In [2], a RN sends all its neighbors’ detection status to the sink and in ours, each RN sends only one neighbor’s ID. Figure 5 compares expected shapes that can be possibly resulted from [2] and our proposed idea. We define a boundary point as a virtual point that is on the half of a RN and one of nodes in its communication range (with strongest signal strength. i.e., the one explained in step6) and also let an expected boundary be a connection between all the expected boundary points. In this paper, we adopt these two new definitions, boundary points and the expected boundary to measure and compare the precision of the shape of an object. As shown in Figure 5, knowing more boundary sensors may not guarantee that we can
332
J.-H. Kim et al.
accurately localize the boundary of an object or get a very likely shape as a real boundary since the sensors are arbitrarily deployed, i.e., not arranged in a horizontal line, and therefore what we can expect is in somewhere between the two nodes, there exists a authentic boundary of the object but we cannot measure or know exactly where the boundary of the object resides in. Certainly, if the more number of sensors deployed in the whole area of network, the more precise localization of the object will be achieved since the distances between nodes in the network will be shorter, which leads to better localization. Nevertheless, it still does not mean that awareness of all neighbors from RNs may create a better shape.
(a) Expected boundary line when a RN knows all nodes that have different status (left) Expected boundary line when a RN knows only one node that has different status (right)
(b) Expected boundary line in COBOM (left) and TOCOB (right)
Fig. 5. Expected shapes
Secondly, we claim that the report data to the sink in our algorithm can be lighter than or equal to [2] when we assume that each node’s id is not more than 1 byte and node density is high. There are numerous research that achieve id assignments with few bits in wireless sensor networks [10][11]. In our architecture, as mentioned before, each RN transmits its own ID and reading and a neighbor’s id whereas in [2], the report message consists of a RN’s ID, its reading and BN-array. The size of the BNarray can vary, especially when the sensors are densely deployed, it will be large as the number of neighbors for each sensor is high.The report message size estimation is performed in section 5.2.
Energy-Efficient Tracking of Continuous Objects
333
5 Performance Evaluations In this section, we evaluate the performance of TOCOB based on simulation results. We developed a simulator using Java to evaluate and compare the performance of Dynamic Structure [1], the COBOM algorithm [2] and our proposed idea. Unlike COBOM algorithm, we do not consider static objects in this paper due to the limitation of page but we fully focus on moving objects only. Albeit our simulations do not concern possible data loss or contention between nodes and how to route data to the sink, the simulations will experientially assure that TOCOB is more energy-efficient than previous works due to less number of RNs. Each simulation is run 1000 times and we assume the node ID is 1byte. Table 1. Simulation Parameters Parameters
Values
field size (m) number of nodes communication range (m) object radius increase (m/time slot) total time slots sensing and reporting periodicity (time slot)
100 x100 500(sparse setting) or 1000 (dense setting) 6 0.5 60 3
5.1 Simulation Model In our simulation, as the number of BNs and RNs will be surely dependent on density of nodes deployed, we will vary the number of nodes while fixing the field size. Sensor nodes are distributed over a 100 x 100 m2 field. In each experiment, 500 or 1000 sensor nodes are deployed arbitrarily in the field to simulate a sparse or dense setting. The communication range of each node is also fixed to 6 m. We simulate a continuous object in the square area with a circle. The circle is initially centered at (50,50) and it continually expands during 60 time slots. In each time slot, its radius increases by 0.5m and all sensors in the network activate and make local observations every 3 time slots. The simulation parameters for our proposed algorithm are given in table 1. 5.2 Simulation Results 5.2.1 Less Number of Boundary Nodes and Representative Nodes In this part, we compare different number of BNs and RNs selected during simulation where two different densities of nodes are set in fixed size of region. Figure 6 compares numbers of BNs chosen that are directly or indirectly involved in transmitting data to the sink. According to the Figure 6 (a) and (b), it is obvious that COBOM [2] generates many more number of BNs than Dynamic Structure [1] and our proposed algorithm (TOCOB) since any node that has different detection status from its neighbors becomes BNs (i.e., the nodes near the boundary of an object in IN and OUT) whereas in Dynamic Structure and TOCOB, BNs are those near the boundary of the object in either IN or OUT. TOCOB produces slightly more number of BNs than Dynamic Structure because, in case that an object expands (not shrinking or merely changes), nodes that are outside will be determined as BNs in our approach
334
J.-H. Kim et al.
whereas for Dynamic Structure, no matter the object shrinks or expands, it will get BNs from inside the object. However, from the fact that 1) each CVN in Dynamic Structure broadcasts a request first and nodes that receives the request messages also need to reply on the request and 2) All BNs selected are eventually necessitated to be participated in transmitting data to the sink, it is unreasonable to allege that less number of BNs in Dynamic Structure saves more energy than in TOCOB.
(a) Sparse setting (N=500)
(b) Dense setting (N=1000)
Fig. 6. Comparing the number of Boundary Nodes selected
(a) Sparse setting (N=500)
(b) Dense setting (N=1000)
Fig. 7. Comparing the number of Representative Nodes selected
In Figure 7, we compare the number of RNs selected in COBOM and TOCOB. Since Dynamic Structure does not select RNs, we do not take it into consideration. As apparently illustrated, difference of the number of RNs becomes larger as the object’s radius increases. As the object grows, the region that BNs and RNs should cover also gets larger and therefore, it is natural that RNs involved in detection increase in both algorithms. For COBOM algorithm, however, since the RNs in COBOM should cover both IN and OUT whereas only the nodes that are either IN or OUT can become the RNs in our proposed architecture and thus the number of RNs in TOCOB slowly
Energy-Efficient Tracking of Continuous Objects
335
increases compared to that of COBOM as the uncovered area (i.e., OUT when IN is chosen and vice versa) also becomes larger. This status quo becomes more apparent when density of nodes deployed is high. As seen in Figure 7(b), when the object’s radius becomes 30meters, the difference of the number of RNs becomes more than doubled once there was almost no difference when the object was small. In summary, the difference of the RNs get significantly large when the continuous object expands over the network where the density of nodes deployed is high. According to the Figure 6 and 7, hence, we can conclude that our algorithm produce less number of BNs and RNs and the difference of the numbers become more significant when the object expands where the high density of nodes is present. According to those facts, it is obvious that our algorithm, TOCOB can reduce more traffic between the RNs and the sink as well as the communication between the RNs and their neighbors. 5.2.2 Comparison of Average Report Data Size In this part, we compare average report data size in COBOM and TOCOB. We consider two density settings, i.e., when the number of nodes deployed is 1000 and 500. The left three columns in table 2 indicate average report data size where the number of nodes is assumed to be 1000 and the rest of the columns on the right hand side are when assuming 500 nodes are deployed. Meta data in density column implies the object’s radius and each figure in Proposed and COBOM columns signifies average number of total report data where each experiment is done 1000 times. Since, in our algorithm, each RN sends its own ID (1byte) , reading(1bit) and only one neighbor’s ID(1byte), we assume that the RNs in our architecture sends 3 bytes for each report message to the sink. For COBOM, we assume that each report message size can vary since it contains a RN’s ID (1byte) and reading (1bit) and a BN-array. Because the size of the BN-array depends on number of neighbors around, if a RN has more than 7 neighbors, it means the message size becomes the same as ours and if the neighbors are more than 15, the message size becomes 4bytes, which is bigger than the report message in our architecture and so on. Each figure in Proposed and COBOM columns is computed as follows: Total average data size = average # of RNs × corresponding report data size
(2)
For example, in COBOM, if the total average number of RNs is 10.367 when the object’s radius is 7.5meters and the average number of RNs that keeps 2, 3 and 4 bytes are 1.175, 7.815 and 1.377 respectively. Then the total average data size computation would be 2 (bytes) * 1.175 + 3 * 7.815 + 4* 1.377 that is 31.303 bytes whereas in our algorithm, it is 10.367 * 3(bytes), which is 31.101 bytes. After the calculation of the total average report data size for each period, we simply add up all the computed values and compare the total values. In this way, we might be able to induce which method is more effective to reduce the report data size. It is shown in table 2 that as the density of nodes in the network increases, our proposed architecture produces less report data size overall. However, it is also clearly demonstrated that when the density is low, our average report message is heavier than that of COBOM. Therefore, we can conclude that our idea for reducing the size of report data is as effective as COBOM when the density is low and becomes more effective as the density is high.
336
J.-H. Kim et al. Table 2. Average report data size Density N=1000 1.5m 3.0m 4.5m 6.0m 7.5m 9.0m 10.5m 12.0m 13.5m 15.0m 16.5m 18.0m 19.5m 21.0m 22.5m 24.0m 25.5m 27.0m 28.5m 30.0m
Proposed (3bytes) 2.172 11.421 21.27 25.155 31.101 37.479 43.764 49.887 56.43 63.567 69.873 77.157 82.908 89.343 96.897 102.375 110.214 116.673 122.394 128.889
COBOM (2,3,4 or 5bytes) 2.189 11.509 21.312 25.189 31.303 37.759 43.985 49.596 57.265 64.749 69.913 77.358 83.008 89.637 97.067 102.597 110.232 117.033 122.904 129.406
Density N=500 1.5m 3.0m 4.5m 6.0m 7.5m 9.0m 10.5m 12.0m 13.5m 15.0m 16.5m 18.0m 19.5m 21.0m 22.5m 24.0m 25.5m 27.0m 28.5m 30.0m
Proposed (3bytes) 0.948 3.798 7.974 10.653 13.515 16.332 19.194 22.143 24.843 27.534 30.879 33.552 36.582 39.042 41.745 44.502 47.709 50.568 53.286 56.649
COBOM (2,3,4 or 5bytes) 0.704 2.807 5.866 7.835 9.88 11.965 14.058 16.237 18.196 20.181 22.609 24.526 26.764 28.538 30.587 32.612 34.963 37.054 39.01 41.511
total
1338.969
1344.011
total
581.448
425.903
6 Conclusion This paper proposes the TOCOB algorithm for boundary detection and monitoring of continuously moving phenomena in wireless sensor networks. The algorithm selects only few representative nodes, which transmit report data to the sink among a small subset of boundary nodes and thus reduces traffic between the RNs and the sink node as well as the communication between the RNs and their neighbors. Furthermore, by sending only one of neighbor nodes’ ID, which is possibly the closest node among its neighbors that have different current reading, we have verified that the report message size can be smaller than the previous work, especially when the density of the deployed nodes is high. We also believe that the expected shape of the object can be as precise as the ones in previous works. In Section 5, we presented some simulations that support our allegation. The results show that our proposed idea greatly outperforms the previous works in terms of energy-efficiency, especially when the density of the nodes deployed is high, our algorithm is absolutely dominant. Future work will include verification of precision of expected boundary and invention of a new algorithm that considers residual energy of each node.
References 1. Ji, X., Zha, H., Metzner, J.J., Kesidis, G.: Dynamic Cluster Structure for Object Detection and Tracking in Wireless Ad-Hoc Sensor Networks. In: ICC conference (2004) 2. Zhong, C., Worboys, M.: Energy-Efficient Continuous Boundary Monitoring in Sensor Networks. Technical Report (2007), http://www.spatial.maine.edu/~czhong/boundary_monitoring.pdf
Energy-Efficient Tracking of Continuous Objects
337
3. Liao, P.-K., Chang, M.-K., Jay Kuo, C.-C.: Distributed Edge Detection with Composite Hypothesis Test in Wireless Sensor Networks. IEEE Communication Society Globecom (2004) 4. Chintalapudi, K., Govindan, R.: Localized Edge Detection In Sensor Fields. Ad-hoc Networks Journal, 273–291 (2003) 5. US Naval Observatory (USNO) GPS Operations (2001), http://tycho.usno.navy.mil/gps.html 6. Bulusu, N., Heidemann, J., Estrin, D.: GPS-less low cost outdoor location for very small devices. IEEE Pers. Commun. 7, 28–34 (2000) (Special Issue on Smart Space and Environments) 7. Singh, J., Madhow, U., Kumar, R., Suri, S., Cagley, R.: Tracking Multiple Targets Using Binary Proximity Sensors. In: IPSN (2007) 8. Jeong, J., Hwang, T., He, T., Du, D.: MCTA: Target Tracking Algorithm based on Minimal Contour in Wireless Sensor Networks. In: IEEE INFOCOM proceedings (2007) 9. Liu, J., Cheung, P., Guonidas, L., Zho, F.: A dual-space approach to tracking and sensor management in wireless sensor networks. In: ACM International Workshop on Wireless Sensor Networks and Applications Workshop, Atlanta (2002) 10. Kang, J.H., Park, M.-S.: Structure-based ID Assignment for Sensor Networks. IJCSNS International Journal of Computer Science and Network Security 6(7) (2006) 11. Oud-Amed-Vall, E., Blough, D.M., Heck, B.S., Riley, G.F.: Distributed Unique Global ID Assignment for Sensor Networks. In: Proceeding of IEEE International Conference on Mobile Ad-Hoc and Sensor Systems, Washington DC (2005)
Abstract. The proliferation of research on target detection and tracking in wireless sensor networks has kindled development of tracking continuous objects such as fires, bio-chemical material diffusion. In this paper, we propose an energy-efficient algorithm that detects and monitors a moving event region by selecting only a subset of nodes near object boundaries. The paper also shows that we can effectively reduce report message size. It is verified with simulation results that overall size of the report message as well as the number of nodes that transmit the report message to the sink can be significantly reduced especially when the density of nodes deployed over the network field is high. Keywords: wireless sensor networks, object tracking, target tracking, boundary, edge, continuous objects, energy-efficient.
1 Introduction Large scale wireless sensor networks have been enabled by rapid technological advances in MEMS and wireless communication. They are used in a wide variety of monitoring applications, ranging from habitat/environmental monitoring to military surveillance. One of typical and famous research area in wireless sensor network is target tracking. There have been enormous research achievements on target tracking with sensor networks. The majority of them are to identify and track one or multiple number of small targets. Sensors collaboratively detect and track the objects through the emittance of energy, noise, light, or seismic waves of the objects. However, there have been relatively small research efforts on detection and tracking large phenomena or objects such as forest fires, mud flows, bio-chemical material diffusion and oil spills. The phenomena can span large geographic extents and with monitoring such continuous and randomly changeable objects, we will be able to prevent possible assaults in advance such as diffusion of hazardous gas or find safe routes in such a circumstance for instance. * **
This work was supported by the Second Brain Korea 21 Project. Corresponding author.
F.E. Sandnes et al. (Eds.): UIC 2008, LNCS 5061, pp. 323–337, 2008. © Springer-Verlag Berlin Heidelberg 2008
324
J.-H. Kim et al.
To estimate such type of objects, it usually requires inordinate amount of message exchanges to collaboratively estimate the objects’ movement and location information in real time. Furthermore, it is well known that the cost for communication between sensors is much higher than that for computation. Therefore, it is judicious to invent an efficient algorithm that minimizes the communication costs as much as possible in operating sensor networks in order to prolong the lifetime.
(a) The simplest approach
(c) COBOM in [2]
(b) boundary sensors in [1]
(d) TOCOB (our approach)
Fig. 1. Comparison of number of nodes that send report messages to sink
There would be various ways to monitor the continuous objects. The simplest way to collect data from sensors would be to let every node that is actually detecting an object transmit its reading status to the sink node. Figure 1 (a) illustrates the straightforward approach. All sensors inside a phenomenon report data to the sink. However, this approach will cause sensors inside the phenomenon to dissipate energy at breakneck pace. In [1], Xiang Ji et al. propose a mechanism that selects only a few nodes nearby object boundaries to save energy. Solid dots in Figure 1(b) depict the boundary sensors selected nearby the boundary of an object. It is clear that there would be much more energy saving compared to Figure 1 (a) since only those which are near the object boundary are chosen to send data to the sink. [1] also discusses about
Energy-Efficient Tracking of Continuous Objects
325
forming clusters among the boundary sensors. We will discuss about it in more detail in section 2. In [2], Cheng Zhong and Michael Worboys propose a new energyefficient boundary node selection algorithm described in Figure 1 (c). According to [2], although more boundary sensors are selected than previous research, actual nodes that report to the sink are representative nodes that are drawn red dots in Fig. 1. (c). In this paper, we present an energy-efficient algorithm that detects boundaries of moving phenomena so as to monitor their shapes and movement in wireless sensor networks. We will focus on reducing the number of boundary sensors that eventually results in lowering the number of representative sensors to reduce traffic to the sink node as well as communication between sensor nodes. Figure 1(d) illustrates our proposed representative node selection. In our proposed architecture, we assume that if a node detects a phenomenon in its local area and the sensing value is beyond the predetermined threshold, then the node is regarded to be inside the phenomenon. Moreover, although it is more realistic to consider three-dimensional space, we examine only two-dimensional space in our architecture for simplicity. The rest of this paper is organized as follows: Section 2 discusses about previous works and criticizes them. In section 3, we present definitions related to our proposed idea as well as assumptions required for our algorithm. Section 4 presents the proposed algorithm in detail and Section 5 analyzes simulations and finally in Section 6, we conclude our dissertation with pointing out directions of the future work.
2 Related Work There has been a lot of research on detecting and tracking single or multiple targets in sensor networks [7][8]. Since last few years, some researchers have begun to analyze continuous object detection and monitoring such as forest fires, mud flows, biochemical material diffusion and oil spills. In [4][9], authors analyze detecting some non-local events which are closely related to our topic. The main difference in the research is that they try to estimate not the boundary of continuously changing objects but non-local phenomena that is always static. Therefore, they do not examine the situation where the phenomena move randomly and unexpectedly in real time fashion. Furthermore, [4] may lead to massive quantity of energy consumption since all boundary sensors report data to the sink. Xiang Ji et al. [1] propose a dynamic cluster based algorithm that tracks the movement of continuous objects with monitoring boundary of the objects. In [1], when a sensor detects the emergence of any phenomena at current time, it immediately broadcasts a query message to its neighbors to ask for the neighbors’ readings and the neighbors reply with sending their current readings to the sensor. If any neighbor that has different detection status exists, which means if the sensor receives at least one different detection status from any neighbor, the sensor becomes a boundary sensor. In a word, only those sensors inside the phenomena that are nearby (i.e., in
326
J.-H. Kim et al.
communication range) the boundary of objects will be selected as boundary sensors. After boundary node selection, cluster formation process takes place. However, the cluster formation explanation in [1] is somewhat ambiguous and also we claim that the clustering formation itself is not a good approach when the goal of forming clusters is to save energy since the application we consider in this paper is tracking objects that randomly and unexpected diffuse and drift with gust in real time. If, for example, the objects move, expand or shrink fast, cluster reformation through all boundary sensors should then happen every moment that the object changes or moves. Further, if a cluster is wide and huge, then propagation through members and a cluster head can be attenuated rapidly which necessitate more energy. Above all, the main drawback of this paper appears where every boundary sensor is directly or indirectly involved in routing data to the sink. Since each boundary sensor will be either a member or a cluster head, the boundary sensor in the network should at least send data to the distance up to its cluster head, which means every boundary node should consume a certain amount of energy anyhow. In [2], COBOM, an energy-efficient algorithm for boundary detection and continuous monitoring is proposed. If any sensor’s detection status is changed, then the sensor broadcasts its reading and ID. A neighbor node that receives the reading and ID stores the received reading into its array (called BN-array). Any sensors that have different detection status in BN-array from itself become boundary nodes. Among those boundary sensors, a few representative nodes will be selected. The more number of different detection status a sensor has in its array, the more likely it becomes a representative node that eventually report all the gathered detection result data to the sink. This algorithm is energy-efficient in a way such that 1) only a few representative nodes, which actually send report to the sink are chosen and 2) by using the BN-array, the report message size is not increased considerably since each message contains all of its neighbors’ detection status information only rather than keeping the neighbors’ IDs also, which requires only few bits while the precision of boundary monitoring is guaranteed. Figure 2 illustrates the BN-array.
Fig. 2. BN-aray is shown where c is the start node. It is assumed that the sink knows the start node and so do the other neighbors.
In this paper, however, we claim that we can get less number of boundary nodes than [2], which will also lead to less number of representative nodes. As few
Energy-Efficient Tracking of Continuous Objects
327
representative nodes will be chosen, more energy saving will be achieved. We will discuss about this more in detail through section 4.1.2 and 4.1.3. Furthermore, we also argue that if we can truly get more precise prediction on the shape of the objects when representative nodes report all their neighbors’ detection status to the sink. We will discuss more details about this argument in section 4.2.
3 Preliminaries In this section, we make fairly general assumptions about the capabilities of sensor nodes and the framework of sensor networks. And further, we discuss definitions employed in our algorithm. 3.1 Assumptions y y y y y y
Sensor nodes and a sink are stationary. The nodes are homogeneous. I.e., every node has the same capability. Each node has the same communication range R. The sensor nodes are densely and arbitrarily deployed in the network. The sink knows all the nodes’ IDs and locations. We do not concern possible data loss or contention. I.e., all communications between sensors are error free. y Each sensor node knows its own location by possibly using the global positioning system (GPS) [5] or other techniques such as triangulation [6] or localization [3]. 3.2 Definitions Network Model: Our sensor network can be modeled by a graph G=(V, E) in 2-D plane, where vertices V = {v1, v2, …,vn} represent sensor nodes. An edge eij exists between two vertices (nodes) vi and vj when they are within each other’s communication range and E symbolizes the set of all edges in the network. Definition 1. Interior (IN) of a phenomenon We define the interior of a phenomenon (IN) to be the spatial region ℜ2 such that a sensor detects an event, that is, identify a higher value than a predefined threshold at present time (i.e., at time slot t) and thus it is supposed to be placed inside the region. Its reading function would have evaluated to 1(or true). Definition 2. Exterior (OUT) of a phenomenon The exterior (OUT) of the phenomenon can be analogously defined as definition 1. The spatial region ℜ2 where no phenomenon or the reading lower than the threshold discovered is called OUT. The reading function of the nodes in OUT would have evaluated to 0 (or false).
328
J.-H. Kim et al.
Definition 3. Neighbors (Nu) Let u and Nu represent a node and neighbors of u respectively. The neighborhoods Nu are those nodes that are within communication range R of u. Definition 4. Changed Value Nodes (CVNs) Knowing the assumption that all sensors in the network periodically activate and make local observations (sensing) for detecting target objects, we define Changed Value Nodes (CVNs), detecting emergence of the object at the current time slot t when they did not identify any phenomena at previous time(t-1) or vice versa. That is, for any sensor u, if dt-1 = 0 and dt =1 or dt-1 = 1 and dt = 0 the sensor u becomes a CVN, where dt-1 and dt denote detection result from its local area at previous and current sensing time and 0 and 1 represent the reading result, true and false respectively. Figure 3 shows the CVNs when the continuous object expands and shrinks. Definition 5. CompareOneZero message (COZ message) CVNs broadcast a COZ message to their neighbors. This message includes a CVN’s ID and detection status. Definition 6. Boundary Nodes (BNs) A boundary node u is a node that receives at least one COZ message with different detection status. For example, if u’s current detection status is 0(false) and it receives from its one of neighbors, say v, a COZ message that includes v’s reading 1(true), now u is a boundary node since u’s current detection status is different from v’s. Definition 7. Representative nodes (RNs) A representative node is a node that actually sends data to the sink. Only a few representative nodes will be selected among BNs to save energy.
(a) CVNs when expanding
(b) CVNs when shrinking Fig. 3. CVNs in definition 4
4 Tracking Continuous Objects with Boundary Detection In this section, we illustrate the TOCOB algorithm in detail. In section 4.1, we explain each step one by one carefully and in section 4.2 we will discuss about precision of expected boundary and BN-array in detail.
Energy-Efficient Tracking of Continuous Objects
329
(a) This diagram well describes which sensors become boundary nodes in TOCOB
(b) BNs selected in TOCOB when the object expands (left) and shrinks (right)
(c) BNs selected in COBOM. All nodes closely located in to the boundary will become BNs.
Fig. 4. CVNs and BNs when expanding and shrinking
330
J.-H. Kim et al.
4.1 TOCOB Algorithm 4.1.1 CVNs and COZ Message Exchange Step 1: Each sensor, say u, activates and makes local observations periodically. Step 2: CVNs appear when there are some changes of shape of phenomena in the network. When the detection status of the previous time slot t-1 and the present time slot t of u are different, i.e., when dt-1 = 0 and dt =1 or dt-1 = 1 and dt = 0, u becomes a CVN. Step 3: Each CVN broadcasts a CompareOneZero message (hereafter, COZ) to its neighbors. The COZ message includes its own ID and detection status. The more significant expansion, shrinking or change of the object occurs, the more number of sensors will be implicated in broadcasting the COZ message to its neighbors since more sensors’ readings will be changed 0(false) to 1(true) or vice versa. 4.1.2 Boundary Nodes Step 4: A sensor u may receive the COZ messages sent from its neighbor CVNs ( I.e., the CVNs are in u’s communication range) and compares with its own current detection status. If the detection result is the same, then u ignores the COZ messages. It just stays. If u receives any COZ message that includes different detection status from its own, now u is called a BN and it counts the number of receiving COZ messages during a specific given time and based on the number of COZ messages received, the sensor will set a different waiting time, which is a way to select few RNs among the BNs. The following step demonstrates more on representative node selection. Figure 4 is presented for clarification of the explanation for CVNs and BNs. Figure 4 (a) shows which sensors become BNs when the continuous object is expanding. As clearly illustrated, among the sensors which are adjacently located to the boundary of the object, only the nodes that are in OUT will get privileges to become BNs when the object expands. Conversely, only the nodes in IN will become BNs when the object shrinks according to our algorithm. This is the main difference between COBOM algorithm in [2] and our proposed algorithm. As depicted in Figure 4 (b) and (c), COBOM algorithm selects all the nodes that are proximately situated to the boundary of objects as BNs since the nodes that have any different reading in its BN-array from that of itself all become BNs whereas in our proposed architecture, only those nodes that receive different reading from its neighbors can become BNs. Therefore, the number of BNs will be surely different no matter how or how much the object changes. 4.1.3 Representative Node Selection and Back-Off Time Step5: It would be still energy inefficient if all the BNs are involved in sending their data to the sink. Since our main objective in this paper is to construct an energy efficient algorithm that monitors changeable continuous objects, we try to choose as few RNs as possible that actually transmit report to the sink while tolerable accuracy is reserved. An effective method to determine the RNs without excessive message exchanges would be to figure in the number of received COZ messages. Based on the number of received COZ messages that includes different detection status from its
Energy-Efficient Tracking of Continuous Objects
331
own, each BN sets a waiting time. The higher number of the COZ messages a BN receives, the shorter the random back-off time it will be set such that BNs more proximately situated to the boundary of an object get higher probability to become RNs. The back-off timer is set as the following equation (1): ⎧ ⎪ ⎪ ⎛ ⎪ W W W ⎞ ⎪ D = COZ + U ⎜⎜ COZ − 1 − COZ ⎟ COZ total 〉 2 total total total ⎠, ⎪ ⎝ ⎪ ⎛ W W ⎞ ⎪ − ⎜ COZ − 1 COZ ⎟ W ⎪ total total ⎜ ⎟ COZ total = 2 = + D U ⎨ COZ total 2 ⎜ ⎟ ⎪ ⎟ ⎜ ⎪ ⎠, ⎝ ⎪ ⎪ ⎛ W W ⎞ − ⎜ COZ + 1 COZ ⎪ ⎟ total total ⎪D = W ⎟ COZ total = 1 −U ⎜ COZ total 2 ⎜ ⎪ ⎟ ⎟ ⎜ ⎪ ⎠, ⎝ ⎪⎩
(1)
Where D denotes back-off time, W is the maximum waiting time, COZtotal implies total number of COZ messages received and U represents the uniform distribution in [0,W). By using this equation for setting the back-off time, there would be very low chance that two or multiple RNs in communication distance to one another will be set with the same back-off time. Step 6: The BNs that have shorter back-off time will wake up earlier and broadcast a message to suppress its neighbors to become RNs. The node that sent the broadcasting message becomes a RN and on behalf of the nodes around, only the RNs send report data to the sink. The data to the sink includes the RN’s own ID and an ID of a CVN with the most powerful signal strength, which is one of the CVNs that sent a COZ message to the RN in step 4. I.e., whenever the RN receives a new COZ message in a given time, it compares signal strength of the previous COZ message with that of the new COZ message and drops one with lower signal strength. In this way, the RN might keep the closest CVN’s COZ message that will result in better estimation of the continuous object. 4.2 Discussion on the Precision of Expected Boundary and More on BN-Array In this part, we debate on the precision of boundary tracking through algorithms used in our proposed idea and in [2]. We claim that the expected shape of a typical object derived from our algorithm is as accurate as one generated by [2] while ours is more energy efficient. In [2], a RN sends all its neighbors’ detection status to the sink and in ours, each RN sends only one neighbor’s ID. Figure 5 compares expected shapes that can be possibly resulted from [2] and our proposed idea. We define a boundary point as a virtual point that is on the half of a RN and one of nodes in its communication range (with strongest signal strength. i.e., the one explained in step6) and also let an expected boundary be a connection between all the expected boundary points. In this paper, we adopt these two new definitions, boundary points and the expected boundary to measure and compare the precision of the shape of an object. As shown in Figure 5, knowing more boundary sensors may not guarantee that we can
332
J.-H. Kim et al.
accurately localize the boundary of an object or get a very likely shape as a real boundary since the sensors are arbitrarily deployed, i.e., not arranged in a horizontal line, and therefore what we can expect is in somewhere between the two nodes, there exists a authentic boundary of the object but we cannot measure or know exactly where the boundary of the object resides in. Certainly, if the more number of sensors deployed in the whole area of network, the more precise localization of the object will be achieved since the distances between nodes in the network will be shorter, which leads to better localization. Nevertheless, it still does not mean that awareness of all neighbors from RNs may create a better shape.
(a) Expected boundary line when a RN knows all nodes that have different status (left) Expected boundary line when a RN knows only one node that has different status (right)
(b) Expected boundary line in COBOM (left) and TOCOB (right)
Fig. 5. Expected shapes
Secondly, we claim that the report data to the sink in our algorithm can be lighter than or equal to [2] when we assume that each node’s id is not more than 1 byte and node density is high. There are numerous research that achieve id assignments with few bits in wireless sensor networks [10][11]. In our architecture, as mentioned before, each RN transmits its own ID and reading and a neighbor’s id whereas in [2], the report message consists of a RN’s ID, its reading and BN-array. The size of the BNarray can vary, especially when the sensors are densely deployed, it will be large as the number of neighbors for each sensor is high.The report message size estimation is performed in section 5.2.
Energy-Efficient Tracking of Continuous Objects
333
5 Performance Evaluations In this section, we evaluate the performance of TOCOB based on simulation results. We developed a simulator using Java to evaluate and compare the performance of Dynamic Structure [1], the COBOM algorithm [2] and our proposed idea. Unlike COBOM algorithm, we do not consider static objects in this paper due to the limitation of page but we fully focus on moving objects only. Albeit our simulations do not concern possible data loss or contention between nodes and how to route data to the sink, the simulations will experientially assure that TOCOB is more energy-efficient than previous works due to less number of RNs. Each simulation is run 1000 times and we assume the node ID is 1byte. Table 1. Simulation Parameters Parameters
Values
field size (m) number of nodes communication range (m) object radius increase (m/time slot) total time slots sensing and reporting periodicity (time slot)
100 x100 500(sparse setting) or 1000 (dense setting) 6 0.5 60 3
5.1 Simulation Model In our simulation, as the number of BNs and RNs will be surely dependent on density of nodes deployed, we will vary the number of nodes while fixing the field size. Sensor nodes are distributed over a 100 x 100 m2 field. In each experiment, 500 or 1000 sensor nodes are deployed arbitrarily in the field to simulate a sparse or dense setting. The communication range of each node is also fixed to 6 m. We simulate a continuous object in the square area with a circle. The circle is initially centered at (50,50) and it continually expands during 60 time slots. In each time slot, its radius increases by 0.5m and all sensors in the network activate and make local observations every 3 time slots. The simulation parameters for our proposed algorithm are given in table 1. 5.2 Simulation Results 5.2.1 Less Number of Boundary Nodes and Representative Nodes In this part, we compare different number of BNs and RNs selected during simulation where two different densities of nodes are set in fixed size of region. Figure 6 compares numbers of BNs chosen that are directly or indirectly involved in transmitting data to the sink. According to the Figure 6 (a) and (b), it is obvious that COBOM [2] generates many more number of BNs than Dynamic Structure [1] and our proposed algorithm (TOCOB) since any node that has different detection status from its neighbors becomes BNs (i.e., the nodes near the boundary of an object in IN and OUT) whereas in Dynamic Structure and TOCOB, BNs are those near the boundary of the object in either IN or OUT. TOCOB produces slightly more number of BNs than Dynamic Structure because, in case that an object expands (not shrinking or merely changes), nodes that are outside will be determined as BNs in our approach
334
J.-H. Kim et al.
whereas for Dynamic Structure, no matter the object shrinks or expands, it will get BNs from inside the object. However, from the fact that 1) each CVN in Dynamic Structure broadcasts a request first and nodes that receives the request messages also need to reply on the request and 2) All BNs selected are eventually necessitated to be participated in transmitting data to the sink, it is unreasonable to allege that less number of BNs in Dynamic Structure saves more energy than in TOCOB.
(a) Sparse setting (N=500)
(b) Dense setting (N=1000)
Fig. 6. Comparing the number of Boundary Nodes selected
(a) Sparse setting (N=500)
(b) Dense setting (N=1000)
Fig. 7. Comparing the number of Representative Nodes selected
In Figure 7, we compare the number of RNs selected in COBOM and TOCOB. Since Dynamic Structure does not select RNs, we do not take it into consideration. As apparently illustrated, difference of the number of RNs becomes larger as the object’s radius increases. As the object grows, the region that BNs and RNs should cover also gets larger and therefore, it is natural that RNs involved in detection increase in both algorithms. For COBOM algorithm, however, since the RNs in COBOM should cover both IN and OUT whereas only the nodes that are either IN or OUT can become the RNs in our proposed architecture and thus the number of RNs in TOCOB slowly
Energy-Efficient Tracking of Continuous Objects
335
increases compared to that of COBOM as the uncovered area (i.e., OUT when IN is chosen and vice versa) also becomes larger. This status quo becomes more apparent when density of nodes deployed is high. As seen in Figure 7(b), when the object’s radius becomes 30meters, the difference of the number of RNs becomes more than doubled once there was almost no difference when the object was small. In summary, the difference of the RNs get significantly large when the continuous object expands over the network where the density of nodes deployed is high. According to the Figure 6 and 7, hence, we can conclude that our algorithm produce less number of BNs and RNs and the difference of the numbers become more significant when the object expands where the high density of nodes is present. According to those facts, it is obvious that our algorithm, TOCOB can reduce more traffic between the RNs and the sink as well as the communication between the RNs and their neighbors. 5.2.2 Comparison of Average Report Data Size In this part, we compare average report data size in COBOM and TOCOB. We consider two density settings, i.e., when the number of nodes deployed is 1000 and 500. The left three columns in table 2 indicate average report data size where the number of nodes is assumed to be 1000 and the rest of the columns on the right hand side are when assuming 500 nodes are deployed. Meta data in density column implies the object’s radius and each figure in Proposed and COBOM columns signifies average number of total report data where each experiment is done 1000 times. Since, in our algorithm, each RN sends its own ID (1byte) , reading(1bit) and only one neighbor’s ID(1byte), we assume that the RNs in our architecture sends 3 bytes for each report message to the sink. For COBOM, we assume that each report message size can vary since it contains a RN’s ID (1byte) and reading (1bit) and a BN-array. Because the size of the BN-array depends on number of neighbors around, if a RN has more than 7 neighbors, it means the message size becomes the same as ours and if the neighbors are more than 15, the message size becomes 4bytes, which is bigger than the report message in our architecture and so on. Each figure in Proposed and COBOM columns is computed as follows: Total average data size = average # of RNs × corresponding report data size
(2)
For example, in COBOM, if the total average number of RNs is 10.367 when the object’s radius is 7.5meters and the average number of RNs that keeps 2, 3 and 4 bytes are 1.175, 7.815 and 1.377 respectively. Then the total average data size computation would be 2 (bytes) * 1.175 + 3 * 7.815 + 4* 1.377 that is 31.303 bytes whereas in our algorithm, it is 10.367 * 3(bytes), which is 31.101 bytes. After the calculation of the total average report data size for each period, we simply add up all the computed values and compare the total values. In this way, we might be able to induce which method is more effective to reduce the report data size. It is shown in table 2 that as the density of nodes in the network increases, our proposed architecture produces less report data size overall. However, it is also clearly demonstrated that when the density is low, our average report message is heavier than that of COBOM. Therefore, we can conclude that our idea for reducing the size of report data is as effective as COBOM when the density is low and becomes more effective as the density is high.
336
J.-H. Kim et al. Table 2. Average report data size Density N=1000 1.5m 3.0m 4.5m 6.0m 7.5m 9.0m 10.5m 12.0m 13.5m 15.0m 16.5m 18.0m 19.5m 21.0m 22.5m 24.0m 25.5m 27.0m 28.5m 30.0m
Proposed (3bytes) 2.172 11.421 21.27 25.155 31.101 37.479 43.764 49.887 56.43 63.567 69.873 77.157 82.908 89.343 96.897 102.375 110.214 116.673 122.394 128.889
COBOM (2,3,4 or 5bytes) 2.189 11.509 21.312 25.189 31.303 37.759 43.985 49.596 57.265 64.749 69.913 77.358 83.008 89.637 97.067 102.597 110.232 117.033 122.904 129.406
Density N=500 1.5m 3.0m 4.5m 6.0m 7.5m 9.0m 10.5m 12.0m 13.5m 15.0m 16.5m 18.0m 19.5m 21.0m 22.5m 24.0m 25.5m 27.0m 28.5m 30.0m
Proposed (3bytes) 0.948 3.798 7.974 10.653 13.515 16.332 19.194 22.143 24.843 27.534 30.879 33.552 36.582 39.042 41.745 44.502 47.709 50.568 53.286 56.649
COBOM (2,3,4 or 5bytes) 0.704 2.807 5.866 7.835 9.88 11.965 14.058 16.237 18.196 20.181 22.609 24.526 26.764 28.538 30.587 32.612 34.963 37.054 39.01 41.511
total
1338.969
1344.011
total
581.448
425.903
6 Conclusion This paper proposes the TOCOB algorithm for boundary detection and monitoring of continuously moving phenomena in wireless sensor networks. The algorithm selects only few representative nodes, which transmit report data to the sink among a small subset of boundary nodes and thus reduces traffic between the RNs and the sink node as well as the communication between the RNs and their neighbors. Furthermore, by sending only one of neighbor nodes’ ID, which is possibly the closest node among its neighbors that have different current reading, we have verified that the report message size can be smaller than the previous work, especially when the density of the deployed nodes is high. We also believe that the expected shape of the object can be as precise as the ones in previous works. In Section 5, we presented some simulations that support our allegation. The results show that our proposed idea greatly outperforms the previous works in terms of energy-efficiency, especially when the density of the nodes deployed is high, our algorithm is absolutely dominant. Future work will include verification of precision of expected boundary and invention of a new algorithm that considers residual energy of each node.
References 1. Ji, X., Zha, H., Metzner, J.J., Kesidis, G.: Dynamic Cluster Structure for Object Detection and Tracking in Wireless Ad-Hoc Sensor Networks. In: ICC conference (2004) 2. Zhong, C., Worboys, M.: Energy-Efficient Continuous Boundary Monitoring in Sensor Networks. Technical Report (2007), http://www.spatial.maine.edu/~czhong/boundary_monitoring.pdf
Energy-Efficient Tracking of Continuous Objects
337
3. Liao, P.-K., Chang, M.-K., Jay Kuo, C.-C.: Distributed Edge Detection with Composite Hypothesis Test in Wireless Sensor Networks. IEEE Communication Society Globecom (2004) 4. Chintalapudi, K., Govindan, R.: Localized Edge Detection In Sensor Fields. Ad-hoc Networks Journal, 273–291 (2003) 5. US Naval Observatory (USNO) GPS Operations (2001), http://tycho.usno.navy.mil/gps.html 6. Bulusu, N., Heidemann, J., Estrin, D.: GPS-less low cost outdoor location for very small devices. IEEE Pers. Commun. 7, 28–34 (2000) (Special Issue on Smart Space and Environments) 7. Singh, J., Madhow, U., Kumar, R., Suri, S., Cagley, R.: Tracking Multiple Targets Using Binary Proximity Sensors. In: IPSN (2007) 8. Jeong, J., Hwang, T., He, T., Du, D.: MCTA: Target Tracking Algorithm based on Minimal Contour in Wireless Sensor Networks. In: IEEE INFOCOM proceedings (2007) 9. Liu, J., Cheung, P., Guonidas, L., Zho, F.: A dual-space approach to tracking and sensor management in wireless sensor networks. In: ACM International Workshop on Wireless Sensor Networks and Applications Workshop, Atlanta (2002) 10. Kang, J.H., Park, M.-S.: Structure-based ID Assignment for Sensor Networks. IJCSNS International Journal of Computer Science and Network Security 6(7) (2006) 11. Oud-Amed-Vall, E., Blough, D.M., Heck, B.S., Riley, G.F.: Distributed Unique Global ID Assignment for Sensor Networks. In: Proceeding of IEEE International Conference on Mobile Ad-Hoc and Sensor Systems, Washington DC (2005)
