Habitat and environmental monitoring represent a class of sensor network applications that benefit scientific communities and societies [3][4][5]. This technology ...
AN APPROACH FOR TRACKING WILDLIFE USING WIRELESS SENSOR NETWORKS Francine Lalooses*†, Hengky Susanto*, Chorng Hwa Chang* † The MITRE Corporation *Tufts University 202 Burlington Road 161 College Avenue Bedford, MA 01730 Medford, MA 02155 USA USA {Francine.Lalooses, Hengky.Susanto, Chorng.Chang}@tufts.edu ABSTRACT - The advancement of mobile sensor networks yields a variety of wireless sensor network applications. One typical application for wireless sensors networks is area monitoring in wildlife environments. Sensor networks are utilized in wildlife to track animals. A significant number of studies have been done in tracking moving targets with sensor networks. However, at this moment, we believe there is no study has been done on recovering method from failure of monitoring mobile target. In this paper, we propose a simple recovery strategy to locate the lost target when tracking failure occurs.
there are many movement tracking techniques for single targets, there are other challenges which are equally valuable. One of the most important challenges is the problem of what to do when a target is lost and what is the best way to find the lost target. Our paper will discuss how to recover and rediscover targets when failures occur during the habitat and wildlife monitoring operation. We also specify the necessary conditions for the recovery algorithm to work successfully and better understand the performance of our recovery algorithm.
KEYWORDS - Algorithms, Sensor Networks, Tracking, Monitoring
Tracking wildlife has been a hot topic for the past few years. There has even been talk [22] from lawmakers to force the government to track livestock nationwide. An identification system has been introduced to implant an RFID chip called VeriChip into the animals. This chip, injected by a syringe, is the size of a grain of rice.
1. INTRODUCTION Wireless sensor networks have recently come into prominence because of its potential in revolutionizing many aspects of our economy and day to day life from environmental monitoring, to the health care industry, manufacturing, etc. Sensors are often deployed in constrained environments, such as rainforests, mountains or construction sites, for monitoring or detecting particular events [1] [8] [9]. Habitat and environmental monitoring represent a class of sensor network applications that benefit scientific communities and societies [3][4][5]. This technology allows researchers and biologists to more closely investigate a given area by providing [7] local measurement, sampling, and detailed information that are otherwise difficult and expensive to obtain. Traditional approaches in obtaining this information often require a human presence at the monitored location, which may disrupt normal animal behavior. For example, [10] seabird colonies are extremely sensitive to human presence. In Maine, biologists have discovered that their daily 15 minute visits to the colony increased the egg and chick mortality rate by 20% in a given breeding year. If the disturbance is continuously repeated, the entire colony may abandon the breeding site altogether. Numerous studies have been done for tracking movements and the population of animals in their natural habitat. While
2. RELATED WORK
On Maine’s Great Duck Island (GDI) [2], biologists have placed sensor devices in ducks’ underground nests and on four inch stilts placed just outside of duck burrows for a nine month monitoring period. This deployed network of thirtytwo nodes continuously streams data onto the web. At Princeton University [6], researchers are investigating advances with wireless sensor networks and applying this technology to support wildlife tracking for biological research. This tracking system, called ZebraNet, can forward data to a mobile base station by using peer-to-peer networking techniques. Their design decisions include custom tracking collars worn by the animals. These collars weigh approximately 2.4 pounds and are enabled with global positioning systems (GPS). Many additional tracking techniques have been proposed [13] [14] [16] to track moving targets. Researchers have formulated an estimation of signals received from sensors which calculate time dependent measurements to represent the location and characteristics of the target. One approach is to adopt the classic Bayesian formulation, which computes the measurement and communication costs to minimize tracking failure. Another research study [18] tracks and investigates the
temporal relationship between objects in tracking multiple target environments, such as monitoring the interaction of numerous animals. It is also very useful to consider predatory activities, such as a wolf trying to attack its prey of vulnerable, innocent lambs. Researchers at RPI developed a distributed protocol for target tracking in sensor networks [15]. This algorithm organizes sensors in clusters and uses sensor triplet triangulation to predict the target’s present location. The target’s next location is predicted using a linear predictor based on the last two locations of the target.
selected group. For example, the current group predicts the target is heading north, but the target may make a drastic direction change and head northeast. Hence, prediction error occurs. Failure is detected when the group leader, who is currently monitoring the target about to enter another area, does not receive an acknowledgement from the next selected group. The newly selected group is where the target was predicted to go, so this results in a lost confirmation. 3.1. Lost target conditions The possibility of not being able to find a target could be a direct result of several conditions. These conditions may include network failure, prediction failure, multiple sensor interaction, or hardware malfunction. 3.1.1. Network failure Network failure occurs when packets from the cluster head currently monitoring the target fails to reach the next cluster head. This results in a lost acknowledgement.
Figure 1: Sensor triplet triangulation Figure 1 also brings up a notion of normal beam and high beam. These sensors are usually set to normal beam if the cluster is active. Most clusters are in hibernation mode when the cluster head does not sense the target in its cluster. High beam is only activated when the target is lost and the cluster radius of the predicted location does not sense the target. By having sensors operate with normal beam, the predictive mechanism works without always having to consume more energy. A relay message is sent from one cluster head to another. Each cluster head activates the appropriate sensors before the target arrives. As the target is leaving the original cluster head’s group, that cluster head alerts the next cluster by handing over a target descriptor to the next cluster head. This target descriptor consists of the target’s identity, the present and predicted locations, and the time stamp of when the target entered the first cluster. By having this information relayed, clusters are able to obtain facts about the animal before it even arrives in its cluster.
3. RECOVERY ALGORITHM In this section, we discuss our proposed failure recovery scheme in tracking a moving target. This failure can occur due to issues such as network failure, prediction error, or node failure. The basic protocol for tracking a moving target is to select and wake up a group of nodes based on the movement of the target. The current group of nodes hands off the current monitoring task to the next group of nodes. However, the next group might never wake up and see the target enter their monitored area due to network failure or the target may select to go towards the opposite area of the
3.1.2. Prediction failure Prediction failure of the next location where the animal is heading may lead to the inability to find the target. Predicting the target’s next location helps reduce the possibility of losing a target. However, if the prediction of the next location is incorrect, then the cluster that was expecting the animal sends out alerts of the missing target. For example, the animal may suddenly change its path resulting in an inaccurate predicted location. 3.1.3. Multiple sensors interaction Another possibility of not being able to find a target could be a direct result of the failure to recognize multiple targets at once. For example, a cluster might be monitoring two horses standing next to each other. The cluster nodes might mistakenly recognize the two horses as one entity. When the horses leave the monitored area and enter another cluster group, the horses decide to split up. This split results in two distinct directions and different cluster groups. The current sensor nodes realize there is more than one target, but do not notify the second cluster head. Hence, one of the two targets is not monitored.
3.1.4. Hardware malfunction Another condition resulting in a lost target is the sensor node hardware may malfunction or the battery is weakening. Someone must be available to perform routine maintenance checks on the deployed hardware to ensure the integrity of the system.
3.2. Hierarchical Clustering The missing target can be quickly recovered by performing a space decomposition search. The basic concept of this algorithm is to take advantage of hierarchy clustering, wake up all of the nodes in the area at once, and perform a single simultaneous search. The root will wake up its subordinates and the order will be propagated to leaf level. Those leaf nodes subsequently perform an instantaneous search and will propagate the finding result to their superior until it reaches the root level. The running time of this search algorithm is O (log n). However, waking up the entire system of nodes could be too expensive and energy consuming. We propose a recovery algorithm that takes advantage of hierarchical clustering without waking up the entire network. The algorithm establishes a search region and wakes up the relevant nodes. Simultaneously waking up the necessary nodes at one particular region reduces network traffic and avoids the unnecessary extra energy costs. Establishing a search area is much simpler if the node positions are known. GPS might be required to determine the precise position of each node. However, our proposed algorithm does not require prior knowledge of the node positions or GPS. Additionally, targets are not tagged as in ZebraNet [6]. Tagging the target actually removes the risk of prediction failure of the target’s next location and thus, the recovery procedure for lost targets will not be necessary. This approach is less opportune because it is not practical to tag every animal in wildlife and potentially disturb the life of the animal. Moreover, tagging individual animals defeats the purpose of habitat monitoring. Our proposed algorithm takes advantage of hierarchical clustering structures for more efficient utilization of resources, simpler routing protocol, and management [19]. Clustering also allows some nodes to play watchdog or managerial roles over other nodes. This manager node is called the cluster head (CH). Each CH joins another cluster and selects another CH. This process continues until it forms a tree with one root, as shown in figure 2. This approach is known as hierarchical clustering technique. CH CH
Master CH d
CH
First level CH d
Figure 2: Hierarchical clustering technique After the initial deployment, sensor nodes often are compelled to organize themselves into a group of clusters and select a leader for every cluster as it is shown in figure 3. Several protocols exist on how to form a cluster. [29] [30] proposed that a CH is selected based on their node ID. Currently there is no theoretical proof that guarantees the quality of clustering. Clustering is just a constant bound to
measure the effectiveness of reaching and discovering all nodes in the network. The author describes in [20] that taking turns being a CH for every periodic cycle could significantly reduce extra energy costs.
Figure 3: Clusters of sensors Figure 3 depicts a Voronoi diagram where each polygon has exactly one point, shown as black and white circles, and every point in the polygon is closer to its central point than to any other. [21] proposed an algorithm to determine the location and radius of the monitored region by a cluster of nodes. We modify the algorithm such that after the initial deployment process, every CH reports the radius of its monitored region to its immediate parent. The parent computes the radius of its region based on the information from its subordinates and it propagates its newly computed region radius until it reaches the root. For example, consider a large company with many remote sites. The manager reports their respected department size to their immediate supervisor, the director. All directors compute the size of their region based on information from their subordinates and report the size of their region based on information from their subordinates and report the radius of their area to a higher superior, the general manager. This process is repeated until it reaches the chief executive officer, the root of the hierarchy tree. 3.3. Popular place We apply a new approach to find the lost target, the wildlife animal, by searching where their likely destination would be and establish a perimeter surrounding that area. The CH contacts its peers to establish a perimeter and sends a request to its subordinates to find the popular place. Animals tend to visit places where food and water is available. Other factors which commonly dictate animals’ locations are those which provide ideal climate or some degree of safety from predators. This is where animals typically prefer to rest, nest, or reproduce. One might ask what characterizes a popular place. Consider tourists visiting Boston, MA. A popular place would be places like Faneuil Hall, Boston Common, Fenway Park, or the USS Constitution. Now that we have four of the most popular places, we can try to locate a tourist by first asking these four places. By asking the most popular places, the amount of energy in a sensor network is minimized by avoiding the energy-consuming operation of turning all sensors on.
2h
2d
the target was last seen. If the target cannot be found, a designated CH will report these results to the base station. Otherwise, it will hand off the monitoring task to the new group, which will attempt to find the target.
1h 4h
Figure 4: Tracking algorithm based on popular place The popular place convention can be further explained by computing the distance d that the target travels. Once d is determined, then d needs to be multiplied by 2 simply because there are two opposing possible directions and the target may select either one of these directions. Next, that 2d value is propagated from the CH, where the target was last seen, to its superior. The superior compares the value of d with the diameter of its region. If the diameter of its region is roughly equal to 2d, then it will wake up all of its subordinates and leaf nodes to perform a search event. Otherwise, the 2d information is propagated to the higher superior until the proper region is found. Figure 4 shows the 2d value with the black circle being the last position the animal was seen and the dotted circles being the next popular places with their hop counts denoted. 3.4. Broadcasting Once popular places are defined, each CH of popular places sends beacon packets and computes distance, or number of hops, to the CH where the target was last seen. Each beacon packet contains descriptions of the targets. After receiving the packet from CHs of the popular place, the CH where the target was last seen broadcast packets with the life span equal to the distance, or hops, from one of the popular places. The purpose of broadcast is to establish and wake every node in the search region. Packet life span is decreased by one in every forward event. Flooding based mechanisms in the hierarchy tree are described in [11] [12] and the authors claim performing search operations with flooding based techniques is energy efficient. In fact, the flooding technique in [12] is specifically designed to support querying information in sensor networks for bird habitat monitoring. The rules of broadcasting should be that a node only forwards packet if and only if the new packet sequence is greater than the previously received packet sequence. Also, the packet sequence should be decreased by one at each node. Once the broadcast operation is executed, the search region is established and all the leaf level nodes are awake. For some period of time, all leaf nodes perform a simultaneous search operation to find the lost target based on their target description. Subsequently, all leaf nodes will report their findings to their CH and then this information will be propagated to the CH responsible for the area where
3.5. The Recovery The recovery is when the search region fails to capture the target and the target’s actual position is outside the search region. These failure occurrences are caused by the lack of proper information in finding popular places, especially when nodes are recently deployed in the new environment. If the nodes are recently deployed, performing a recovery operation by finding the perimeters of where the lost target might go is not effective since there is not enough data to capture the places that have been previously visited. For example, during the first minute after node deployment, a tiger passes the monitored region and the sub-regions are at most visited once. Therefore, every place has equal potential in becoming the popular stopping place for the tiger and it is difficult to determine the popular place given very limited information. Our approach is to perform another search using the maximum search area to find the popular place. The suggested search region is velocity of the moving target v * time t * constant c. All nodes in the radius of the targeted area are activated and perform the search. The success of the search relies heavily on time t. Thus, finding the proper t is critical for a successful search. 2d X Region
Search Region
Figure 5: X is the actual position of the target Figure 5 illustrates the maximum search area 2d which is double the diameter of the initial search region, X as the actual location of the animal hiding, and the search region as the location the animal was last seen. 3.6. Necessary Conditions As of today, we understand there are different algorithms and approaches on how to monitor different animals or their habitats because different animals have different behavior and live in different environments. For this reason, we need to understand the performance of our recovery algorithm before discussing the energy saving impact of our approach. In this section, we will describe the conditions that must exist for the recovery algorithm to work successfully. In order for the recovery algorithm to be successful, the following conditions are necessary: a well-defined coverage area, where the sensors are placed and what kind of terrain
they are in, sensor limitations and their reliability, and how multiple sensors interact when trying to identify the lost target among a group of targets with sensors.
procedure would affect the battery conservation when the sensors are not reliable. A study will be conducted on whether the recovery procedure would effect the battery conservation when failures occur very often.
3.6.1. Coverage area 3.6.5. Multiple sensor interaction and identification It is known that some animals can run faster than other animals. Computing the accurate speed of the animal is very crucial for computing the size of the search area. Therefore, depending on the speed of the animal, we need to understand at what probability the sensor will make the wrong estimation of animal speed. Minor mistakes are tolerable for animals that do not run and typically move slowly. For example, computing the speed of a cow could be easily bounded by some number based probability because their movement is predictable and they cover less distance. However, monitoring horse movement would be a challenge because horses cover more distance than cows. Our next accomplishment is to understand how our approach performs with different groups or types of animals. 3.6.2. Sensor placement and terrain Different animals cover different distances when the animal is in motion, like the cow and the horse. For this reason, biologists would use different sensors to monitor certain types of animals. In addition, depending on the environment and the terrain of where the animal lives, sensors are deployed differently. Furthermore, some animals have more limitations than others. For example, a horse could jump over an obstacle better and more easily than cow. Consequently, we have to be very careful in finding the place an animal is likely to stop (to rest, eat, etc) because some areas that appear reachable from the sensor networks point of view might be impossible for the animal, or a destination area that appears to be a short distance could physically turn out to be a longer distance because the animal is not able to jump over the obstacles and has to find a longer route to reach its destination. We believe these aspects need to be considered and fully investigated on how they would affect the performance of our recovery approach. 3.6.3. Sensor limitation It is possible that the recovery procedure can be difficult to implement due to the limitations of the sensors themselves. For example, the target animal could become out of reach because the animal climbed a very tall tree, which is beyond the reach of the sensor itself. 3.6.4. Sensor reliability We also need to understand how much our recovery
We also need to understand how the recovery algorithm would perform when there are multiple animals in a given area. Depending on the current technology, we need to answer how to identify the target and whether the sensors find the desired target or a different target, which is similar to the desired target. For example, we are very interested in understanding how the recovery algorithm would perform when there are six horses. This scenario makes the identification of the lost target even more challenging.
4. THE QUANDARY OF RECOVERING FAILURE The implementation of the recovery algorithm is quite complex. We need to understand how this recovery algorithm would respond to the unpredictable, real life network traffic conditions. Since recovering the lost target is a real time system application and the accuracy of establishing the proper search region depends on the current state of the target, we must consider the circumstances when packets are not delivered properly and timely. struct targetDescription { int8_t sensorID; // The time when the failure occur int16_t timeLostDetected; int16_t timeLatest; // Time int8_t targetID; // current target’s velocity int16_t currentTargetVelocity; //history of target’s velocity int16_t* historyTargetVelocity; int16_t constantC; //preset by user int32_t targetState; int32_t* lastTargetLocation; // localization measurement history up time t int32_t* measurement; // sensor uses acoustic or infrared to detect target int32_t* amplitude; // sensor uses more than one instruments. int32_t* amplitude2; }targetInfo; Figure 6: Sample of target description The common types of sensors for tracking are acoustic amplitude sensors, direction of arrival (DOA) sensors, and radio frequency identification (RFID). The acoustic amplitude sensor node measures sound amplitude to estimate the distance to the target based on the sound
reduction. DOA uses beam-forming techniques to determine the direction of the where the sound originates. Similarly, RFID uses radio waves to beam and determine the position of the target similar to DOA. This technique would work best when a single target is being monitored, but not when multiple targets are present at the same location. Moreover, monitoring multiple targets requires more information. The information currently received from sensors would not be sufficient to identify the desired target. For this reason, we have to associate a plethora of distinct information and data with one desired target in order to avoid mistakes. We are currently using the standard structure data type from TinyOS to describe a single target as it is shown in figure 6. The target description would be included in a packet with other network information (such as sensor ID, routing information, etc) and transmitted into the network. From our experience, it turns out that processing this information requires a lot of computing power which consumes a lot of energy. In addition to that, sending such large amounts of information creates traffic overhead.
so on. The target’s behavior can vary depending on the condition of the environment. There are many factors that could drive the target to behave differently and it is very difficult to make such a prediction with very limited information. That is why another discipline, such as machine learning, could be quite valuable in learning the pattern behavior of a particular target. We are currently working on testing our algorithm and making any improvements as we see fit. We would like to discover whether there is a softer tradeoff between energy consumed and prediction accuracy. In addition, our major goal is to minimize energy consumption when searching for the target. By having most sensor nodes in the system in hibernation, the risk of losing the target may increase. Future work includes testing this on the field and seeing how multiple animals interact with the sensor network we have described.
ACKNOWLEDGEMENTS
Therefore, in order to minimize the traffic overhead and conserve more energy, we believe the size of the packet transmitted through the network must be minimized, meaning that the information sent has to be compressed even more. Therefore, we are currently conducting a study to design a structure that describes the lost target and the current stage of recovery efficiently. Furthermore, due to limited storage space in sensor networks, minimizing information means less data to be stored.
The authors would like to thank all those who participated in our wildlife tracking efforts. Special thanks to the MITRE Corporation (Approved for Public Release; Distribution Unlimited. Case Number 07-0512) for their support on this project. The author’s affiliation with The MITRE Corporation is provided for identification purposes only, and is not intended to convey or imply MITRE’s concurrence with, or support for, the positions, opinions or viewpoints expressed by the author.
Another challenge is determining the speed of the target. In [15], the author proposes a method of computing the target’s speed based on the detection time between sensors and the position of nodes. Another approach described in [16] shows how sensors should determine the underlying target state of its position and velocity based on the sensor measurements up to time t. A prediction-based tracking algorithm is described in [17] where the authors use the velocity estimation of the target to select which sensors to query. Other approaches have been developed using Kalman Filters, which assume Gaussian observation models and linear state dynamics, and Bayesian Filtering techniques. However, there is yet a reliable method which can provide accurate estimations of the target’s speed. Therefore, this quandary needs to be seriously considered in our research especially when trying to recuperate from sensors not being able to provide accurate data.
REFERENCES [1] Ganesan D., Cerpa A., Ye W., Yu Y., Zhao J., and Estrin D., “Networking Issues in Wireless Sensor Networks”, Journal of Parallel and Distributed Computing (JPDC), 2004. [2] Mainwaring A., Polastre J., Szewcyk R., Culler D., & Anderson J., “Wireless sensor networks for habitat monitoring”, Proc. of workshop on wireless sensor networks and applications, Atlanta, GA, 2002. [3] Cerpa A., Elson J., Hamilton M., & Zhao J., “Habitat monitoring: application driver for wireless communications technology”, Proc. SIGCOMM 2001, 2001.
5. CONCLUSION
[4] Szewczyk R., Polastre J., Mainwaring A., & Culler D., “Lesson from a sensor network expedition”, Proc. EWSN, 2004.
The key to a successful recovery operation is determining the appropriate time t to compute distance d = time t * velocity v. Parameter t depends on many conditions, such as the processor power, network traffic, target’s behavior, and
[5] Wang H., Estrin D., & Girod L., “Preprocessing in a tiered sensor network for habitat monitoring”, EURASIP JASP Special Issue on Sensor Networks 2003, 4, 2003, 392-401.
[6] Juang P., Oki H., Wang Y., Martonosi M., Peh L., Rubenstein D., “Energy-efficient computing for wildlife tracking: design tradeoffs and early experiences with zebranet”, SIGOPS Oper. Syst. Rev., 36(5), 2002, 96107. [7] Polastre J. R., Design and implementation of wireless sensor networks for habitat monitoring, Master’s thesis, University of California at Berkeley, 2003. [8] Elson J. & Estrin D., Sensor networks: A bridge to the physical world (Norwell, MA: Kluwer Academic Publishers, 2004). [9] Estrin D., Culler D., Pister K., & Sukhatme G., “Connecting the physical world with pervasive networks”, IEEE Pervasive Computing, 1(1), 2002, 5969. [10] Anderson J., Pilot survey of mid-coast Maine seabird colonies: An evaluation of techniques. Report to the State of Maine Department of Inland Fisheries and Wildlife, Bangor, ME, 1995. [11]
Baryshnikov Y., Coffman E., Jelenkovic P., Momcilovic P., & Rubenstein D., “Flood search under the California split rule”, Operations Research Letters, 32(3), 2004.
[12] Sadagopan N., Krishnamachari B., & Helmy A., “The ACQUIRE mechanism for efficient querying in sensor networks”, Proc. IEEE International Workshop on Sensor Network Protocols and Applications (SPNA), Anchorage, AK, 2003, 149-155. [13] Liu J.J., Liu J., Reich J., Cheung P., & Zhao F., “Distributed group management for track initiation and maintenance in target localization applications”, Proc. of 2nd International Workshop on Information Processing in Sensor Networks (IPSN'03), Palo Alto, CA , 2003. [14] Shin J., Guibas L., Zhao F., “A distributed algorithm for managing multi-target identities in wireless ad-hoc sensor networks”, Proc. of 2nd International Workshop on Information Processing in Sensor Networks (IPSN'03), Palo Alto, CA, 2003. [15] Yang H. & Sikdar B., “A protocol for tracking mobile targets using sensor networks”, Proc. IEEE International Workshop on Sensor Networks Protocols and Applications, Anchorage, AK, 2003, 71-81. [16] Zhao F., Liu J., Liu J., Guibas L., & Reich J., “Collaborative signal and information processing: An information directed approach”, Proc. of the IEEE, 91(8), 2003, 1199-1209.
[17] Brooks R., Griffin C., & Friedlander D., “Selforganized distributed sensor network entity tracking”, International J. of High Performance Computing Application, 16(3), 2002, 207-220. [18] Guibas L., “Sensing, tracking, and reasoning with relations”, IEEE Signal Processing Magazine, 19(2), 2002, 73-85. [19] Marti S., Giuli T. J., Lai K., & Baker M., “Mitigating routing misbehavior in mobile ad hoc networks”, Proc. 6th Annual International Conference on Mobile Computing and Networking (MobiCom 2000), Boston, MA, 2000, 255-265. [20]
Bandyopadhyay S. & Coyle E., “Minimizing communication cost in hierarchically clustered network of wireless sensors”, IEEE INFOCOM 2003, San Francisco, CA, 2003.
[21] Savarese C., Rabaey J. M., & Beutel J., “Locationing in distributed ad-hoc wireless sensor networks”, Proc..International Conference Acoustics, Speech, and Signal Processing (ICASSP '01). 4, Salt Lake City, UT, 2001, 2037 - 2040. [22] Gossett S., “’Spy chips' for nation's livestock?”, WorldNetDaily.com , February 28, 2004.
