Thesis Committees: Dr. T. A. Yang, Chair Dr. L. Shih Dr. G. C. Collins APPLYING IMAGE PROCESSING TECHNIQUES TO SIMULATE A SELF-ORGANIZED SENSOR NETWORK FOR TRACKING OBJECTS Thesis Committees: Dr. T. A. Yang, Chair Dr. L. Shih Dr. G. C. Collins Date: 12/15/2005 UHCL - FALL 2005

CONTENTS INTRODUCTION RELATED WORK THESIS METHOD MODULE SIMULATION MATRICS SIMULATION RESULTS CONCLUSION LIMITATIONS AND FUTURE WORK REFERENCES

I. INTRODUCTION A wireless sensor network (WSN) is a network of wireless sensor nodes. Each node is a computer with attached sensors that can process, exchange sensing data, as well as communicate wirelessly among themselves to perform various tasks. Figure 1: A typical sensor node structure and some of commercial sensor nodes

I. INTRODUCTION Wireless sensor networks have many efficiency applications in military and civil, which may be classified into three classes: data collection, surveillance, and object tracking Figure 2: An example of wireless sensor network for data collection in agriculture Figure 3: Indoor Navigator project at MIT Figure 4: A scenario enemy tracking using sensor network

I. INTRODUCTION Current challenges: Small size of sensor node  Limit battery capacity and lower hardware performance. The network is formed by randomly throwing thousands or even millions of sensor nodes in an area  Overlapping sensing areas (redundancy). The network is usually installed in a large area with many physical effects, such as earthquake, explosive, or etc.  The topology of the network has to be changed frequently. Sensor nodes typically communicate via wireless links. Due to the limited bandwidth and interferences in the environment  Problems such as dropped packets and disconnected links.

So, what we expect … Thesis target A sensor network that Employs multi-hop communication with. Is reconfigurable. Self-organizes (including re-routing, redundancy reduction, and sensor deployment). Thesis target Develop a method, called OCO (Optimized Communication and Organization), which achieves the goals: Redundancy reduction. Efficient energy dissipation. Low computation overhead on the sensor nodes. Ensures maximum accuracy of object tracking

Figure 5: A scenario of direct communication II. RELATED WORK According to a survey of Chuang (National Tsing Hua Uni.) in 5/2005, there are three main approaches for target tracking in sensor network: tree-based, cluster-based, and prediction-based. Of course, the Direct Communication is also a kind of approach. Direct Communication The sensor module of all nodes is ON. Nodes send data directly to the base. Advantages Give the best accuracy. Disadvantages Unrealistic because the base has limited number of channels. Node energy is limited. Cannot applied to a large area. Suffer redundancy. Figure 5: A scenario of direct communication

Figure 6: Some methods to build hierarchy tree in the tree-based II. RELATED WORK Tree-based method Build a hierarchy tree by using mathematic model such as graph theory or Voronoi diagram. Nodes send data to the base through its ancestors. Advantages Can applied for a large area because of using multi-hop communication. Data could be aggregated at intermediate nodes Disadvantages Demand heavy calculation on the nodes for building the tree and routing. Suffer redundancy. Figure 6: Some methods to build hierarchy tree in the tree-based

Figure 7: A picture of cluster-based methods II. RELATED WORK Cluster-based method Build a hierarchy tree by using LEACH algorithm: Nodes randomly self-elect to become cluster heads. The cluster head invites its neighbors to join to the group. Re-elect cluster heads after a period of time for energy balancing. Nodes send data to the base through the cluster heads. Cluster heads communicate to the base directly. Advantages Simple. Disadvantages All nodes are supposed communicate directly to the base  Cannot applied to a large area and suffer the channel limitation of the base. Suffer redundancy. Figure 7: A picture of cluster-based methods

II. RELATED WORK Prediction-based method Based on tree-based or cluster-based method with added prediction models: The moving objects will stay at the current speed and direction for the next few seconds. The object’s speed and direction for next few seconds can be deduced from the average of the object’s movement history. Different weights can be assigned to the different stages based on the history. Advantages Give more efficient results. Disadvantages Keep disadvantages of the base method (tree or cluster). Prediction modes  unstable results. Demand heavy calculation for the nodes. Suffer redundancy.

III. THESIS METHOD OCO (Optimized Communication and Organization) includes 4 phases: position collection, processing, tracking, and maintenance. In the position collection phase, the base-station collects positions of all reachable nodes in the network. In the processing phase, it applies image processing techniques to clean up the redundant nodes, detect border nodes, and find the shortest path from each node to the base. In the tracking phase, the sensors in the network all work together to detect and track intrusion objects. The maintenance phase involves re-organizing the network when, for example, a change in the topology of the network occurs, or some of the sensor nodes die (i.e., running out of power).

Figure 8: Position Collection phase in simulating status III. THESIS METHOD Position collection phase: The base station sends a message to its neighbors to gather their IDs and positions, and at the same time advertise its own ID as the father ID of the neighbor nodes. The base’s neighbor nodes, after sending its ID and position to its father (the base), marks itself as recognized, and then performs the same actions as the base does by collecting IDs and positions from their neighbors, and advertising itself as the father node, and so on. When a node gets the position and ID information from its neighbor, it forwards the information to its father. Figure 8: Position Collection phase in simulating status

III. THESIS METHOD Processing phase: Consists of three steps: Clean up redundant nodes; Define the border nodes; Find the shortest path from each node to the base. Clean up redundant nodes: Firstly, we build a geographic image of the network by assign color value = 1 for all points that is covered by at least one sensor node. The rest points are assigned color value = 0. Initialize a list of nodes that are supposed to cover the whole network area, called Area_List. Assign Area_List = null. Add the base node to the Area_List. For all nodes in the area, if a node is not overlapping with any node in the Area_List, add it to the Area_List. The purpose of this step is to optimize the node distribution. For each point in the network area, if the point is not covered by any node in the Area_List, add the node that contains the point to the Area_List. Nodes that are not in the Area_list after the “for” loops in steps 3, 4, and 5 are redundant nodes and will be turned off. Figure 9: The results of cleaning up redundant nodes (the white in middle)

Figure 10: An example of border node detection III. THESIS METHOD Processing phase…. B. Define border nodes: Nodes that are positioned along the border of the network area are called border nodes. To define these nodes, firstly, apply the border detection algorithm to find out a list of points that traverse the border of the geographic image, called border points. Finally, find a minimum set of nodes in the that contain all the border points. The algorithm bellow is used to find the border of a monochrome image: For each pixel in the image, check if the color value =1. If true (meaning this pixel belongs to the object), scan all its neighbors to see if any of them having the color value = 0. If true, this pixel belongs to the border. Note: In this thesis, to optimize the border nodes, the image border is moved toward inside of the network area by a distance of sensing_radius/2 (half of the sensing radius). By doing so, the number of border nodes will decrease significantly without sacrificing any major characteristics of the network. This change may cause the accuracy of object detection to decrease a little bit, because the objects will be recognized a little bit late. The delay is acceptable though, in light of the gained benefit of reduced number of border nodes. Figure 10: An example of border node detection

Figure 11: Shortest path from node to the base III. THESIS METHOD Processing phase…. C. Find the shortest path from each node to the base: The following algorithm is used to find the shortest path (the least hops) from every node in the list of nodes (after the cleaning up) to the base: Work only with cleaned nodes. Assign father_ID = 0 for all nodes. Assign father_ID = the base’s ID for all neighbors of the base and add these node to a list, call processing list. For each node in the processing list. Consider all its neighbors. If the neighbor having father_ID = 0, assign the neighbor’s father_ID = the node’s ID. Add the neighbor to the processing list. Repeat step 4 until all nodes are assigned father_ID. Figure 11: Shortest path from node to the base

III. THESIS METHOD Processing phase…. After the processing phase, all nodes are assigned missions. The base broadcasts messages with node IDs to assign task for them. The summary is as follow: The redundant nodes are turned off totally. They just wake up after a long period (predefined) to receive commands from the base. If there is no command or the commands do not relate to them, they again switch to off totally. The border nodes have the sensor modules and the radio receiver modules are ON (called ACTIVE state). The rest of the nodes in the sensor network are called forwarding nodes, which have their sensor modules is OFF but the radio receiver modules is ON (called FORWARD state).

III. THESIS METHOD Tracking phase: Objects are assumed to have come from the outside. Normally, only the border nodes are ACTIVE. When a border node detects an object, it periodically sends its position information to the base by first forwarding the information to its father. We have two different types of sensor nodes, so , there are two tracking solution appropriated for each of them: Nodes are capable of sensing distinct multiple objects Nodes can accurately track each of the multiple objects; thus it only needs to activate its neighbors when a particular object is leaving its coverage area. The activated nodes will automatically return to the original status after an interval of sensing nothing. Nodes can not distinguish multiple objects (regular node) Without the capability to identify each of the multiple objects  Need to periodically activate its neighbors, assuming one or more of the multiple objects may leave its coverage area at any time.

III. THESIS METHOD Maintenance phase: The purpose is to reconfigure the network in the case of topology change. The following are example cases showing when such a topology change may be necessary: Exhausted nodes  When energy level of a node is below a threshold, it turns all its sons to SLEEP and sends a report to the base. When the base gets the report, it enters the processing phase to reconfigure the whole network, with dead nodes being removed and the network restructured. Damaged nodes  After a predefined interval of time, nodes require their child nodes to send their IDs to them (via a small size message). Children nodes that do not report to their parents are assumed to be damaged and will be reported to the base. Also, if a child node did not receive any asking from its father after the predefined interval of time (meaning the father is damaged), it will turn to SLEEP mode and wait for the command from the base. Re-positioned nodes (due to physical events, such as earthquake, explosion, etc)  When a node’s position changes, it will be considered as damaged by its parent (case b.). After being changed position, the node will do the following jobs: Automatically turns to SLEEP mode. Broadcast a message indicating that its position needs to be updated. Any node that has received the broadcast will forward the information to the base, which then updates the given node’s position.

IV. MODULE SIMULATION The tool that is used for simulation is OMNET++. It is selected because it enables to put simulated modules at any place. That is why we can simulate the random location feature of a sensor network as well as build moving objects. The are 3 basic components needed for the simulation: sensor node, intruder object, and sensor network. In addition, we need a module called manager to help the simulation such as: Put nodes to locations based on the result files of the processing phase. Making connection among nodes. Making connection between the objects and the nodes. Simulate the sensing behavior. Detect the first dead node in the network. Collect energy level of all nodes after the simulation.

Figure 12: Simulation structure of an sensor node IV. MODULE SIMULATION Sensor node simulation: Application MAC Layer 0 Coordinator Module Sensor Energy Radio Figure 12: Simulation structure of an sensor node The Layer 0 module represents the physical layer of a sensor node. MAC module represents pre-processing packet layers. The Application module represents the application layer. The Coordinator module is an interface to connect all modules together. It categorizes an incoming message in order to deliver it to the right module. The Sensor module represents the sensor board in a sensor node. The Radio module represents the radio board in a sensor node. The Energy module represents battery in a sensor node.

Figure 13: A sensor network structure with an intruder o under OMNET++ IV. MODULE SIMULATION Intrusion object simulation: Similar to the sensor node, an object has only two layers: the application layer on top of the physical layer. Simulating a sensor network with intruder objects: Figure 13: A sensor network structure with an intruder o under OMNET++

Selected method for simulation: IV. MODULE SIMULATION Selected method for simulation: In evaluating the performance of OCO, two methods are selected as comparisons: the Naive method (DC, Direct Communication) and the cluster-based method (LEACH). In DC, the sensor modules of all nodes are ON and the radio receiver modules are OFF. When having detected an intruder object, the node sends the information about intruders directly to the base node. In LEACH, the sensor modules of all nodes are ON. The radio receiver modules of the client (leaf) nodes are OFF, but the radio receiver modules of the cluster heads are ON.

V. METRICS There are four types of metrics that are considered when comparing the performance of the three selected methods: total energy consumption, accuracy, cost per detected point, and time before the first dead node. The total energy consumption is defined as the total energy that the network spends for tracking in a scenario. The accuracy is a percentage of the number of detected object positions of the method over the number of detected positions of the DC. The underlying assumption is that the DC method, due to its direct communication to the base, should exhibit the highest accuracy in detecting objects. The cost per detected point is an average number of energy units that are spent for a detected position. The time before first dead node is the time when the first node of the network runs out of energy. This matric is a significant indication of the sensor network’s ‘well-being’ or longevity.

Table 1: Summary of energy consumption for the simulation V. METRICS Energy consumption calculation Create/Receive a data message Create/Receive a signal message 100 µJ 3 µJ Send a data message (d<= 60m) Send a signal message (d<=60m) 820 µJ 26 µJ Send a message (d > 60m) 100 µJ + 0.1*d^2 Sensor board (full operation) 66 µJ/s Radio board (idle/receive mode) 100 µJ/s Table 1: Summary of energy consumption for the simulation

V. METRICS Accuracy calculation Cost per detected point calculation Standard number of detected point A sensor network with all nodes in the tracking mode (i.e. the sensor board is in full operation mode) is a useful base for comparison, because it provides the best possible quality of tracking. So, we consider the total number of detected points in this case as 100%, and call it the standard number of detected point. It also means the number of detected points in DC is 100%. The accuracy of each method is a percent ratio between the number of detected points of the method and the standard number of detected point. Cost per detected point calculation Cost per detected point is a ratio between the total energy dissipation and the total number of detected points of the method. Time before first dead node calculation The manager module periodically (every 0.1s) in the simulation network checks all nodes in the network to see if any node has run out the energy. If it finds any, the simulation time at that time will be recorded as time before the first dead node.

VI. SIMULATION RESULTS Simulation summary: The simulation environment is built as an area of 640x540. The number of nodes in the network is 200, 250, 300, 350, 400, 450, 500,550, 600, 650,700, 750, 800, 850, 900, 950, and 1000 with 2J (Joule) of energy for each node. The sensing radius of each node is 30m and the communication radius is 60m. - Randomly generate nodes. - Do collect ion position phase for OCO. - Generate text file results for OCO, DC, and LEACH. (OMNeT++) Processing program: - Generate text files for OCO after do the processing phase (C#) DC algorithm simulation program  Text files (OMNET++) OCO algorithm simulation program  Text files LEACH algorithm simulation program  Text files Evaluation programs (MATLAB) Figure 14: Summary of the simulation steps

Figure 15: Four intruder object moving paths VI. SIMULATION RESULTS Simulation summary… Intruder objects are supposed moving follow specific paths and come from outside of the network area. No data aggregation is allowed. The moving paths of objects are created by draw images. A MATLAB program reads the images and generates appropriate text files of positions of the path images. 4 paths 1, 2, 3, 4 are shown respectively as in Figure 15 Figure 15: Four intruder object moving paths

Figure 16: Simulation status of 3 methods in the case of 700 nodes VI. SIMULATION RESULTS Simulation results: Figure 16: Simulation status of 3 methods in the case of 700 nodes

Table 2: Summary of node deployment of the 3 methods VI. SIMULATION RESULTS Simulation results... The detail results are shown in the thesis document. This presentation just shows the main results of the simulation. Num of nodes DC LEACH-based OCO 200 178 (126 border nodes) 250 212 (136 border nodes) 300 230 (126 border nodes) 350 251 (131 border nodes) 400 269 (110 border nodes) 450 280 (101 border nodes) 500 285 (101 border nodes) 550 291 (86 border nodes) 600 287 (88 border nodes) 650 278 (83 border nodes) 700 299 (72 border nodes) 750 297 (74 border nodes) 800 294 (73 border nodes) 850 293 (73 border nodes) 900 279 (66 border nodes) 950 315 (61 border nodes) 1000 295 (60 border nodes) Table 2: Summary of node deployment of the 3 methods

VI. SIMULATION RESULTS Simulation results in the case of no intruder object Figure 17: Energy consumption (left) and time before first dead node (right) of the 3 methods in 9000 seconds (Note that if the time before first dead node is 1000, it means there are no dead node in the simulation) The energy consumption of OCO is the higher when the number of node is 200. When the number of nodes increases, the energy dissipation goes to a constant. That is because, when the number of node increases, the size of the border decreases (less gaps in the network coverage and more straight in the border line)

VI. SIMULATION RESULTS Simulation results in the case of 4 intruder objects moving in the 1st, 2nd, 3rd, and 4th paths, respectively. Figure 18: Energy consumption (upper left), time before first dead node (upper right), accuracy (lower left), and cost per detected points (lower right) of the 3 methods in the case of 4 intruder objects (Note that if the time before first dead node is 1000, it means there are no dead node in the simulation)

VII. CONCLUSION We have devised a method, OCO, for efficient target tracking in wireless sensor networks, and have evaluated its performance in various simulation scenarios against two other methods (DC and LEACH). Based on the evaluations, OCO appears to consume less energy than the other methods while achieving superior accuracy. The main strengths of OCO include its efficiency and easy maintenance, meaning that, when too many nodes have exhausted their energy, new nodes can be refilled to the tracking area and the OCO method will be able to dynamically build up a new network.

IIX. LIMITATIONS AND FUTURE WORK OCO needs to be implemented in a real sensor network to further verify its performance. The OCO did not yet support node-to-node communication which could be needed for other applications. The sensor network usually works in hostile environments, therefore, security features of OCO need to be seriously considered. Future Work: Implement OCO in a real sensor network. The node to node multi-hops communication is very important. We are considering adding this feature to OCO. We are also investigating security features for OCO.

IX. REFERENCES [1] Guo, Weihua, Zhaoyu Liu, and Guangbin Wu (2003). “An Energy-Balanced Transmission Scheme for Sensor Networks”. Dept. of Software and Information Systems - Univ. of North Carolina at Charlotte. Retrieved 9/8/2005 at http://www.cens.ucla.edu/sensys03/proceedings/p300-guo.pdf [2] S.C.Chuang (5/26/2005) . “Survey on Target Tracking in Wireless Sensor Networks ”. Dept. of Computer Science – National Tsing Hua University. Retrieved 11/8/2005 at http://mnet.cs.nthu.edu.tw/paper/934355tbl/050526--Survey%20on%20Target%20Tracking%20in%20wireless%20sensor%20newworks.pdf. [3] H. T. Kung and D. Vlah. (2003) “Efficient Location Tracking Using Sensor Networks.” WCNC, March 2003. Retrieved 11/7/2005 from http://www.eecs.harvard.edu/~htk/publication/2003-wcnc-kung-vlah.pdf [4] Wensheng Zhang and Guohong Cao (9/2004), “DCTC: Dynamic Convoy Tree-Based Collaboration for Target Tracking in Sensor Networks” IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 3, NO. 5, SEPTEMBER 2004. Retrieved 11/7/2005 from http://mcn.cse.psu.edu/paper/zhang/Twireless04.pdf [5] Wendi Rabiner Heinzelman, Anantha Chandrakasan, and Hari Balakrishnan (2000). “Energy-Efficient Communication Protocol for Wireless Microsensor Networks”. THE HAWAII INTERNATIONAL CONFERENCE ON SYSTEM SCIENCES, JANUARY 4-7, 2000, MAUI, HAWAII. Retrieved 6/20/05 from http://academic.csuohio.edu/yuc/mobile03/0403-heinzelman.pdf [6] Wei-Peng Chen, Jennifer C. Hou, and Lui Sha, (2004) “Dynamic Clustering for Acoustic Target Tracking in Wireless Sensor Networks” IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 3, NO. 3, JULY SEPTEMBER 2004. Retrieved 11/7/2005 from http://www-rtsl.cs.uiuc.edu/papers/icnp03_final.pdf [7] Yingqi Xu Winter, J. Wang-Chien Lee (2004). “Prediction-based strategies for energy saving in object tracking sensor networks” Mobile Data Management, 2004. Proceedings. 2004 IEEE International Conference. Retrieved 11/7/2005 from: doi.ieeecomputersociety.org/10.1109/MDM.2004.1263084 [8] Xu, Y., Winter, J., Lee, W.-C. “Dual predictionbased reporting for object tracking sensor networks” MOBIQUITOUS 2004. Retrieved 11/7/2005 from: doi.ieeecomputersociety.org/10.1109/MOBIQ.2004.1331722 [9] Andreas Savvides Mani Srivastava. “A SELF-CONFIGURING LOCATION DISCOVERY SYSTEM FOR SMART ENVIRONMENTS”. Retrieved 11/7/2005 from: http://www.eng.yale.edu/enalab/publications/cpcn_chapter.pdf [10] OMNET Official Website. “OMNET++ User Manual Version 3.2”. Retrieved from: http://www.omnetpp.org/doc/manual/usman.html

IX. REFERENCES [11] Sundeep Pattem, Sameera Poduri, and Bhaskar Krishnamachari (2003). “Energy-Quality Tradeoffs for Target Tracking in Wireless Sensor Networks”. DEPARTMENT OF ELECTRICAL ENGINEERING AND DEPARTMENT OF COMPUTER SCIENCE,UNIVERSITY OF SOUTHERN CALIFORNIA. Retrieved 7/10/05 from http://www-scf.usc.edu/~pattem/PattemKrishnamachari_Tracking.pdf [12] Shin Yoshizawa. “Voronoi Diagram”. Retrieved 11/7/2005 from: http://www.mpi-sb.mpg.de/~shin/Research/CCurve/node22.html [13] Rev B (2005). “MPR/ MIB User’s Manual”. DOCUMENT 7430-0021-06. Retrieved 7/10/05 from http://www.xbow.com/Support/Support_pdf_files/MPR-MIB_Series_Users_Manual.pdf [14] Matlab Image Processing Toolbox User's Guide. Retrieved 11/7/2005 from: http://www.mathworks.com/access/helpdesk/help/toolbox/images/intro5.htm [15] Die.Net Dictionary Online. Retrieved 11/7/2005 from: http://dict.die.net/joule/ [16] Unit of Measure conversion table. Retrieved 11/7/2005 from: http://www.themeter.net/conv5_e.htm [17] Jason Lester Hill (2003). “System Architecture for Wireless Sensor Networks”. Ph.D. thesis - UNIVERISY OF CALIFORNIA, BERKELEY. Retrieved 11/7/2005 from: http://www.motelab.org/papers/jhill-thesis.pdf [18] Cricket Project Group (2005). “The Cricket Indoor Location System” website, MIT. Retrieved 11/7/2005 from: http://nms.lcs.mit.edu/projects/cricket/ [19] VigilNet Project group (2004). “VigilNet: An Integrated Sensor Network System for Energy-Efficient Surveillance” website, Computer Science Dept. – University of Virginia. Retrieved 11/7/2005 from: http://www.cs.virginia.edu/~control/SOWN/index.html [20] Sensor Network Research Group at Louisiana State University (2/1/2005), “Simulating Wireless Sensor Networks with OMNeT++”. Retrieved 11/7/2005 from http://bit.csc.lsu.edu/sensor_web/final_papers/SensorSimulator-IEEE-Computers.pdf [21] Son Tran (2000). “Parallel Skeletonization for 3D binary image”, Son Tran Master Thesis. Retrieved 11/11/2005 from UHCL Library. [22] Chih-Yu Lin, Wen-Chih Peng, and Yu-Chee Tseng (2004). “Efficient In-Network Moving Object Tracking in Wireless Sensor Networks”. Department of Computer Science and Information Engineering - National Chiao Tung University. Retrieved 11/20/2005 from http://www.csie.nctu.edu.tw/~yctseng/papers.pub/sensor02-tracking-ieee-tmc.pdf

QUESTION?

ACKNOWLEDGMENTS I WOULD LIKE TO THANK DR. T. A. YANG, WHO SHOWED ENTHURIASTIC AND CONSISTENT SUPPORT AS THE CHAIR OF MY THESIS COMMITTEE. I APPRECIATE THE TIME HE SPENT IN OUR WEEKLY MEETINGS AND THE VALUABLE ADVICES HE HAD OFTEN GIVEN ME DURING THE THESIS JOURNEY. I ALSO LIKE TO THANK HIM FOR HAVING PROVIDED COMPUTER EQUIPMENTS FOR RUNNING THE SIMULATIONS. LAST BUT NOT THE LEAST, I THANK HIM FOR HIS PATIENCE OF SPENDING MANY HOURS IN HELPING TO REVISE THE GRAMMAR OF MY THESIS. I WOULD LIKE TO THANK DR. L. SHIH WHO ALWAYS SUPPORTS AND HELPS ME DURING THE TIME IN THE SCHOOL. I WOULD LIKE TO THANK DR. G. C. COLLINS WHO LED ME TO THE SENSOR NETWORK AREA. I WOULD LIKE TO THANK MY CLASSMATES. DURING THE PROCESS OF WORKING ON THE THESIS, I LEARN MANY THINGS FROM THEM AND FROM THEIR WORKS.

THANK YOU!