Wireless Sensor Networking for Intelligent Transportation Systems

Transcription

1 Wireless Sensor Networking for Intelligent Transportation Systems A DISSERTATION SUBMITTED TO THE FACULTY OF THE GRADUATE SCHOOL OF THE UNIVERSITY OF MINNESOTA BY Jaehoon Jeong IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FORTHEDEGREEOF Doctor of Philosophy Prof. David H.C. Du, Co-Advisor Prof. Tian He, Co-Advisor December, 29

2 c JaehoonJeong 29 ALL RIGHTS RESERVED

3 Acknowledgements I acknowledge the helps and contributions from many people during my Ph.D at the University of Minnesota. First of all, I sincerely appreciate the education and guidance from my advisors Professor David H.C. Du and Professor Tian He. They taught me how to research in the viewpoint of academia and encouraged me to be a better researcher. Also, I give my sincere thanks to my Ph.D committee for their good suggestion and sincere help for my research: Professor Zhi-Li Zhang, Professor Nicolai Krylov, and Professor Henry X. Liu. There are many other people that cheered and helped me. Especially, I sincerely appreciate the help and contributions of Minnesota Embedded Sensor System (MESS) group members, that is, Ziguo Zhong, Shuo Guo, Yu Gu, Ting Zhu, Yongle Cao, etc. Also, I thank my friend and colleague, Dongchul Park for his kindness and support during my Ph.D study. I would like to say thanks to my family. Especially, I sincerely thank my wife, Younkyeong Nam for her full support and deep love for the last five years. Also, I would like to acknowledge that my son, Younseong has given me his sincere love, has endured well the tough life in the US, and has been doing a good job at school. I thank my daughter, Younjee for giving me bright smile when I was tired. I sincerely love my wife, son and daughter. Also, I would like to give my sincere thanks to my father-in-law (Gichun Nam), mother-in-law (Younghee Lee), sister-inlaw (Hyekyeong) and brother-in-law (Soohyun) for their love and support for my family during my Ph.D study. I am missing my father (Kyuchang Jeong) and my mother (Soonduck Cho) in heaven. I would like to share my delight with them. They encouraged me to become a good scholar and a good person for our community. I would like to thank my sisters (Yeonho, Soonho, Kyeongho, Myeongho and Kyeongsook) and my brothers (Jaehyun, Sehyun and Sanghoon) for their love and support. Also, I would like to thank my best friends (Hyunjoo Jeong, Jeonghoon Shin, Woodong Park, Jinho Seo and Jongkwan Lee) for their support and encouragement. Last, I sincerely thank God for leading me and my family in the new voyage in the US. i

4 Dedication To my wife Younkyeong Nam, my son Younseong and my daughter Younjee. ii

5 Wireless Sensor Networking for Intelligent Transportation Systems by Jaehoon Jeong ABSTRACT This dissertation studies the Wireless Sensor Networking for Intelligent Transportation Systems, tailored and optimized for road networks. For military scenarios, since the road networks are used for main maneuver of military troops in cities or urban areas, they need to be protected for military operations. For civil engineering scenarios, the Intelligent Transportation Systems have been developed and been evolving to support the driving safety and transportation efficiency through the information computing and communications among transportation infrastructures and vehicles. Roadways are mainly used for the transportation of people and goods and also are nowadays equipped with intelligent devices, such as electronic tollgates and variable message signs for driving. In addition to this, vehicles are equipped with GPS-based navigation systems and emergency notification systems for the driving efficiency and safety. With this trend, Wireless Sensor Networks have been considered new parts for the Intelligent Transportation Systems and are being deployed into road networks in order to enhance further the driving safety and security. This dissertation studies the key technologies in the wireless sensor networking for the security and communications in the road networks as follows: (i) Localization for sensor location, (ii) Road Surveillance for vehicle monitoring, (iii) Data Forwarding for road sensing data delivery and (iv) Reverse Data Forwarding for road condition information sharing. In order to design the technologies to be tailored for road networks, this dissertation investigates the characteristics of road networks and takes advantage of the characteristics for the wireless networking. The first characteristic is the predictable vehicle mobility within the roadways. The second is the abstract representation of the layouts of the road networks into road maps. The third is the vehicular traffic statistics representing the vehicle density on the roadways and intersections. The fourth is the vehicle trajectory representing the future vehicle mobility along the roadways, guided by the GPS navigation systems. These four characteristics open a door of new research on wireless sensor networks. Therefore, using these road network characteristics, this dissertation designs and evaluates the wireless sensor networking technologies for road networks. iii

6 Contents Acknowledgements Dedication Abstract List of Tables List of Figures i ii iii viii ix 1 Introduction 1 2 APL: Autonomous Passive Localization Algorithm for Sparse Road Sensor Networks Introduction Problem Formulation Definitions Assumptions APL System Design System Architecture Step 2: Traffic Analysis for Road Segment Length Estimation Step 3: Prefiltering Algorithm for Virtual Graph Step 4: Graph Matching Step 5: Node Location Identification Practical Issues Sensor Time Synchronization Error iv

7 2.4.2 Vehicle Detection Missing and Duplicate Vehicle Detection Subgraph Matching for Non-isomorphism Performance Evaluation Performance Comparison between Road Segment Estimation Methods Performance Comparison among Prefiltering Types APL Operational Region Related Work Conclusion VISA: Virtual Scanning Algorithm for Road Network Surveillance Introduction Problem Formulation Virtual Scanning for Surveillance Analytical Network Lifetime Comparison Analytical Detection Time Comparison Configuring VISA for Better Delay and Longer Lifetime Virtual Scanning Algorithm Design Definitions and Assumptions VISA Scheduling on Road Network Handling of Sensing Holes Handling of Detection Failure Probability Handling of Time Synchronization Error Performance Evaluation System Behavior over Time Performance Comparison Achieving Shorter Delay and Longer Lifetime Simultaneously The Effect of Hole Handling Related Work Conclusion TBD: Trajectory-Based Data Forwarding for Light-Traffic Vehicular Ad-Hoc Networks Introduction Problem Formulation v

8 4.3 Link Delay Model TBD: E2E Delay Model and Protocol End-to-End Delay Model Forwarding Protocol Design Forwarding for Multiple Access Points Performance Evaluation Verification of Probability Model Forwarding Behavior Comparison The Impact of Vehicle Number N The Impact of Vehicle Speed µ v The Impact of Vehicle Speed Deviation σ v The Impact of Packet Time-To-Live TTL The Impact of Internet Access Point Number M Related Work Conclusion TSF: Trajectory-based Statistical Forwarding for Infrastructure-to-Vehicle Data Delivery Introduction Problem Formulation Assumptions About Stationary-Node-Assisted Forwarding Concept of Operation in TSF Target Point Selection for Data Delivery Delay Models Link Delay Model E2E Packet Delay Model Vehicle Delay Model TSF Protocol Forwarding Protocol Data Forwarding with Multiple APs Performance Evaluation Forwarding Behavior Comparison vi

9 5.6.2 The Impact of Vehicle Number N The Impact of Vehicle Speed µ v The Impact of Vehicle Speed Deviation σ v The Impact of Delivery Probability Threshold α The Impact of Internet Access Point Number M Related Work Conclusion Conclusion and Future Work 129 Bibliography 132 Appendix A. APL Appendix 14 A.1 Valid Edge Selection based on Minimum Spanning Tree A.2 Graph Matching Independent of Vehicle Mean Speed Appendix B. VISA Appendix 143 B.1 Average Detection Time in Virtual Scanning B.2 ADT Computation for Bounded Variable Vehicle Speed Appendix C. TBD Appendix 146 C.1 Average Forwarding Distance for Infinite Road Segment C.2 Average Forwarding Distance for Finite Road Segment Appendix D. TSF Appendix 149 D.1 Average Forwarding Distance for Infinite Road Segment D.2 Average Forwarding Distance for Finite Road Segment D.3 Variance of Forwarding Distance for Infinite Road Segment D.4 Variance of Forwarding Distance for Finite Road Segment D.5 Average Link Delay D.6 Variance of Link Delay vii

10 List of Tables 2.1 Outdoor Test Results APL Simulation Configuration Notation of Parameters for Analysis Performance Analysis for Three Approaches Performance Analysis for Three Approaches in Redundant Sensing VISA Simulation Configuration TBD Simulation Configuration Probabilities and Estimates for E2E Delay over Time Delay Average Estimation of VADD Delay Standard Deviation (STD) of VADD TSF Simulation Configuration viii

11 List of Figures 1.1 Intelligent Transportation Systems Road Network Characteristics Wireless Sensor Network deployed in Road Network APL System Architecture Time Difference on Detection (TDOD) Operation Estimation of Movement Time through TDOD Operation Road Network for Outdoor Test Comparison between Non-aggregation Method and Aggregation Method Procedure of Prefiltering for Obtaining Virtual Graph Procedure of Filtering out Path Estimates Construction of Reduced Virtual Subgraph Localization of Non-intersection Nodes Maximum Time Sync Error vs. Matrix Error Ratio according to Aggregation Type The Impact of Maximum Time Synchronization Error (ǫ max ) The Impact of Vehicle Speed Deviation (σ v ) The Impact of Vehicle Interarrival Time (1/λ) Operational Region for Vehicle Speed Deviation and Time Synchronization Error Operational Region for Detection Missing and Duplicate Detection Probabilities Road Network Surveillance Randomized Linear Deployment Sensor Sensing Sequence Scheduling Time Diagram for Node k Performance Comparison according to Working Time Performance Comparison under Different α Values ix

12 3.7 Road Network Graph G Virtual Scanning on Road Network for Working Schedule Establishment Virtual Scanning on Road Networks Augmentation of Road Network Graph for Holes Clustering for Sensing Hole Labeling Scheduling Time Diagram for Redundant Sensing Virtual Scanning for Redundant Sensing Handling of Time Synchronization Error The Comparison of Sleeping Time T sleep over Time The Impact of Working Time w The Impact of Sensor Density d The Impact of Time Synchronization Error ǫ max The Impact of Working Time w under Redundant Sensing The Impact of Sensor Density d under Redundant Sensing The Impact of Detection Failure Probability q The Impact of Silent Time α The Comparison of Hole Labeling Algorithms Packet Delivery Scenarios Forwarding Distance l f and Carry Distance l c Forwarding Distance (l f ) for Vehicle Arrivals Validation and Comparison of Analytical Models Link Delay Comparison for Model Validation Road Network Graph for Data Forwarding EDD Computation for Edge e 1, Average Forwarding Probability P 2,3 at Intersection EDD Computation for Vehicle Trajectory TBD Forwarding Protocol in VANET Road Network Graph with Two APs CDF Comparison for Delivery Delay The Impact of the Number of Vehicles under Infinite TTL The Impact of Vehicle Speed under Infinite TTL The Impact of Vehicle Speed Deviation under Infinite TTL x

13 4.16 The Impact of the Number of Vehicles under Finite TTL The Impact of Vehicle Speed under Finite TTL The Impact of Vehicle Speed Deviation under Finite TTL Performance Comparison between TBD and VADD according to AP Number Data Forwarding from AP to Target Point in Road Network Delay Histogram for VADD Data Forwarding for Infrastructure-to-Vehicle Data Delivery Packet Delay Distribution and Vehicle Delay Distribution Link Delay Modeling for Road Segment Packet Delay Model from AP to Target Point Vehicle Delay Model from Current Position to Target Point TSF Packet Format TSF Forwarding Protocol Reverse Path Forwarding for Vehicle Trajectory Data Forwarding with Multiple APs CDF Comparison for Delivery Delay The Impact of Vehicle Number N The Impact of Vehicle Speed µ v The Impact of Vehicle Speed Deviation σ v The Impact of Delivery Probability Threshold α The Impact of AP Number M B.1 Vehicle Detection Cases in Virtual Scanning B.2 The Impact of Vehicle Speed Deviation on Average Detection Time D.1 Forwarding Distance Statistics according to Vehicle Interarrival Time xi

14 Chapter 1 Introduction Intelligent Transportation Systems (ITS) are important systems for both military scenarios and civil engineering scenarios. For the military scenarios, since the road networks are used for main maneuver of military troops in cities or urban areas, they need to be protected for military operations. For the civil engineering scenarios, the Intelligent Transportation Systems have been developed and been evolving to support the driving safety and transportation efficiency through the information computing and communications among transportation infrastructures and vehicles [1]. Nowadays, roads are equipped with electronic tollgates and variable message signs for driving and also vehicles with GPS-based navigation systems and emergency notification systems. With this trend, Wireless Sensor Networks (WSN) can be deployed into road networks in order to enhance the driving safety and mobility further. The U.S. Department of Transportation (DOT) and many automobile companies (e.g., General Motors and Toyota) have recognized the promising features of ITS and recently started to apply the WSN technology to the ITS infrastructures [1, 2, 3]. As shown in Figure 1.1, the European Telecommunications Standards Institute (ETSI) has been making globally-applicable standards for Intelligent Transportation Systems [4], such as telematics and all types of communications between vehicles (e.g., vehicle-to-vehicle) and between vehicle and stationary location (e.g., vehicle-toinfrastructure or infrastructure-to-vehicle). In this figure, the wireless sensor networks can be integrated into ITS to monitor the road condition for driving safety (e.g., road construction sites or obstacles) and to announce such road condition to vehicles through vehicle-to-vehicle or infrastructure-to-vehicle communications. 1

15 2 Figure 1.1: Intelligent Transportation Systems Along with this trend, this dissertation studies the wireless sensor networking on road networks for both military applications and civil engineering applications. In order to design the wireless sensor networking technologies to be tailored and optimized for road networks, this dissertation investigates the characteristics of road networks and takes advantage of these characteristics for the wireless sensor networking on road networks. This dissertation observes the following four characteristics of the road networks, as shown in Figure 1.2: (i) Predictable Vehicle Mobility, (ii) Road Network Layout, (iii) Vehicular Traffic Statistics and (iv) Vehicle Trajectory. The first characteristic is the predictable vehicle mobility, as shown in Figure 1.2(a). Vehicles move along the roadways in the road networks with bounded speeds (i.e., speed limits). The second is the road network layout, as shown in Figure 1.2(b). The layouts of the road networks are available as road maps and these road maps can be represented as road network graphs where the vertices are intersections and the edges are road segments. The third is the vehicular traffic statistics, as shown in Figure 1.2(c). The vehicular traffic statistics (e.g., vehicle inter-arrival time) can be collected for the road segments or intersections due to the the mobility

16 3 p13 p14 p p Moving along the roadway 2. Speed Limit = 35MPH p 18 p 17 p p p 8 p 9 p 7 p 1 p 6 p 1 p 3 p 4 p 5 (a) Predictable Vehicle Mobility p 2 (b) Road Network Layout d1 = 1 Road Segment Inter-distance T v Tv d = Arrival Vehicle Density Inter-arrival time T T 1 t 1 t2 t 3 Time Vehicle Trajectory GPS Navigator (c) Vehicular Traffic Statistics (d) Vehicle Trajectory Figure 1.2: Road Network Characteristics within the road network layouts. The fourth is the vehicle trajectory, as shown in Figure 1.2(d). Vehicles tend to follow their vehicle trajectories representing the travel paths along the roadways, guided by the GPS navigation systems for the efficient travel. These four characteristics open a door of new research on wireless sensor networks. Therefore, this dissertation designs wireless sensor networking technologies for road networks on the basis of these four characteristics of road networks. To the best of our knowledge, the designs proposed in this dissertation are the first attempts to use these characteristics of the road networks. They are different from the legacy designs for wireless sensor networking based on the two-dimensional space. In this dissertation, the following four important technologies are addressed in depth for the security and communications in military applications and civil engineering applications, respectively: (i) Localization for sensor location, (ii) Road Surveillance for vehicle monitoring, (iii) Data Forwarding for road sensing data delivery and (iv) Reverse Data Forwarding for

17 4 road condition information sharing. First, for the wireless sensor localization, this dissertation introduces an Autonomous Passive Localization (APL), tailored for road networks. This sensor localization is the prerequisite function for wireless sensor networking. It is assumed that sensors are deployed sparsely (hundreds of meters apart) to save costs in road networks. This makes the existing localization solutions based on the ranging ineffective. To address this issue, our work is inspired by the fact that vehicles move along routes with a known map, which means the characteristics of the predictable vehicle mobility and the road network layout. Using binary vehicle-detection timestamps, distance estimates are obtained for any pair of sensors on roadways to construct a virtual graph composed of sensor identifications (i.e., vertices) and distance estimates (i.e., edges). This virtual graph (i.e., sensor network topology) is then matched with the topology of road map (i.e., road network topology), in order to identify where sensors are located in road networks. Second, for the road network surveillance, this dissertation proposes a Virtual Scanning Algorithm (VISA), tailored and optimized for road network surveillance. This road network surveillance is very important to the battles in urban areas or cities. The proposed design uniquely leverages upon the facts that (i) the movement of targets (e.g., vehicles) is confined within roadways and (ii) the road network maps are normally known, which are the predictable vehicle mobility and the road network layout, respectively. The detection of moving targets is guaranteed before the mobile targets reach designated protection points (such as temporary base camps), while maximizing the lifetime of the sensor network. The main idea in the proposed surveillance algorithm is to use virtual scan - waves of sensing activities scheduled for road network protection. Importantly, to the best of our knowledge, this is the first attempt to study how to guarantee target detection while sensor network deteriorates, using a novel sensing hole handling technique. Third, for the data forwarding of sensing data, this dissertation proposes a Trajectory-Based Data Forwarding (TBD), tailored for the data forwarding in light-traffic vehicular ad-hoc networks. This data forwarding is very important in the delivery of the road sensing data towards the infrastructure nodes, such as Internet access point or roadside unit. This is because these infrastructure nodes can play the role of hub for sharing the road condition information for the driving safety. State-of-the-art schemes have demonstrated the effectiveness of their data forwarding strategies by exploiting known vehicular traffic statistics (e.g., densities and speeds) in these vehicular networks. Even though these results are encouraging, further improvements

18 5 can be made by taking advantage of GPS-based navigation systems with the growing popularity. This dissertation presents the first attempt to investigate how to effectively utilize vehicles trajectory information in a privacy-preserving manner. In the proposed design, the trajectory information (i.e., vehicle trajectory) is combined with the vehicular traffic statistics to improve the performance of data forwarding in road networks. Fourth, for the reverse data forwarding for sharing road condition information, this dissertation proposes a data forwarding scheme called Trajectory-based Statistical Forwarding (TSF), tailored for the infrastructure-to-vehicle data delivery in vehicular networks. This reverse forwarding is very important to deliver such road condition information from infrastructure nodes to moving vehicles for the driving safety. TSF forwards packets to a multihop-away target point where the vehicle is expected to pass by, which is defined as rendezvous point of the packet and the destination vehicle. Such a target point is optimally selected to minimize the packet delivery delay while satisfying the required packet delivery probability. The optimality is analytically achieved by utilizing both distributions of the packet delay and the destination vehicle delay, which are obtained from the vehicular traffic density and the vehicle trajectory. To the best of our knowledge, the scheme proposed in this dissertation is the first attempt to investigate how to effectively utilize the destination vehicle s trajectory to compute such an optimal target point. The rest of this dissertation is organized as follows: Chapter 2 describes an autonomous passive localization scheme called APL for the localization of wireless sensors deployed on the road networks. In Chapter 3, a road network surveillance scheme called VISA is described for the mobile target detection in road networks. Chapter 4 describes a data forwarding scheme called TBD to deliver the road sensing data to the infrastructures, such as Internet access points on road networks. In Chapter 5, a reverse forwarding scheme called TSF is proposed for infrastructure-tovehicle data delivery in road networks for sharing the road condition information. Chapter 6 concludes this dissertation and presents future work of the wireless sensor networking for the future Intelligent Transportation Systems.

19 Chapter 2 APL: Autonomous Passive Localization Algorithm for Sparse Road Sensor Networks Road networks are one of important surveillance areas in military scenarios. In these road networks, sensors will be sparsely deployed (hundreds of meters apart) for the cost-effective deployment. This makes the existing localization solutions based on the ranging ineffective. To address this issue, this chapter proposes anautonomous Passive Localization (APL) 1 tailored and optimized for the localization of wireless sensors sparsely deployed into road networks. This scheme is based on the passive vehicular traffic measurement. Our work is inspired by the fact that vehicles move along routes with a known map. Using binary vehicle-detection timestamps, we can obtain distance estimates between any pair of sensors on roadways to construct a virtual graph composed of sensor identifications (i.e., vertices) and distance estimates (i.e., edges). The virtual graph is then matched with the topology of the road map, in order to identify where sensors are located on roadways. We evaluate our design outdoors on Minnesota roadways and show that our distance estimate method works well despite traffic noises. In addition, we show that our localization scheme is effective in a road network with eighteen intersections, where we found no location matching error, even with a maximum sensor time synchronization error of.7 sec and a vehicle speed deviation of 1 km/h. 1 Note that the preliminary version of APL was published at IEEE Infocom 28 [5]. 6

20 2.1 Introduction 7 Road networks are one of important infrastructures under surveillance in military operations. For the surveillance of these road networks, the localization of sensors is a prerequisite, providing target positions. In the military scenarios, it has been envisioned that for the fast and safe deployment, unmanned aerial vehicles drop a large number of wireless sensors into road networks around a target area. For the localization, many solutions have been proposed, using (i) precise range measurements (e.g., TOA [6], TDOA [7], and AOA [8]) or (ii) connectivity information (e.g., Centroid [9], APIT [1], SeRLoc [11], and Robust Quads [12]) between sensors for sensor localization. To cover a large area in road networks, sensors have to be sparsely deployed (hundreds of meters apart) to save costs. In this sparse deployment, since sensors cannot reach each other either through ranging devices (e.g., Ultrasound signals can only propagate 2 3 feet) or single-hop RF connectivity, the previous solutions become ineffective. To address this issue, we propose an Autonomous Passive Localization (APL) algorithm for extremely-sparse wireless sensor networks. This algorithm is built upon an observation: Military targets normally use roadways for maneuver; therefore, only the sensors near the road are actually useful for surveillance. The sensors away from the roadway can only be used for communication, since targets are out of their sensing range. In other words, their localization is unimportant. In such a scenario, the research question becomes how sensors on/near a road can identify their positions in a sparse deployment without any pair-wise ranging or connectivity information. The high-level idea of our solution is to use vehicles on roadways as natural events for localization. The solution would be trivial if all nodes are equipped with sophisticated vehicle identification sensors, because measuring the distance between two sensors by multiplying vehicles average speed by Time Difference on Detection (TDOD) between two sensors corresponding to the same vehicle is relatively easy. Obviously vehicle identification sensors would be costly in terms of hardware, energy and computation. Therefore, the challenging research question becomes how to obtain locations of the sensors, using only binary detection results without the vehicle identification capability in sensors. Our main idea is as follows. Through statistical analysis of vehicle-detection timestamps, we can obtain distance estimates between any pair of sensors on roadways to construct a virtual graph composed of sensor identifications (i.e., vertices) and distance estimates (i.e., edges).

21 8 This virtual graph is then matched with the topology of the known road map. Thus, this unique mapping allows us to identify where sensors are located on roadways. Specifically, our localization scheme consists of three phases: (i) the estimation of the distance between two arbitrary sensors in the same road segment; (ii) the construction of the connectivity of sensors on roadways; and (iii) the identification of sensor locations through matching the constructed connectivity of sensors with the graph model for the road map. Our key contributions in this chapter are as follows: A new architecture for autonomous passive localization using only the binary detection of vehicles in the road networks. Unlike previous approaches, APL is designed specially for sparse sensor networks where long distance ranging is difficult, if not impossible. A statistical method to estimate road-segment distance between two arbitrary sensors, based on the concept of Time Difference on Detection (TDOD). For the distance estimation, the TDOD operation uses the correlation between the timestamps of sensors geographically close to each other. A prefiltering algorithm for selecting only robust edge distance estimates between two arbitrary sensors in the same road segment. Unreliable path distance estimates are filtered out for better accuracy. A graph-matching algorithm for matching the sensor s identification with a position on the road map of the target area. The graph matching uses the isomorphic structure between the road network and the sensor network. Considerations on practical issues, such as time synchronization error, vehicle detection missing, duplicate vehicle detection, and the subgraph matching for the missing of sensor nodes or edge estimates. The rest of this chapter is organized as follows. Section 2.2 describes the problem formulation for our Autonomous Passive Localization (APL). Our APL system design is described in Section 2.3. In Section 2.4, we discuss practical issues that can affect our localization scheme in practice. Section 2.5 evaluates our APL algorithm in realistic settings. We summarize related work in Section 2.6 and conclude this chapter along with future work in Section 2.7.

22 2.2 Problem Formulation 9 We consider a network model where sensors are placed at both intersection points and nonintersection points on road networks. The objective is to localize wireless sensors deployed in road networks only with a road map and binary vehicle-detection timestamps taken by sensors as shown in Figure 2.1(a). Section lists definitions for APL and Section lists assumptions. b a c d Base Camp (a) Road Network with Wireless Sensors H v s 5 s 6 s 4 s 3 s 7 s 8 s 2 s 1 s 51 s 17 s 16 s 12 s 9 s s11 1 Road Segment Path s 13 s 15 s 14 (b) Virtual Topology of Wireless Sensors: H v = (V v, M v) s s 6 s 25 s 7 s 27 s G 49 s 48 8 s 47 v s s s 5 19 s 24 s s s 5 s 23 s s s 1 s 2 s3 s4 s s 29 s 1 s 2 s 3 11 s s s13 s 21 s 36 s 35 s 45 s s s s s 16 s 18 s s s 15 s 33 s s s 37 s 4 s s s Intersection Node Non-Intersection Node (c) Virtual Graph representing Sensor Network: G v = (V v, E v) Base Camp (d) Road Network only with Intersection Nodes of Virtual Graph G ~ v s 1 s 2 s 18 s 16 s 5 s3 s4 s 17 s 6 s s 1 11 s 12 s7 s8 s 15 s 9 s 13 s 14 (e) Reduced Virtual Subgraph consisting of Intersection Nodes of Virtual Graph: G v = (Ṽv, Ẽv) G r p 18 p 17 p 1 p 3 p 14 p p p 2 p 13 p p 9 8 p 7 p12 p11 p 4 p 1 p 6 p 5 (f) Real Graph corresponding to Road Map: G r = (V r, E r) Figure 2.1: Wireless Sensor Network deployed in Road Network Definitions We define eight terms as follows: Definition 2.2.1(Intersection Nodes) Sensors placed at an intersection and having more than two neighboring sensors (i.e., degree 3). In Figure 2.1(a), sensors a and c are intersection nodes.

23 1 Definition 2.2.2(Non-intersection Nodes) Sensors placed at a non-intersection and having one or two neighboring sensors. In Figure 2.1(a), sensors b and d are non-intersection nodes. Definition 2.2.3(Virtual Topology) Let Virtual Topology be H v = (V v,m v ), where V v = {s 1,s 2,...,s n }isasetofsensorsintheroadnetwork, and M v = [v ij ]isamatrixofpathlength v ij forsensors s i and s j. Figure2.1(b)showsavirtual topology ofsensors intheroadnetwork, shown in Figure 2.1(a). M v is a complete simple graph, since there is an edge between two arbitrary sensors. We define the edge of the virtual topology as virtual edge. In Figure 2.1(b), among the virtual edges, a solid thick line represents an edge estimate (i.e., road segment) between two sensors, which means that they are adjacent on the road network. The dotted thin line represents a path estimate between two sensors, which means that they are not adjacent on the road network. Definition 2.2.4(Virtual Graph) LetVirtualGraphbe G v = (V v,e v ),where V v = {s 1,s 2,...,s n } is a set of sensors in the road network, and E v = [v ij ] is a matrix of road segment length v ij between sensors s i and s j. Figure2.1(c) shows a virtual graph of the sensor network deployed on the road network shown in Figure 2.1(a), where the black node represents an intersection node and the gray node represents a non-intersection node. Definition 2.2.5(Reduced Virtual Subgraph) LetReducedVirtualSubgraphbe G v = (Ṽv,Ẽv), where Ṽv = {s 1,s 2,...,s m }isaset ofsensors placed only at intersections intheroadnetwork, and Ẽv = [v ij ]isamatrixofroadsegment length v ij betweenintersection nodes s i and s j. The reduced virtual subgraph G v is obtained by deleting non-intersection nodes and their edges from the virtual graph G v through the degree information in G v. Refer to Section For example, Figure 2.1(e) shows a reduced virtual subgraph consisting of only intersection nodes of virtual graph in Figure 2.1(c). Definition 2.2.6(Real Graph) LetRealGraphbe G r = (V r,e r ),where V r = {p 1,p 2,...,p n } is a set of intersections in the road network around the target area, and E r = [r ij ] is a matrix of road segment length r ij for intersections p i and p j. Real Graph can be obtained through map services, such as Google Earth and Yahoo Maps. Figure 2.1(f) shows a real graph corresponding to the road network whose intersection points have intersection sensor nodes, shown infigure2.1(d). Therealgraphis isomorphic tothereducedvirtual subgraphgraph G v shown in Figure 2.1(e) [13].

24 11 Definition 2.2.7(Shortest Path Matrix) Let Shortest Path Matrix for G = (V,E) be M such that M = [m ij ]isamatrixoftheshortestpathlengthbetweentwoarbitrarynodes iand j in G. M is computed from E by the All-Pairs Shortest Paths algorithm, such as the Floyd-Warshall algorithm [14]. We define M r as the shortest path matrix for the real graph G r = (V r,e r ), and define M v asthe shortest path matrix for the virtual graph G v = (V v,e v ). Definition 2.2.8(APL Server) A computer that performs the localization algorithm with binary vehicle-detection timestamps collected from the sensor network Assumptions The localization design of APL is based on the following assumptions: Sensors have simple sensing devices for binary vehicle detection without any costly ranging or GPS devices [15]. Each detection is a tuple (s i,t j ), consisting of a sensor ID s i and timestamp t j. There is an ad-hoc network or a delay tolerant network for wireless sensors to deliver vehicle-detection timestamps to the APL server. Sensors are time-synchronized at the millisecond level. This can be easily achieved because many state-of-the-art solutions [16, 17] can achieve microsecond level accuracy. The APL server has road map information for the target area under surveillance and can construct a real graph consisting of intersections in the road network. Vehicles pass through all road segments on the target road networks. The vehicle mean speed is close (but not identical) to the speed limit assigned to roadways. The standard deviation of vehicle speed is assumed to be a reasonable value, based on real road traffic statistics obtained from transportation research [18]. 2.3 APL System Design In this section, we explain our system architecture for autonomous passive localization, the estimation method to measure distance between two arbitrary sensors, the prefiltering algorithm

25 12 to convert a virtual topology into a virtual graph, the graph-matching algorithm to find a permutation matrix letting the reduced virtual subgraph and real graph be isomorphic, and sensor location identification using the found permutation matrix System Architecture 2 APL Server ( s, T ) Traffic Analysis 3 4 H v Prefiltering G v Graph Matching P G ~ P v Location Identification Vehicle Detection Timestamps Node Location Notification ( s, l) 1 ( i s i, T ) ( i s i, l ) 5 6 Sensor Node Repository t k Vehicle Detection s i Figure 2.2: APL System Architecture We use an asymmetric architecture for localization as in Figure 2.2 in order to simplify the functionality of sensors for localization. As simple devices, sensors only monitor road traffic and register vehicle-detection timestamps into their local repositories. A server called the APL server processes the complex computation for localization. Specifically, the localization procedure consists of the following steps as shown in Figure 2.2: Step 1: After road traffic measurement, sensor s i sends the APL server its vehicledetection timestamps along with its sensor ID, i.e., (s i,t i ), where s i is sensor ID and T i is timestamps. Step 2: The traffic analysis module estimates the road segment length between two arbitrary sensors with the timestamp information, constructing a virtual topology H v = (V v,m v ), where V v is the vertex set of sensor IDs, and M v is the matrix containing the distance estimate of every sensor pair. Step 3: The prefiltering module converts the virtual topology H v into a virtual graph G v = (V v,e v ), where V v is the vertex set of the sensor IDs, and E v is the adjacency matrix of the estimated road segment lengths.

26 13 Step4: The graph-matching module constructs a reduced virtual subgraph G v = (Ṽv,Ẽv) from the virtual graph G v, where Ṽv is a set of only intersection nodes among V v, and Ẽ v is a set of edges whose endpoints both belong to Ṽv. Gv is isomorphic to the real graph G r = (V r,e r ). The graph-matching module then computes a permutation matrix P, making the reduced virtual subgraph G v = (Ṽv,Ẽv) be isomorphic to the real graph G r = (V r,e r ). Step 5: The location identification module determines each sensor s location on the road map by applying the permutation matrix P to both the reduced virtual subgraph G v and the real graph G r. Through this mapping, node location information (s,l) is constructed such that s is the sensor ID vector, and l is the corresponding location vector; that is, l i = (x i,y i ), where s i is the sensor ID, x i is the x-coordinate, and y i is the y-coordinate in the road map. Step 6: With (s,l), the APL server sends each sensor s i its location with a message (s i,l i ). In the rest of this section, we describe the technical content of each step. We start with the second step, because the operations in step 1 are straightforward Step 2: Traffic Analysis for Road Segment Length Estimation In order to estimate road segment lengths, we found a key fact that vehicle arrival patterns in one sensor are statistically maintained at neighboring sensors close to the sensor. This means that the more closely the two sensors are located, the more correlated the vehicle-detection timestamps are. Consequently, we can estimate road segment length with estimated movement time between two adjacent sensors using the correlation of the timestamp sets of these two sensors, along with the vehicle mean speed (i.e., speed limit given on the road segment). Through both outdoor test and simulation, we found that we can estimate the lengths of road segments used by vehicles during their travels on roadways only with vehicle-detection timestamps.

27 14 S 1 t1,1 t1, 2 t1, 3 t1, 4 t1, 5 t 1, 6 t S 3 t3,1 t3, 2 t3, 3 t3, 4 t3, 5 t 3, 6 t S 2 t2,2 t2, 3 t2, 6 t2, 8 t2, 1 t 2, 12 t2,1 t2, 4 t2, 5 t2, 7 t2, 9 t 2, 11 (a) Detection Sequence for Vehicles at Sensors s 1, s 3, and s 2 t S 1 t1,1 t1, 2 t1, 3 t1, 4 t1, 5 t 1, 6 t S 2 t2,2 t2, 3 t2, 6 t2, 8 t2, 1 t 2, 12 t2,1 t2, 4 t2, 5 t2, 7 t2, 9 t 2, 11 (b) TDOD between Timestamps t 1,1 and t 2,i t S 1 t1,1 t1, 2 t1, 3 t1, 4 t1, 5 t 1, 6 t S 2 t2,2 t2, 3 t2, 6 t2, 8 t2, 1 t 2, 12 t2,1 t2, 4 t2, 5 t2, 7 t2, 9 t 2, 11 (c) TDOD between Timestamps t 1,2 and t 2,i t Figure 2.3: Time Difference on Detection (TDOD) Operation Time Difference on Detection (TDOD) Operation The Time Difference on Detection (TDOD) for timestamp sets T i and T j from two sensors s i and s j is defined as follows: d ij hk = t ih t jk (2.1) where t ih T i for h = 1,..., T i is the h-th timestamp of sensor s i and t jk T k for k = 1,..., T j is the k-th timestamp of sensor s j. We define a quantized Time Difference on Detection (TDOD) as follows: ˆd ij hk = g(dij hk ) (2.2)

28 where g is a quantization function to map the real value of d ij hk to the discrete value. The interval between two adjacent quantization levels is defined according to the granularity of the time difference, such as 1 second,.1 second or 1 millisecond. The number m of quantization levels (i.e., q k for k = 1,...,m) is determined considering the expected movement time of vehicles in the longest road segment of the relevant road network. We define frequency as the count of a discrete time difference. After the TDOD operation for two timestamp sets from two sensors, the quantization level with thehighest frequency (i.e., ˆd ij ) is regarded as the movement time of vehicles for the roadway between these two sensors s i and s j as follows: 15 ˆd ij arg max q k f(q k ) (2.3) where f is the frequency of quantization level q k for k = 1,...,m; that is, in (2.3), the value of ˆd ij is set to the quantization level q k with the maximum frequency. The movement time on the road segment can be converted into road segment length using the formula l = vt, where l is the road segment s length, v is the vehicle mean speed, and t is the vehicle mean movement time on the road segment. 25 Estimated Movement Time: 7.3 sec Frequency (Time Difference Count) Time Difference [sec] Figure 2.4: Estimation of Movement Time through TDOD Operation For example, Figure 2.3 shows the Time Difference On Detection (TDOD) operation for the vehicle detection sequence. Figure 2.3(a) shows the detection sequence for vehicles at intersection nodes s 1, s 2, and s 3 in Figure 2.1(e), where s 2 is a common neighbor of s 1 and s 3. Figure 2.3(b) and Figure 2.3(c) show the TDOD operation for nodes s 1 and s 2 that is a kind of Cartesian product for two timestamp sets. Figure 2.4 shows the histogram [19] obtained by the

29 TDOD operation for two timestamp sets. The time difference value (7.3 sec) with the highest frequency indicates the estimated movement time between two nodes. 16 A D 8 [m] B C 9 [m] Figure 2.5: Road Network for Outdoor Test We performed outdoor test to verify whether our TDOD operation could give good estimates for road segment lengths in terms of vehicle movement time. The results of outdoor test indicate that our TDOD can give reasonable road segment length indicators. Figure 2.5 shows the road map of local roadways in Minnesota for outdoor test. The test roadways consist of four intersections A, B, C, and D. Speed limit on these road segments is 64 km/h (or 4 MPH). We performed vehicle detection manually for more accurate observation; note that it is hard to get accurate vehicle detections at intersections with the current motes due to the sensor capability and mote s physical size, so the development of the vehicle detection algorithm based on motes is left as future work. Table 2.1: Outdoor Test Results Expected Measured Road Segment Distance Movement Time Movement Time A and B 8 m 45 sec 43 sec C and D 8 m 45 sec 43 sec B and C 9 m 51 sec 54 sec D and A 9 m 51 sec 56 sec Table 2.1 shows the expected movement times and measured movement times for these four

30 17 road segments through TDOD. It can be seen that the estimated movement times are close to the expected movement times; note that even though the manual measurement can introduce some human errors, this experimental result shows the significant evidence that the TDOD can provide us with the estimates accurate enough to perform the localization. Therefore, with the TDOD, the distance estimates between two arbitrary nodes can be obtained for the virtual edges in the virtual topology, as shown in Figure 2.1(b). Frequency Non aggregation Method Estimated Movement Time Frequency Time Difference [sec] Aggregation Method 8 Estimated Movement Time Time Difference [sec] Figure 2.6: Comparison between Non-aggregation Method and Aggregation Method Enhancement of Road Segment Length Estimation We found that an estimate close to a road segment s length cannot always be obtained by the maximum frequency through the TDOD operation discussed previously. The reason is that there are some noisy estimates with higher frequencies than an expected good estimate. In order to resolve this problem, we introduce an aggregation method where the mean of several adjacent time differences becomes a new TDOD value, and the sum of frequencies of those is the corresponding frequency. This is based on an observation that time differences close to a real time difference (i.e., movement time needed by a vehicle with the vehicle mean speed on a road segment) have relatively high frequencies by the TDOD operation for two timestamp series, as shown in Figure 2.3. On the other hand, we observe that a noisy estimate with the highest frequency occurs randomly, and its neighbor estimates have relatively low frequencies.

31 18 This method based on TDOD aggregation is called as theaggregation Method and the previous simple TDOD is called as thenon-aggregation Method. We determine the aggregation window size proportionally to standard deviation σ v of the vehicle speed, such as c σ v for c >. Figure 2.6 shows the comparison between the non-aggregation method and aggregation method through simulation. The aggregation window size is 1 such that the vehicle speed deviation σ v is 1 km/h and the window size factor c is 1. Starting from the time difference value of zero in the histogram for Non-aggregation Method, we choose a representative of the adjacent time difference values within the aggregation window size as the mean of them, and then sum their frequencies into the representative s frequency. We then move the window to the right by the unit of time difference value and repeat the computation of the representative and frequency. Thus, the histogram foraggregation Method is obtained by this moving window. We found that for the road segment between sensors s 2 and s 3 in Figure 2.1(e) whose real time difference is 9.36 sec with the vehicle speed µ v =5 km/h, the non-aggregation method makes a wrong estimate (i.e., 26.8 sec), but the aggregation method makes a correct estimate (i.e., 9.3 sec). Thus, this aggregation method can be used to obtain good estimates for road segment lengths in virtual topology Step 3: Prefiltering Algorithm for Virtual Graph The prefiltering algorithm is performed to make a virtual graph that has only edge estimates among virtual edges (i.e., distance estimates) in the virtual topology (e.g., Figure 2.1(b)) obtained from the TDOD operations in Section Our prefiltering algorithm consists of two prefilterings: (i) Relative Deviation Error and (ii) Minimum Spanning Tree. We explain the prefiltering procedure and the effect of two prefilterings on a virtual topology using Figure 2.7. As shown in Figure 2.7(a), there is a partial road network of the entire one shown in Figure 2.1(a) containing sensors {s 1,s 2,s 3,s 4,s 5,s 19,s 2,s 22 }. In the virtual topology, two arbitrary sensors among them have a distance estimate, as shown in Figure 2.7(b). Using the prefiltering based on the relative deviation error, we remove the virtual topology s edges corresponding to inaccurate path estimates, and we then construct a virtual graph, shown in Figure 2.7(c). Next we apply the prefiltering based on the minimum spanning tree to the virtual graph, so the virtual graph containing only the edge estimates is constructed by removing accurate path estimates, as shown in Figure 2.7(d). In this section, we explain the idea of these two prefilterings for obtaining the virtual graph G v = (V v,e v ) from virtual topology