Vehicle Localization Enhancement with VANETs

Transcription

1 2014 IEEE Intelligent Vehicles Symposium (IV) June 8-11, Dearborn, Michigan, USA Vehicle Localization Enhancement with VANETs Ali Ufuk Peker, Tankut Acarman, Çağdaş Yaman, and Erkan Yüksel Abstract This paper presents an assisted system for vehicle localization and map-matching by utilizing Vehicle ad-hoc Networks (VANETs). Fusion of the GNSS and odometer measurement is augmented by ranging distance. Ranging is computed by exchanging the data packets between the vehicular nodes equipped with Dedicated Short Range Communication (DSRC) modem and GNSS receiver. Time-of-Arrival (ToA) of exchanged data packet between the two vehicular nodes is converted in distance. Map matching enhances accuracy of localization while projecting the result of multilateration created by numerous ranging queries. Realistic simulations are conducted to test the performance of the algorithm. Test results show bounded and acceptable particle filter positioning results. The scenario of GPS outages and low number of vehicles collaborating for positioning are simulated. Tracking performance of the particle filter is illustrated. Algorithm helps dead reckoning when GPS data is not available temporarily. A simple GPS receiver is fused with odometer data during tests. Particularly, expensive sensors are not used to achieve better price/performance towards commercial usage. V I. INTRODUCTION ANETs lead to many collaborative schemes and opportunistic approaches by allowing communication among the decentralized vehicular nodes. The vehicular nodes need to self-localize and be able to timestamp the events to perform networking and routing in VANET. GNSS is used a standard technology for vehicle localization. But line-of-sight occlusions and multipath issues in urban areas degrade the receiver performance and affect VANETs. To solve poor localization performance inherited by standard GNSS receivers, cooperation among the vehicular nodes on a digital map are being investigated. The Peer-to-Peer (P2P) Cooperative Positioning and Communications has been proposed to improve the standalone GNSS receiver performance by allowing information exchange between collaborative users, [1], [2]. The project on hybrid P2P The authors gratefully acknowledge the support of Galatasaray University, scientific research support program under grant # A. U Peker is with the Computer Engineering Dept. Bogazici University, Bebek, Istanbul, Turkey ( upeker@infotech.com.tr). T. Acarman is with the Computer Engineering Dept., Galatasaray University, Ortakoy, Istanbul, Turkey (phone: ext. 112 fax: acarman@ieee.org). Ç. Yaman is with the Computer Engineering Dept. Galatasaray University, Ortakoy, Istanbul, Turkey ( cyaman@infotech.com.tr). E.Yüksel is with the Computer Engineering Dept. Galatasaray University, Ortakoy, Istanbul, Turkey ( erkan.yuksel@ykare.com). cooperative position is achieved by exchanging GPS raw measurements from the other vehicular nodes and road side units in the radio transmission range. The use of the relative positioning at pseudorange level and geometric constellation is performed on a precise map which is called cooperative map matching in the Covel project, [3]. For cooperative positioning, each vehicular node must have a GNSS receiver, a DSRC system and range calculation method to estimate the distance from its neighbors. Several technologies exist for ranging purposes (e.g. received signal strength, two-way time-ofarrival, time-difference-of-arrival, angle-of-arrival or Doppler measurements). Multipath effects degrading GNSS receiver accuracy are detected by the trained neural network. The Accuracy and robustness is improved by vehicular communication, [4]. An overview of localization techniques for wireless communication, i.e., radio localization, is presented along its sources of errors in [5]. Ranging systems are heavily dependent on clock accuracy, hardware-based time-of-arrival delay calibration is presented in [6]. Digital map and road infrastructure is used for better localization. Use of particle filter generating hypotheses on the probability of being in a certain road segment, fusion of the existing onboard odometer and GNSS measurements is presented in [7]. In this paper we present an algorithm which uses particle filter augmented by mobile ranging in VANET. Following [7], fusion of odometer and GNSS measurements is used for localization. A simple map matching is used for localization. Mobile ranging is applied for computation of ToA while exchanging timestamped packets at physical layer with multiple neighbor vehicular nodes. The error sources and characteristics of these sensors are different. Weight update depending on the measurement availability and accuracy is applied in the particle filter algorithm. Realistic simulators such network simulator ns-3 for DSRC with IEEE p MAC protocol and mobility simulator SUMO are used to model the packet and vehicle traffic. Particle filter is implemented in ns-3 with C++ code to achieve computational performance. This paper is organized as follows: Particle filter algorithm is revisited in Section II. Implementation issues are presented in Section III. Simulation study is presented in Section IV. And finally, some conclusions are given. II. PARTICLE FILTER ALGORITHM A. Localization A moving vehicle is considered for localization. Model linear in the state dynamics and non-linear in the /14/$ IEEE 661

2 measurements is given by, (1) ( ) (2) where for the vehicle localization problem state is represented as a two-dimensional vector x t = [Lon, Lat] T, u t = w wheel measured odometer input, w t faults, z t =[ Lon GNSS Lat GNSS ] T denotes measurement vector constituted by the localization algorithm based on the multilateration ( ) and GNSS receiver measurement of the ego vehicle ([Lon, Lat] T ), and e t denotes the masurement error vector. B. Bayesian Filter Consider a system with state x t at a given time t. Initial system state can be modeled by its probability distribution p(x 0 ), where X 0 is the state at time t = 0. Since the system is a Markov process of first order probability distribution at current time t only depends on previous state which can be denoted by p(x t X t-1 ). System can be monitored by sensor measurements z t, with some noise. Probability of receiving measurement z t given the current state of system is donated by p(z t =z t X t ). Computation of the posterior density p (x t Z t ) is memoryless, complete state history is not needed. By using these assumptions the posterior density can be expressed as follows: ( ) ( ) ( ) ( ) Here p (x t Z t-1 ) is the prediction, which shows the probability of the vehicle being in the location denoted by x t with respect to the history of sensor measurements denoted by Z t-1. Prediction is given by integrating motion model p (x t x t-1 ) over posterior density p (x t-1 Z t-1 ). (3) ( ) ( ) ( ) (4) p(z t Z t-1 ) is normalizing constant (k t ). When (4) is inserted into (3), the update statement of the Bayes filter is calculated as follows: ( ) ( ) ( ) ( ) (5) Posterior probability is calculated from previous position by integrating the motion model and the sensor data. Minimum mean square estimate is calculated by use of posterior probability, ( ) (6) C. Particle Filter Numerical approximation of Bayes Filter can be implemented by use of Sequential Monte Carlo (SMC) methods which is also known as particle filters. The sampling importance resampling (SIR) algorithm is one of the most widely used sequential Monte Carlo methods, which allow system state estimates to be computed on-line while the state is changed. A SIR filter usually manages a fixed number of possible system state hypotheses x i t, which are also called particles. Ideally, these particles approximate the distribution of the system state, p(x t ). The SIR algorithm stages are iterated over discrete time steps. With particle filters posterior p (x t Z t ) is approximated by weighted sample set. A basic framework with particle filters for localization problems in different application areas proposed by Gustafson et al., [8]. This approach converts difficult integrals to computable summations. Particle filter implementation is as follows: -- Initialize particles for i = 1,,N. -- Measurement update, update particle weights by use of likelihood. ( ), for i = 1,,N. -- Normalize weights such that total of weights is 0 is an approximation to (6) for calculating current state. -- Resampling: If number of efficient samples fall below a certain value resampling is required. This can be done by calculating as follows. ( ) ( ) -- When is lower than a certain threshold resampling can be done. The value for can be set to a value depending on. Resampling is done by replacing of the particles which has small weights. After resampling all the weights are set to same value. -- Transition: According to the motion model particles will be updated. A. Transition model III. IMPLEMENTATION The vehicle speed is measured by the odometer and it is denoted by. Particles are randomly distributed at each sampling period to predict the displacement of the ego vehicle. A transition is performed as arbitrary movement on 2D plane. Since the heading angle of the ego vehicle is not measured, this transition is considered as follows: ( ) ( ) ( ) ( ) ( ) ( ) xhere X and Y are randomly generated numbers in the interval of [-1, 1]. Transition is enforced by doubling the ego vehicle s speed to prevent the prediction bias. Also the speed value is saturated at a certain minimum value for exciting prediction and tuning the position estimation even when vehicle is stationary. B. Weight update For each particle, initial weight is calculated based on the possibility derived by the measurement of GNSS position and ToA. Probability is calculated under the assumption that the deviations of GNSS and ToA measurements are 662

3 normally distributed and standard deviation is changed with respect to the confidence values. Concerning the measurements of ToA, the number of vehicular nodes, which are being communicated to determine mobile ranging, determines the confidence value. First weights are normalized then location information is calculated by use of linear weights. All the information is stored in estimations array. C. Resampling Resampling improves the estimation of future states by concentrating particles into domains of higher posterior probability current estimate by increasing the variance of the estimate after resampling. So, resampling must be applied with caution. Resampling only occurs if estimated particle size (ESS) is lower than a certain threshold. If the ESS smaller than this threshold means that only small portion of particles are active in the estimation. If this is the case, particles with small weight are deleted. Surviving particles are copied according to their respective weights. Resampling is run after mean calculation. Weights are normalized at mean calculation stage. D. ToA Measurement Small size data packet is sent to each of the surrounding vehicular nodes with a timestamp on it. The receiver vehicular node is replying with an adding of another timestamp denoting the internal delay of the node for processing at physical layer. Each vehicular node in the vicinity is communicated in an unicast scheme for the period of 10 milliseconds. Figure 1: Ranging measurement between the ego vehicle and the other five neighbor vehicles Multilateration algorithm is used to estimate the ego vehicle location by means of communicating with the neighbor vehicles in the radio transmission range and exchanging small size data packets, in which the GNSS position measurement of each vehicle is transmitted with its timestamp. To measure ranging distance with respect to the neighbor vehicles, ToA values of these exchanged packets are calculated. ToA is calculated by means of counting the quartz oscillator s cycles between the instances of sending and receiving the data packet. As illustrated in Fig. 1, mutual positioning and ranging is obtained the result of multilateration and it is an estimation problem. For simplicity, we assume we have 1 measurement to 3 stations: ( ) ( ) ( ) ( ) (10) ( ) ( ) where, i=1,2,3 denotes the unknown measurement noises. ( ) are the unknown coordinates that need to be estimated. The ranging distances are given by, ( ) ( ) (11) Range estimation problem is fit into Batch Least Squares (BLS) problem by linearizing (11) about its approxiated known initial coordinates, i.e., ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) (12) (13) By using Taylor Series approximation and ignoring higher order terms, range estimation problem is defined by, ( ) ( ) ( ) (14) Then, estimation problem of range in VANET is straightforward, where θ=[lon Lat] T is the position of the ego vehicle and needs to be estimated in terms of the ranging with the neighbor vehicles z(k)=[r 1 r 2 r 3... r k ] T. The noise is introduced by the noise vector v(k)=[v 1 v 2 v 3... v k ] T. In order to estimate its position, the ego vehicle exchanes beaconing messages with its neighbor vehicles and for the given number, the raw data replied by the neighbor vehicles are chosen from the beaconing table randomly. In the next section IV, the estimation accuracy is evaluated with respect to the number of neighbor vehicles. IV. SIMULATIONS AND RESULTS A. Simulation Setup Simulator of Urban Mobility (SUMO) is used to generate the position of the vehicular nodes at a given speed interval defined by its minimum and maximum value. Inner city area is chosen and digital map of Kadıköy region, Istanbul is used. The vehicular nodes are artificially generated with a frequency of 1vehicle/seconds at the main intersections. Network simulator is used to validate vehicular communication. IEEE p physical and MAC layers are implemented. Location tables and beaconing protocol is used subject to C2C network and transport layer standards (ETSI). An interface is created by means of 663

4 Figure 2 : OpenGL interface of the network and mobility simulators OpenGL and the digital map attributes such the height of the buildings, the number of lanes is implemented. The urban road topology, generated particles along the ego and neighbour vehicle trajectories are plotted in Fig. 2. List of the parameters is given in Table 1. The positions of the vehicular nodes are controlled by the trace files of SUMO, which acts as a server to the ns-3 client. Particle Filter and multilateration algorithm is implemented on a Matlab server which is connected through a socket to the ns-3 client. Connection of the simulators and their functions are given in Table 1. Figure 3 : Attenuation of received signal strength due to the building obstacles. As the destination vehicular node moves, signal is attenuated by one, then two buildings. Table 1 - List of simulation parameters Parameter Value Scenario Urban Canyon Vehicle Number 240 Vehicle Velocity Max 20 m/s Vehicle Penetration Rate 3 veh/s Path Loss Exponent 3 Propagation Loss Model NLOS + Nakagami Transmit Rate 6 Mbps Transmit Power 19 dbm Carrier Sensing Threshold -99 dbm Transmit Range 550 m Reference Distance 1 m Beaconing Rate 10 Hz Beacon Size 100 bit Simulation Time 100 s In the network simulator, the following data processed by the Particle filter and multilateration algorithm: Individual time-of-arrival value obtained by exchanging packet with the neighbour vehicle, Ego vehicular node position subject to Gaussian white noise, GNSS measurement and its variance of the neighbour vehicular nodes, Road IDs of the neighbour vehicular nodes. Particle filter uses the odometer readings to be used in the transition model. Weights are distributed for GNSS and ToA measurements, and map matching precision. The map matching is simply ensured by finding the shortest distance to the road edge. The inverse of this matching error is multiplied by the weight derived by the measurements of GNSS and ToA. This overall weight is updated for each individual particle towards prediction of localization. Figure 4: Scatter plot of GNSS positioning error (marked with + ), ToA error (marked with o ), and the results of fusion by particle filer (marked with * ) Towards realistic simulation modeling, it is important to include the radio coverage range in VANET. The wireless media is considered to be crucial at the stage of replying to the individual ToA calculation query, line-ofsight occlusions and multipath in urban areas degrade the VANET performance. By using the digital map attributes, the buildings are modeled based on their dimension to simulate the signal propagation. Radio shadowing in urban environments and the used model relies only on building outlines, which are available in digital map. Computationally inexpensive model by considering the LOS between sender and receiver is used, [9]. For each obstacle, [ ] (15) the additional attenuation of a transmission due to an obstacle is given by. This attenuation model is a weighted sum of the number of times n the border of the obstacle is intersected by the line of sight and the total length of the obstacle s intersection. The coefficient, β, given in db per wall and is the attenuation of transmission due to exterior wall of a building. The 664

5 second coefficient, γ, given in db per meter, roughly models the internal structure of the building. These two calibration factors are experimentally calculated in [9] and they are used in our simulation study. Fig. 3 illustrates transmission inside the radio obstacles. B. Simulation Results The simulation scenarios are tested to validate the effectiveness of the particle filter algorithm. Comparisons of the measurement errors coming from GNSS receiver and VANET ranging are plotted versus filter positioning error in Figure 4. GNSS measurement error is subject to normal distribution with standard deviation of 10 meters, whereas ToA measurement error is subject to normal distribution with standard deviation of 10 meters propagation delay. In Fig. 5, the time responses of each positioning error are plotted. The particle filter generates bounded positioning error during the outage and errors are lower than standalone low accuracy GNSS receiver, Fig.6. Figure 5: The error responses of GNSS and ToA measurements in comparison to the Particle filter error result Figure 6: The error responses of particle filter algorithm during the existence and absence of GNSS measurement Multilateration algorithm is using ranging measurements based on ToA of the exchanged packets in VANETs as well as map matching information. Therefore, reliability of self localization with respect to the neighbour vehicles is expected to be dependent on the number of the corresponding neighbour vehicular nodes and their probabilistic distribution on the road. For example, to assure sufficient richness, the ranging information needs to be provided by the neighbour vehicles on the perpendicularly oncoming roads and preferably not from the following neighbours. To assure probabilistic selection, a table is created by the replied raw data including GNSS position, time and road ID of the neighbour vehicles. Then, among these lines, given number of vehicles is randomly selected. This random selection is achieved by matching the randomly generated number with the line of the table. If the does not exist or previously selected, a new number is generated. Following this approach, the simulation scenarios are repeated for different number of ranging measurement selection. In Table 2, localization of the ego vehicle is simulated by using the neighbour vehicles GNSS position, ranging measurement based on ToA calculation and displacement of the ego vehicle by using information provided by its odometer. As the number of neighbour vehicles contributing to the multilateration algorithm is increased, Circular Error Probability (CEP) is decreased where CEP refers to the radius of a circle in which 50% of the values occur. In a similar way, Distance Root Mean Squared (DRMS) is decreased also, where DRMS is related to 2D accuracy. R95 is defined as the radius of circle centred at the true position, containing the position estimate with probability of 95%. Following the results given in Table 2, positioning precision below 3 meters can be achieved when the number of the ranging measurement is more than 5. Table 2 : Positioning error for different suite of measurement sources: GNSS receiver, ToA and wheel odometer Mutual Positioning CEP DRMS R95 2DRMS GNNS TOA FILTER GNNS TOA FILTER GNNS TOA FILTER GNNS TOA FILTER GNNS TOA FILTER The number of vehicles is used to update the weight to ensure the contribution of this measurement into the filter algorithm. The second scenario is repeated in the absence of GNSS measurement. GNSS outage is enforced and filter relies on ToA measurements and 665

6 Table 3 : Positioning error during GNSS outage GNSS Outage CEP DRMS R95 2DRMS 3 TOA FILTER TOA FILTER TOA FILTER TOA FILTER TOA FILTER odometer readings only. The particle filter generates bounded positioning error during the outage and errors are lower than standalone low accuracy GNSS receiver. Accuracy of ToA measurements depend on the number of vehicles. In this GNSS outage scenario, higher number of ranging measurement leads to better accuracy in the positioning responses, Table 3. As an interesting simulation scenario, tunnel passage is simulated and the road side units placed at each 140 meters is used to send their accurate position to the unicast ToA echo request for the purpose of mutual positioning. The tunnel passage is 1000 meters and the mutual positioning metrics are given in Table 4. Table 4 : Performance in Tunnel Passage Scenario Only ToA MEAN CEP %65 %95 TOA FILTER VANET performance plays an important role to fulfill the requirements of mutual positioning by using VANET assistance. During the simulation study, two metrics are observed: measurement success ratio, which is the ratio between the ToA echo requests of the ego vehicle and the replies, the second metric is the ratio of the packet success, which is the number of the received packets versus the overall number of the generated packets. The number of received packets is calculated by counting the sequence number of the ToA echo request packets. The performance metric of VANET is given in Table 5 and it is promising in terms of the success ratio which is above %90 for the largest number of the neighbor vehicles, which is 11 nodes, used in the simulation scenario subject to radio signal obstacles, i.e., buildings in an urban area. V. CONCLUSION All available information provided by VANET, GNSS and the vehicle itself is fused to ensure bounded and available positioning information. New approach such as particle filter is used to adjust contributions of the inputs depending on their availability and accuracy. Realistic error modeling and road scenarios are simulated. Sensor suite of GNSS receiver, vehicular wireless communication modems, commercially available Table 5 : Performance of measurement and packet delivery Number of nodes Measurement Success Ratio Packet Success Ratio navigable digital map and the vehicle odometer is adequate for continuous and reliable localization. Particle filter and weight update with respect to the vehicle speed can be proposed to enhance the reliability and accuracy of vehicular localization. Real-time implementation of ranging measurement and road tests are underway. ACKNOWLEDGMENT This article presents some ideas and results of GLOVE [10], a project co-funded under the 7 th Framework Program, in the context of Exploiting The Full Potential- GALILEO-2011-GSA-1, Grant Agreement n The concepts here presented are investigated within GLOVE s WP3. The article reflects the views only of the authors and the EU cannot be held responsible for any use that may be made of the information contained therein. REFERENCES [1] R. Garello, L. Lo Presti, G. E. Corazza, J. Samson, Peer-to-Peer Cooperative Positioning, Part I: GNSS-Aided Acquisition. In: Inside GNSS, pp , March-April [2] R. Garello, J. Samson, M. A. Spirito, H. Wymeersch, Peer-to- Peer Cooperative Positioning, Part II: Hybrid Devices with GNSS & Terrestrial Ranging Capability. In: Inside GNSS, pp ISSN X. [3] M. Obst, R. Schubert, N. Mattern, C. Liberto,S. Romon, L. Khoudour, "Cooperative GNSS Localization In Urban Environments Results From The Covel Project", The Proceedings of 19th ITS World Congress, Vienna, Austria, October 2012, pp [4] N.M. Drawil, and O. Basir, Intervehicle-Communication- Assisted Localization, IEEE Transactions on Intelligent Transportation Systems, vol.11,no.3, pp ,2010. [5] J.J. Caffery, and G.L. Stüber, Overview of Radiolocation in CDMA Cellular Systems, IEEE Communication Magazine, pp.38-45, April [6] D.D. McCrady, L. Doyle, H. Forstrom, T. Dempsey, and M. Martorana, Mobile Ranging Using Low-Accuracy Clocks, IEEE Transactions on Microwave Theory and Techniques, vol.48, no.6, pp , [7] A.U. Peker, O. Tosun, T. Acarman, Particle filter vehicle localization and map-matching using map topology" IEEE Intelligent Vehicles Symposium (IV), 3-5 June 2011, Baden Baden Germany, pp [8] F. Gustafsson, F. Gunnarson, N. Bergman, U. Forssell, J. Jansson, R. Karlsson, and P. Nordlund, Particle Filters for Positioning, Navigation and Tracking, IEEE Transactions on Signal Processing, vol. 50, No. 2, pp , February [9] C. Sommer, D. Eckhoff, R. German and F. Dressler, Computationally Inexpensive Empirical Model of IEEE p Radio Shadowing in Urban Environments, 8th Int. Conf. On Wireless On-Demand Network Systems and Services, pp.84-90, [10] R. Scopigno, et al. GLOVE s manifesto: leveraging GNSS timespace information to improve VANETs performances and enrich VANETs services, in Proc. of 9 th ITS European Congress 2013, Dublin, Ireland. 666