VT-2005-00647.R2 1 Position and Velocity Tracking in Mobile Networks Using Particle and Kalman Filtering with Comparison Mohammed M. Olama, Seddik M. Djouadi, Ioannis G. Papageorgiou, and Charalambos D. Charalambous Abstract—This paper presents several methods based on signal strength and wave scattering models for tracking a user. The received signal level method is first used in combination with Maximum Likelihood (ML) estimation and triangulation to obtain an estimate of the location of the mobile. Due to non-lineof-sight conditions and multipath propagation environments this estimate lacks acceptable accuracy for demanding services, as numerical results reveal. The 3D wave scattering multipath channel model of Aulin is employed together with recursive nonlinear Bayesian estimation algorithms to obtain improved location estimates with high accuracy. Several Bayesian estimation algorithms are considered such as the extended Kalman filter (EKF), the particle filter (PF), and the unscented particle filter (UPF). These algorithms cope with nonlinearities in order to estimate the mobile location and velocity. Since the EKF is very sensitive to the initial state, we propose to use the ML estimate as an initial state to the EKF. In contrast to the EKF tracking approach, the PF and the UPF approaches do not rely on linearized motion models, measurement relations, and Gaussian assumptions. Numerical results are presented to evaluate the performance of the proposed algorithms when measurement data do not correspond to the ones generated by the model. This shows the robustness of the algorithm based on modeling inaccuracies. Index Terms—Kalman filtering, location tracking, maximum likelihood estimation, multipath fading channels, and particle filtering. T I. INTRODUCTION HE need for an efficient and accurate mobile station (MS) positioning system is growing day by day. This has been stressed by a federal order issued by the federal communications commission (FCC), which mandates all wireless service providers to provide public safety answering points with information to locate an emergency 911 caller with an accuracy of 100 meters for 67% of the cases [1]. It is M. M. Olama and S. M. Djouadi are with the Electrical and Computer Engineering Department, University of Tennessee, Knoxville, TN 37996 USA, (phone: 865-974-5447; fax: 865-974-5483; e-mail: {molama, mdjouadi}@utk.edu). I. G. Papageorgiou and C. D. Charalambous are with the Electrical and Computer Engineering Department, University of Cyprus, Nicosia, 1678 Cuprus, (e-mail: {ioannisp, chadcha}@ucy.ac.cy). also expected that the FCC will tighten its requirements in the near future [2]. Many other applications, such as vehicle fleet management, location sensitive billing, intelligent transport systems, fraud protection, and mobile yellow pages have driven the cellular industry to research new and promising technologies for MS positioning. The problem of determining the location and velocity of a MS has been studied extensively in the last few years. The current standards in estimating the location and velocity are based mostly on time signal information, such as time difference of arrival (TDOA), enhanced observed time differences (E-OTD), observed time difference of arrival (OTDOA), global positioning system (GPS), etc., [3]-[7]. However, not all of these methods meet the necessary needs imposed by certain services. In addition, most of them require new hardware since localization is not inherent in the current wireless systems, for instance, GPS demands a new receiver and TDOA, E-OTD, OTDOA requires additional location measurement units in the network [8]. Adding extra hardware means extra cost for implementation, which can be reflected on both consumers and operators. Researchers have also suggested several MS location methods based on signal power measurements such as in [9] and [10], where a certain minimization problem is solved numerically to get an initial estimate of the MS position, and then a smoothing procedure such as linear regression [9], or the Kalman filter [10] are applied to obtain a more accurate estimate. In this paper, several MS tracking methods based on maximum likelihood estimation (MLE) [11], and recursive nonlinear Bayesian estimation (RNBE) algorithms such as the extended Kalman filter (EKF) [12], the particle filter (PF) [13], and the unscented particle filter (UPF) [14] are proposed. The MLE algorithm employs the average received power measurements based on the lognormal propagation channel model to obtain an initial MS location estimate [3, 16]. These measurements are readily available through network measurement reports or radio measurements, in idle or active mode, for any MS unit in 2G and 3G cellular networks. The RNBE algorithms employ the instantaneous electric field measurements based on the 3D multipath channel model of Aulin [15] to account for multipath and non-line-of-sight (NLOS) characteristics of the wireless channel as well as the VT-2005-00647.R2 dynamicity of the MS. The received instantaneous electric field in this model is a nonlinear function of the position and velocity of the MS. The EKF approach is based on linearizing the nonlinear system model around the previous estimate, and therefore is very sensitive to the initial state. This motivates the use of the ML estimate of the MS location as an initial state to the EKF. Particle filtering approaches approximate the optimal solution numerically based on the physical model, rather than applying an optimal filter to an approximate model such as in the EKF. They provide general solutions to many problems where linearization and Gaussian approximations are intractable or yield low performance. The more nonlinear the model is or the more non-Gaussian the noise is, the more potential PFs have, especially in applications where computational power is rather cheap and the sampling rate is moderate. In this paper, particle filtering is implemented for the generic PF and the more recent UPF. Aulin’s model postulates knowledge of the instantaneous received field at the MS, which is obtained through the circuitry of the mobile unit. The proposed RNBE algorithms take into account NLOS condition as well as multipath propagation environments. They require only one base station (BS) to estimate the MS location instead of at least three BSs as found in the literature [10], [16]. However, an initial MS location estimate that requires at least three BSs, such as the MLE and triangulation method, will improve the convergence of the RNBE filter. Particle filtering has been used in several tracking wireless applications [17]-[20], but the channel models used do not take into account the multipath properties of the wireless channel. To the best of our knowledge, the utilization of the PF and/or the UPF together with the classical wireless channel model to extract the MS location and velocity is new. The performance of the proposed algorithms is computed numerically and in the presence of parameters uncertainty. Numerical results indicate that the proposed UPF algorithm is highly accurate and superior to other approaches. The main contribution of this paper is to develop MS location and velocity estimation algorithms, which use Aulin’s 3D wave scattering model and are based on particle filtering. The performance is compared with that of the EKF. The choice of this channel model is to account for the multipath properties and NLOS of wireless networks. These methods support existing network infrastructure and channel signaling and only use one BS. Moreover, simulations are performed under uncertainties in the model parameters of the channel to evaluate the robustness of the algorithm. The paper is structured as follows. In Section II, we describe the mathematical models for the location and velocity estimation algorithms. The MLE and the EKF approaches for MS location estimation are presented in Sections III and IV, respectively. The PF and the UPF approaches for MS location and velocity estimation, which are the main contribution of this paper, are presented in Sections V and VI, respectively. In Section VII we present numerical results, and evaluate the robustness of the proposed algorithms to uncertainties due to random variations in the channel parameters. Section VIII 2 provides concluding remarks. II. SYSTEM MATHEMATICAL MODELS A. The Lognormal Propagation Channel Model Here we consider a 2D geometry with the MS located at ( x0 , y0 ) and the BSs located at ((x BS1 ) , yBS1 ), ( xBS2 , yBS2 ),..., ( xBSB , yBS B ) . The general lognormal propagation channel model is described by [21]: ⎛d PLsb (db ) = PL(d0b ) + 10ε b log ⎜ b ⎜ d0 ⎝ b ⎞ ⎟ + X bs ⎟ ⎠ (1) where db ≥ d 0b , s ∈ {1, 2,.., S } , b ∈ {1, 2,.., B} , PLsb (db ) is the path loss from the bth BS at distance db for the sth sample, d 0b is the reference distance, εb is the path loss exponent and X bs ∼ N (0; σ b2 ) is a Gaussian random variable (RV) represents the shadowing variance due to gross variations in the terrain profile and changes in the local topography. In cellular networks, the MS preserves and frequently updates, in idle and active mode, the received power of the strongest nonserving BSs (e.g., in GSM the 6 strongest [22]) in addition to the one of the serving cell. Exploiting these measurements from surrounding BSs lead to estimate the location of the MS. The MLE approach described in Section III that employs this channel model is used to estimate the MS location. B. Aulin’s Scattering Model The basic 3D wireless scattering channel model described in [15], which assumes that the electric field, denoted by E(t), at any receiving point ( x0 , y0 , z0 ) is the resultant of P plane waves (see Fig. 1), in which the receiver moves in the X-Y plane having velocity v in a direction making an angle γ with the X-axis, is given by: E (t ) = P P ∑ E (t ) = ∑ r cos (ω t + ω t + θ ) + e(t ) n =1 n n =1 n c n (2) n where Z nth multipath component (x0, y0, z0) MS O Y βn αn γ X v VT-2005-00647.R2 3 Fig.1. Aulin’s 3D multipath channel model. ωn = θn = − 2πυ λ ( cos(γ − α n ) cos β n ) (3) The measurement equation can be found from Aulin’s scattering model (2), (3), and (4), which can be written in discrete form as: P 2π ⎛ x0 cos α n cos β n + y0 sin α n cos β n ⎞ ⎜ ⎟ + φn λ ⎝ + z0 sin β n ⎠ n =1 (4) of the nth component, λ is the wavelength, e(t ) is a white Gaussian noise, and P is the total number of paths. It can be seen from (3) and (4) that the Doppler and phase shifts depend on the velocity and location of the receiver, respectively. Clearly, (2) assumes transmission of a narrowband signal. This assumption is valid only when the signal bandwidth is smaller than the coherence bandwidth of the channel. Nevertheless, the above model is not restrictive since it can be modified to represent a wideband transmission by including multiple time-delayed echoes. In this case, the delay spread has to be estimated. A sounding device is usually dedicated to estimating the time delay of each discrete path such as the Rake receiver [23]. It can be seen that the noisy instantaneous received field in (2) depends parametrically on the location and velocity of the receiver. Consequently, this expression is used to estimate the MS location and velocity by using the EKF, the PF, and/or the UPF. Next, we formulate the location estimation as a filtering problem in state-space form [24]. The general form, once discretized, is given by: (5) where f (.,.) and h(.,.) are known vector functions, k is the estimation step, z k are the output measurements at time step k, and x k is the system state at time step k and must not be confused with location coordinates. Further, w k and v k are the discrete zero-mean, independent state and measurement noise processes, with covariance matrices Q and R , respectively. Now let x k = [ xk , xk , yk , yk ] denote the state of the MS at T time k, where xk and yk are the Cartesian coordinates of the MS, xk and yk are the velocities of the MS in the X and Y directions, respectively. We choose the case where the velocity of the MS is not known and is subject to unknown accelerations. The dynamics of the MS can be written as [17]: 2 0 ⎤ ⎡ xk ⎤ ⎡1 Δ k 0 0 ⎤ ⎡ xk −1 ⎤ ⎡ Δ k / 2 ⎢ ⎥ ⎢ ⎥ ⎢ ⎢ ⎥ ⎥ xk ⎥ ⎢0 1 0 0 ⎥ ⎢ xk −1 ⎥ ⎢ Δ k 0 ⎥ ⎡ wk −1,1 ⎤ ⎢ xk = = + ⎢ ⎥ (6) ⎢ yk ⎥ ⎢0 0 1 Δ k ⎥ ⎢ yk −1 ⎥ ⎢ 0 Δ k2 / 2 ⎥ ⎢⎣ wk −1,2 ⎥⎦ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥⎢ Δ k ⎥⎦ ⎣⎢ yk ⎦⎥ ⎣⎢0 0 0 1 ⎦⎥ ⎣⎢ yk −1 ⎦⎥ ⎢⎣ 0 where Δ k is a (possibly non-uniform) measurement interval between time k –1 and k. 2π xk2 + yk2 ωn = shift, θ n is the phase shift, rn is the amplitude, φn is the phase z k = h( x k , v k ) ) (7) where and α n , β n are spatial angles of arrival, ωn is the Doppler x k = f (x k −1 , w k −1 ) ( zk = h(x k , vk ) = ∑ rnk cos ωc tk + ωnk tk + θ nk + v(tk ) λ k θn = k ( cos(γ − α nk ) cos β nk k −2π ⎛ xk cos α nk cos β nk + yk sin α nk cos β nk ⎜ λ ⎜⎝ + z0 sin β nk ) (8 ) ⎞ ⎟⎟ + φnk ⎠ (9 ) Clearly, the measurement equation h(.,.) is a nonlinear function of the state-space vector, as observed in (7), (8), and (9). If we assume approximate knowledge of the channel, which is attainable either through channel estimation at the receiver (e.g., GSM receiver), or through various estimation techniques (e.g., least-squares, ML), then this problem falls under the broad area of nonlinear parameter estimation from noisy data which can be solved using the RNBE algorithms. These algorithms will be discussed in Sections IV, V, and VI. The MLE algorithm that employs the lognormal propagation channel model is discussed in the next section. III. THE MLE APPROACH FOR MS LOCATION ESTIMATION In this section, The MLE method that employs the lognormal propagation channel model described in section IIA is considered for the MS location estimation. This method exploits the received power measurements at the MS which are available from network measurement reports (NMR). Thus, we write the likelihood function and then maximize it with respect to the distances θ = d = ( d1 , d 2 ,..d B ) from each BS, where θ is the parameter to be estimated. The ML estimator, denoted by θˆ = dˆ = dˆ , dˆ ,.., dˆ , represents the ( 1 2 B ) most possible MS/BS distances based on the measurements available at the MS. Consider the measurement vector for the sth sample from all BSs, denoted by PLs (d ) = ( PL1s ( d1 ), PLs2 ( d 2 ),.., PLsB ( d B ) ) . The distribution function for this vector is the B-variate normal distribution given by: p ( PLs (d) | θ ) = ( 2π ) −B / 2 ( det(Σs ) ) −1 2 ) ( ( T s s ⎛ 1⎛ exp ⎜ − ⎜ PLs (d) − PL (d) Σ−s 1 PLs (d) − PL (d) ⎝ 2⎝ s (10) ) PLs (d) ∼ N B PL (d); Σ s , where s ( ) ⎞⎟⎠ ⎟⎞⎠ ( s s s PL (d) = PL1 (d1 ), PL 2 (d 2 ),.., PL B (d B ) ) is the mean path loss from each BS, and Σ s is the covariance matrix. Assuming the noise is independent identically distributed (iid), then the logarithm likelihood function is the log product of the sample likelihood functions given by: VT-2005-00647.R2 4 ⎛ 1 L ( θ | PLs ( d ) ) = log ⎜ ⎜ (2π ) SB / 2 ( det( Σ ) ) S / 2 s ⎝ ( s ⎛ − ∑ ⎜ PLs ( d ) − PL (d ) s =1 ⎝ S ) ⎞ ⎟ ⎟ ⎠ ) ( s ⎞ Σ −s 1 PLs (d ) − PL ( d ) ⎟ 2 ⎠ T (11) k where S is the total number of samples. Maximizing (11) first s with respect to PL (d) , the score function yields: ^ PLb (db ) = 1 S ∑ PLsb (db ), S s =1 ∀b ∈ {1, 2,.., B} (12) Solving for d̂ using the invariance property of the MLE [11], it can be shown that: dˆb = 10 ⎧⎪ 1 ⎨ ⎪⎩10ε b ⎡1 ⎢ ⎢⎣ S S ⎤ ⎫⎪ s =1 ⎦⎭ ∑ PLsb ( db ) − PL ( d0b )⎥⎥ ⎬⎪ (13) is the MLE for the distance of the bth BS from the MS. Next, we perform triangulation using the least squares error method [25] to estimate the MS location ( x0 , y0 ) , by solving: ( ) 2⎫ ⎧B arg min ⎨∑ db − dˆb ⎬ x0 , y0 ⎩ b =1 ⎭ K k = Pk HTk ⎡⎣ H k Pk HTk + Vk R k VkT ⎤⎦ xˆ k = x k + K k ( zk − h(x k , 0)) Pˆ = (I − K H )P (14) k In the next section, the EKF approach that employs the channel model of Aulin to estimate the MS location and velocity is discussed. IV. THE EKF APPROACH FOR MS LOCATION AND VELOCITY ESTIMATION Consider the general discrete-time dynamical system model described in (5). Let the known probability density functions (PDFs) of the process noise w k and the measurement noise v k be p (w k ) and p (v k ) , respectively. As usual, w k and v k are assumed to be mutually independent. The set of entire measurements from the initial time step to time step k is denoted by Z k = {z i }i =1 . The distribution of the initial k condition x0 is assumed to be given by p (x0 | Z 0 ) = p (x0 ) . The EKF is based on linearizing the nonlinear system models around the previous estimate. The general algorithm for the discrete EKF can be described by the time-update equations given as [12]: x k = f (xˆ k −1 , 0) (15) P = A Pˆ AT + W Q WT k k k −1 k k k −1 k and the measurement-update equations given as: (16) k where ⎡1 ⎢ 0 ∂f Ak = (xˆ k −1 , 0) = ⎢ ⎢ 0 ∂x ⎢ ⎣0 Δk 0 0 ⎤ ⎥ 1 0 0 ⎥ , 0 1 Δk ⎥ ⎥ 0 0 1⎦ ⎡ Δ k2 / 2 0 ⎤ ⎢ ⎥ Δ 0 ⎥ ∂f k Wk = ( xˆ k −1 , 0) = ⎢ , ⎢ 0 ∂w Δ k2 / 2 ⎥ ⎢ ⎥ ⎢⎣ 0 Δ k ⎥⎦ ∂h Vk = ( xˆ k , 0) = 1, ∂v ∂h Hk = (xˆ k , 0) = [ H 1k H 2k H 3k H 4k ] , ∂x P ⎡ ⎛ 2π ⎞⎤ H 1k = ∑ ⎢ rnk sin ωc tk + ωnk tk + θ nk ⎜ cos α nk cos β nk ⎟ ⎥ ⎝ λ ⎠⎦ n =1 ⎣ ( ) ( ) ( ) ⎡ ⎛ 2π t ⎢ − rn sin ωc tk + ωn tk + θ n ⎜ cos β nk k k k ⎜ λ xk2 + yk2 H 2k = ∑ ⎢ ⎝ n =1 ⎢ ⎢ x cos γ − α + y sin γ − α k nk k k nk ⎣ k P ⎡ ⎛ 2π H 3k = ∑ ⎢ rnk sin ωc tk + ωnk tk + θ nk ⎜ sin α nk cos β nk ⎝ λ n =1 ⎣ ( P The performance of this location estimation algorithm is discussed through numerical results and compared to the following algorithms in Section VII. k −1 ( ( ) ) ( ( )) ) ( ( ) ( ) ( ) ⎠⎥ ⎥ ⎦ ⎞⎤ ⎟⎥ ⎠⎦ ⎞⎤ cos β nk ⎟ ⎥ 2 2 ⎟⎥ ⎝ λ xk + yk ⎠⎥ ⎥ γ k − α nk ⎦ ⎛ 2π t ) ⎜⎜ ( ⎞⎤ ( ) ⎟⎟ ⎥⎥ ( ⎡ ⎢ − rn sin ωc tk + ωn tk + θ n k k k H 4k = ∑ ⎢ ⎢ n =1 ⎢ y cos γ − α − x sin k nk k ⎣ k P (17) ( ) )) K is the gain matrix, P̂ is the estimation error covariance, and γ k = arctan ( yk / xk ) . The notation x k denotes the a priori state estimate at time step k and xˆ k the a posteriori state estimate given measurement zk. P and Pˆ are defined k k similarly. As in any nonlinear estimation problem, the convergence of the EKF to the true value of the location depends on the initial parameter value; therefore we first develop the MLE method to obtain an initial estimator of adequate accuracy for the EKF. This hybrid algorithm, as numerical results indicate, has improved accuracy for the final MS location estimate. The EKF described above utilizes the first term in a Taylor expansion of the nonlinear measurement model in (7). It always approximates p (x k | Z k ) by a Gaussian distribution. However, if the true density is non-Gaussian, then a Gaussian model may not describe it precisely. In such cases PFs yield an improvement in performance in comparison to that of an EKF. The design of the PF is discussed in the next section. VT-2005-00647.R2 5 V. THE PF APPROACH FOR MS LOCATION AND VELOCITY ESTIMATION The PF is a technique for implementing a recursive Bayesian filter by Monte Carlo simulations. The key idea is to represent the required posterior density function by a set of random samples {ω ( j )} k N j =1 {xˆ ( j )} k N j =1 with associated weights N p (x k | Z k ) ≈ ∑ ωk ( j ) δ ( x k − xˆ k ( j ) ) (18) j =1 We therefore have a discrete weighted approximation to the true posterior p (x k | Z k ) . The weights are chosen using the principle of importance sampling [26]: ωk ( j ) ∝ p ( z k | x k ( j ) ) p ( x k ( j ) | x k −1 ( j ) ) (19) q ( x k ( j ) | x k −1 ( j ) , z k ) where q ( x k ( j ) | x k −1 ( j ) , z k ) is the importance proposal distribution function that generates the samples {xˆ ( j )} k N j =1 . The choice of this distribution function is one of the most critical design issues and determines the type of the PF. The optimal proposal distribution function that minimizes the variance of the weights conditioned on x k −1 ( j ) and z k is q ( x k | x k −1 ( j ) , z k )opt = p ( x k | x k −1 ( j ) , z k ) [26]. However, analytical evaluation of the optimal proposal function is not possible for many models, and thus has to be approximated using local linearization [26] or the unscented transformation [14]. In this paper, the unscented transformation method is considered and the resulting filter is called the unscented particle filter (UPF) that is described in Section VI. Nonetheless, the most popular choice of proposal function is the transition prior q ( x k | x k −1 ( j ) , z k ) = p ( x k | x k −1 ( j ) ) . This filter is called the generic PF and is discussed herein. Although this choice of proposal function results in higher Monte Carlo variations than the optimal, it is usually simple to implement. The time-update stage of the generic PF [27] is performed by passing the random samples {xˆ ( j )} k −1 N j =1 through the system model (6) to obtain the time-updated samples k by: N j =1 which yields: ωk ( j ) = and to compute estimates based on these samples and weights. In this case the posterior density at time k can be approximated as: {x ( j )} the time updated PDF p (x k | Z k −1 ) . The measurement-update stage can be described by substituting the choice of proposal distribution q ( x k | x k −1 ( j ) , zk ) = p ( x k | x k −1 ( j ) ) into (19) and normalizing . Namely, the time-updated samples are obtained x k ( j ) = f ( xˆ k −1 ( j ) , w k −1 ( j ) ) (20) p ( zk | x k ( j ) ) ∑ N j =1 (21) p ( zk | x k ( j ) ) We define a discrete density over {x ( j )} k N j =1 with probability mass ωk ( j ) associated with each sample x k ( j ) . Then we get the measurement-update samples {xˆ ( j )} k N j =1 through a resampling process, such that Pr {xˆ k ( i ) = x k ( j )} = ωk ( j ) for any i. Several resampling schemes are presented in the literature such as: systematic [28], stratified, and residual resampling [29]. However, the specific choice of resampling scheme does not significantly affect the performance of the PF. Therefore, systematic resampling is used in all of the experiments in Section VII since it is simple to implement. The estimate of the PF at time k is chosen to be the mean of the samples {xˆ k ( j )} j =1 . N In the next section, an approximate version of the optimal proposal distribution is considered in order to have a more accurate MS location estimate. VI. THE UPF APPROACH FOR MS LOCATION AND VELOCITY ESTIMATION S The UPF results from using a scaled unscented transformation (SUT) method to approximate the optimal proposal distribution within a particle filter framework. The SUT provides more accurate approximation than linearization methods [14]. In particular, the SUT calculates the posterior covariance accurately to the 3rd order, whereas linearization methods such as the EKF rely on a first order biased approximation. The SUT method is introduced next. A. The SUT Method The SUT method still approximates the proposal distribution by a Gaussian distribution, but it is specified using a minimal set of deterministically chosen sample points. These sample points completely capture the true mean and covariance of the Gaussian distribution, and when propagated through the true nonlinear system, captures the posterior mean and covariance accurately to the 3rd order for any nonlinearity. Consider the state equation described in (5). For simplicity, let x k = f (x k −1 ) , where x k −1 an nx dimensional random vector and assume x k −1 has mean xk −1 and covariance Pk −1 . Then, a set of 2nx + 1 weighted samples or sigma points where w k −1 ( j ) is a sample drawn from the PDF p (w k −1 ) of Si = {Wi , X i } the system noise. The samples {x k ( j )} j =1 are distributed as completely capture the true mean and covariance of the prior N are deterministically chosen so that they VT-2005-00647.R2 6 x0 ( j ) = E ⎡⎣ x 0 ( j ) ⎤⎦ random vector x k −1 . A selection scheme that satisfies this requirement is [14]: X k0−1 = xk −1 X i k −1 X ki−1 = xk −1 ( m) W0 ( −( = xk −1 + ( nx + λ ) Pk −1 )i , ) T P0 ( j ) = E ⎡( x 0 ( j ) − x0 ( j ) ) ( x 0 ( j ) − x0 ( j ) ) ⎤ ⎣ ⎦ x0a ( j ) = E ⎡⎣ x 0a ( j ) ⎤⎦ = ⎡( x0 ( j ) ) ⎣ T i = 1,… , nx ( nx + λ ) Pk −1 , i = nx + 1,… , 2nx = λ / ( nx + λ ) W0( c ) = λ / ( nx + λ ) + (1 − α 2 + β ) Wi ( m ) = Wi ( c ) = 1/ {2 ( nx + λ )} , where λ =α parameters, ( 2 ( nx + κ ) − nx , ( nx + λ ) Pk −1 )i i = 1,… , 2nx α , β , and κ are scaling is the ith row or column of the propagated through the nonlinear function X ki = f (X ki−1 ), i = 0,… , 2nx . And the estimated mean and covariance of x k are computed as follows: i =0 i =0 {X i k (23) − xk }{X ki − xk } T ( na + λ ) Pka−1 ( j ) ⎤⎦ (26) • Performing the time update stage as: i =0 2 na { B. The UPF Design The UPF uses the same framework as the regular PF, except that it approximates the optimal proposal distribution by a Gaussian distribution using the SUT method. In particular, the SUT is used to generate and propagate a Gaussian proposal distribution for each particle to get: q ( x k ( j ) | x k −1 ( j ) , z k )opt ≈ N ( xk ( j ) , Pk ( j ) ) (24) and j = 1,… , N . That is, at time k − 1 the SUT is used with the new data, to compute the mean and covariance of the importance distribution for each particle. Next, the jth particle is sampled from this distribution. The description of the UPF approach in this section is mainly based on [14]. In the implementation of the UPF, the augmented state vector is defined as the concatenation of the T original state and noise variables as x ak = ⎡⎣ xTk wTk vk ⎤⎦ . Then the SUT sigma point selection scheme is applied to this new augmented state vector to calculate the corresponding sigma matrix, X ka . The complete UPF is described as follows [14]: 1. Initialization ( k = 0 ) : Draw the particles {x0 ( j )} j =1 from N }{ } Pk ( j ) = ∑ Wi ( c ) X i ,xk ( j ) − xk ( j ) X i ,xk ( j ) − xk ( j ) i =0 These estimates of the mean and covariance are accurate to the 3rd order for any nonlinear function. In comparison, the EKF only calculates the posterior mean and covariance accurately to the first order with all higher order moments truncated the prior p (x0 ) and set: • Calculating sigma points: X ka−1 ( j ) = ⎡ xka−1 ( j ) xka−1 ( j ) ± ⎣ X kx ( j ) = f (X kx−1 ( j ) ,X kv−1 ( j ) ) , xk ( j ) = ∑ Wi ( m )X i ,xk ( j ) xk = ∑ Wi ( m )X ki c) 2. Now for k = 1, 2,… , the importance sampling step is performed by the following steps: 2 na 2 nx Pk = ∑ Wi ( (25) where E[.] is the expectation operator. matrix square root of ( nx + λ ) Pk −1 , Each sigma point is now 2 nx T T P0a ( j ) = E ⎡⎢( x 0a ( j ) − x0a ( j ) ) ( x 0a ( j ) − x0a ( j ) ) ⎤⎥ ⎣ ⎦ 0⎤ ⎡ P0 ( j ) 0 ⎢ ⎥ 0⎥ =⎢ 0 Q ⎢ 0 0 R ⎥⎦ ⎣ (22) i 0 0⎤ ⎦ T (27) 2 na Zk ( j ) = h (X kx ( j ) ,X kn−1 ( j ) ) , zk ( j ) = ∑ Wi ( m ) Zi , k ( j ) i =0 • Performing the measurement update stage as: 2 na Pz k zk = ∑ Wi ( c ) {Zi , k ( j ) − zk ( j )} {Zi , k ( j ) − zk ( j )} T i =0 2 na Pxk z k = ∑ Wi ( c) i =0 K k = Pxk z k Pz−k1z k {X ( j ) − x ( j )}{Z ( j ) − z ( j )} , x ( j ) = x ( j ) + K ( z − z ( j )) T x i, k k k k Pk ( j ) = Pk ( j ) − K k Pz k z k K and then i, k k k k k T k sampling ( (28) ) xk ( j ) from q ( x k ( j ) | x k −1 ( j ) , zk ) = N xk ( j ) , Pk ( j ) . • Evaluating the importance weights as: p ( zk | x k ( j ) ) p ( x k ( j ) | x k −1 ( j ) ) ωk ( j ) ∝ q ( x k ( j ) | x k −1 ( j ) , zk ) (29) and then normalizing the importance weights for j = 1,… , N . 3. Finally, a resampling process such as systematic resampling is performed to obtain N random particles xˆ k ( j ) , Pˆ k ( j ) , and the output is generated in the same ( ) manner as for the generic PF. In the next section, numerical examples are presented to illustrate the accuracy of the proposed algorithms. VT-2005-00647.R2 7 RMSE ( k ) = 1 MC i ∑ xˆ k − xktrue MC i =1 ( ) ( xˆ T i k − x ktrue ) (30) where MC is the number of Monte Carlo simulations performed, and xˆ ik is the filter position estimate ( x, y )T (or velocity estimate ( x, y )T ), at time k in Monte Carlo run i. The overall RMSE is defined as: RMSE = 1 L 1 MC i ∑ ∑ xˆ k − xktrue L k =1 MC i =1 ( ) ( xˆ T i k −x true k ) (31) where L is the total number of simulation time steps after the convergence of the filter. Figure (2a) and (2b) show one realization illustrating the convergence of the proposed algorithms to the real position and velocity of a moving MS, respectively. Figure (3) shows X (meters) 3000 3185 3180 2800 3175 2600 3170 35 2400 2400 Y (meters) [30], cell radii is 5000 m, number of samples S is 10, number of BSs for triangulation is 5, radio-frequency is 900MHz, and 100 Monte Carlo simulations were performed. Next, we consider the simulation setup for the EKF, the PF, and the UPF approaches that employ Aulin’s channel model for MS location and velocity estimation. The simulation setup for the MLE approach remains the same, only now we are trying to locate a single MS. The envelope of the received signal for all paths, rn’s, are generated as Rayleigh iid RVs with parameter 0.5. an , β n , and φn are generated as uniform iid RVs in [0, 2π], [0, 0.2π], and [0, 2π], respectively. The total number of paths P is 6 (represents urban environment). The filters have the following parameters: Number of time steps (measurements) is 50 with Δ k = 0.1 seconds, process noise covariance Q and measurement noise variance R are I 2× 2 and 0.01, respectively, where I 2× 2 is the twodimensional identity matrix, the initial PDF of the MS position is assumed to be uniform over the entire cell size which represents the worst-case as far as choosing an initial PDF is considered, the initial PDF of the MS velocity is Gaussian distributed with mean 65 meters/sec and variance 10, number of particles is 500. The SUT parameters are set to α = 1, β = 0, and κ = 0, and finally the mean estimate of all particles is used as the final estimate. The position (or velocity) root mean square error (RMSE) is used as a performance measure and is defined as: 3200 36 37 2200 2000 2140 1800 1400 Actual PF EKF EKF/MLE UPF 2130 1600 2120 25 0 26 10 27 20 28 29 30 40 50 Filter Steps (a) Vx (meters/second) variance σ b2 is 8 dB, reference distance d 0b is 200 m for all b 3400 Vy (meters/second) In this numerical example, the performance of the proposed MS location and velocity estimation algorithms is determined. We consider first the ML estimate of the MS location in which we employ a typical, yet realistic, wireless communication simulation setup. The service area consists of a 19-cell cluster. The BSs are placed over a uniform hexagonal pattern of cells which are centrally equipped with omni-directional antennas. MSs are placed randomly in the central cell and the number of arranged users is 1000. Path-loss exponent ε b is 3.5, path-loss the position and velocity RMSE for each time according to (30), respectively, and the overall position and velocity RMSE for the convergent runs using (31) are shown in Table (1). 200 55 150 48 42 100 47 50 0 150 51 Actual PF EKF EKF/MLE UPF 100 47 33 34 35 50 0 0 10 20 30 40 50 Filter Steps (b) Fig. 2. (a) Location and (b) velocity estimates of a moving MS generated by the different filters. Velocity RMSE (m/s) Location RMSE (m) VII. NUMERICAL RESULTS 300 EKF PF MLE EKF/MLE UPF 200 15 0 40 100 0 5 10 15 20 25 30 35 43 40 45 50 150 10 100 0 40 45 50 0 5 10 15 20 25 30 35 40 45 50 Filter Steps Fig. 3. Location and velocity estimates RMSE (k) generated by the different algorithms. Table 1. Performance comparison for MS location and velocity estimation algorithms using the MLE, EKF, EKF/MLE, PF, and the UPF approaches. MLE EKF EKF/ PF UPF MLE _ 39 6 2 2 Diverged runs 73.46 142.38 11.23 4.31 3.81 Position RMSE (m) _ 51.36 16.52 1.01 0.96 Velocity RMSE (m/sec) VT-2005-00647.R2 8 From Figure (3) and Table (1), it can be noticed that the accuracy of the MLE approach is satisfactory. However, in realistic NLOS and multipath conditions this method does not perform well. Nevertheless, it can be used as an initial condition for the EKF to find a more accurate estimator. It has been also observed that the accuracy increases as the number of samples, S, increases and σ b2 , ε b decrease, as expected. For more accurate estimates Aulin’s channel model is employed together with the EKF, PF, and UPF. We observe in Figure (2) that the EKF/MLE, PF, and UPF estimators converge to the actual location and velocity within a few iterations (less than 5). While the EKF position and velocity estimates oscillate with large deviation around the actual position and velocity. This is because the EKF truncates higher order series expansion terms and is sensitive to the initial state. However, the latter can be improved by using the ML estimate as an initial estimate for the EKF. Since it takes less than 5 iterations for the filters to converge near the actual value as shown in Figures (2), the RMSE (k) in (30) is calculated starting from the iteration k = 5. Only convergent runs are used in the RMSE calculations. Figure (3) shows that the performance of the PF and the UPF approaches are about the same and superior to other approaches. The superior performance of the UPF is clearly evident. Table (1) shows the number of runs that diverged and the performance for each approach. The latter shows the appropriateness of choosing the PF and the UPF for this kind of problems. We have observed that using fewer particles does not affect significantly the unscented particle filter, while the performance of the particle filter deteriorates. The high accuracy is due to the appropriateness of Aulin’s channel model and the efficiency of the particle filtering in this particular application. Figure (4) shows how robust the particle filtering approach is if we assume that we only know the channel parameters {rn , α n , β n } within certain tolerances. Specifically, rn = rn0 (1 + δ rn0 ), δ rn ≤ 5%, 10%, 20% and 30% 0 α n = α n (1 + δα n ), δα n ≤ 5%, 10%, 20% and 30% 0 0 0 β n = β n (1 + δβ n ), Velocity RMSE (m/sec) Location RMSE (m) 0 δβ n ≤ 5%, 10%, 20% and 30% 0 0 25 20 15 10 5 0 5 10 15 20 25 30 35 40 25 45 50 Exact 5% Error 10% Error 20% Error 30% Error 20 15 (32) where rn0 , α n0 and β n0 are the nominal (actual) values of the channel parameters. Figure (4) is generated by assuming that the real channel has parameters rn0 , α n0 and β n0 , while in the estimation stage the channel model parameters used are uniformly distributed about their nominal values as in the uncertainty model (32), and varying the uncertainty percentage from 5% to 30%. It can be noticed that the location and velocity RMSE still converge even if the channel parameters have errors. The higher the error is, the longer time it takes for the filter to converge. It can also be seen that the final RMSE increases for higher errors in channel parameters as expected. The high accuracy, consistency and performance of the proposed UPF approach, makes it suitable to be used in any location and velocity estimation applications, particularly those which require high accuracy such as emergency services. VIII. CONCLUSION New estimation methods are proposed to track the position and velocity of a MS in a cellular network. They are based on Aulin’s scattering model combined with the EKF, PF, and UPF estimation algorithms. Since the instantaneous electric field is a nonlinear function of the MS location and velocity, the EKF, PF, and UPF are appropriate for the estimation process. Numerical results for typical simulations including in the presence of parameters uncertainty show that they are highly accurate and consistent. The performance of the PF and the UPF estimation methods are superior to the EKF. This is due to the sensitivity of the EKF to the initial condition and Gaussian assumptions. An alternative is to use the ML estimate that employs the lognormal channel model, as the initial EKF state. The use of nonlinear models and/or nonGaussian noise is the main explanation for the improvement in accuracy over linear algorithms such as the EKF. These methods also excel in using inherent features of the cellular system, i.e., they support existing network infrastructure and channel signaling. The assumptions are knowledge of the channel and access to the instantaneous received field, which are obtained through channel sounding samples from the receiver circuitry. Future work will focus on generating efficient channel estimation algorithms, to remove the assumption on partial knowledge of the channel. Work on building a pilot application to test the performance of the PF and/or the UPF in realistic conditions is on-going together with the incorporation of channel model parameters estimation algorithms. 10 5 0 5 10 15 20 25 30 35 40 45 REFERENCES 50 Filter Steps Fig. 4. The UPF location and velocity estimates RMSE (k) for imperfect knowledge of channel parameters. [1] FCC Docket No.96-264, “Revision of the Commission Rule to ensure compatibility with Enhanced 911 emergency calling system,” FCC Reports and Orders, 1996. VT-2005-00647.R2 [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20] [21] [22] [23] [24] [25] [26] [27] [28] [29] [30] J. Reed, K. Krizman, B. Woerner, and T. Rappaport, “An overview of the challenges and progress in meeting the E-911 requirement for location service,” IEEE Commun. Mag., vol. 36, pp. 30-37, Apr. 1998. I. Papageorgiou, “Mobile receiver localization with an enhanced received signal level,” M.S. thesis, ECE Department, University of Cyprus, May 2006. Virtuele Haven Consortium, “Location based services, services and technologies,” pp. 9-21, 2002. CELLO Consortium, “Cellular location technology,” pp. 1-21, 2001. 3GPP TS 03.71 V8.8.0 (2004-03), pp. 101-110, 2004. 3GPP TS 25.305 V6.1.0 (2004-06). Symmetricom, “Location of mobile handsets – the role of synchronization and location monitoring units,” White paper, 2002. M. Hellebrandt, R. Mathar and M. Scheibenbogen, “Estimating position and velocity of mobiles in a cellular radio network,” IEEE Trans. Veh. Technol., vol. 46, no. 1, pp. 65-71, Feb. 1997. M. Hellebrandt and M. Scheibenbogen, “Location tracking of mobiles in cellular radio networks,” IEEE Trans. Veh. Technol., vol. 48, no. 5, pp. 1558-1562, Sep. 1999. P. Zehna, “Invariance of Maximum Likelihood Estimators,” Annals of Mathematical Statistics, 37, 744, 1966. G. Bishop and G. Welch, An Introduction to the Kalman Filter. University of North Carolina, 2001. M. Arulampalam, S. Maskell, N. Gordon, and T. Clapp, “A tutorial on particle filters for online nonlinear/non-Gassian Bayesian tracking,” IEEE Transactions on Signal Processing, vol. 50, no. 2, Feb. 2002. R. van der Merwe, A. Doucet, J. de Freitas, and E. Wan, “The unscented particle filter,” Advances in Neural Information Processing Systems, Dec. 2000. T. Aulin, “A modified model for fading signal at a mobile radio channel,” IEEE Trans. Veh. Technol., vol. 28, no. 3, pp. 182-203, 1979. I. Papageorgiou, C. Charalambous, and C. Panayiotou, “An enhanced received signal level cellular location determination method via maximum likelihood and Kalman filtering,” in Proc. of the WCNC, vol. 4, pp. 2524-2529, 2005. F. Gustafsson, F. Gunnarsson, 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, Feb. 2002. H. Jwa, S. Kim, S. Cho, and J. Chun, “Position tracking of mobiles in a cellular radio network using the constrained bootstrap filter,” in Proc. Nat. Aerosp. Electron. Conf., Dayton, OH, Oct. 2000. R. Karlsson and N. Bergman, “Auxiliary particle filters for tracking a maneuvering target,” in Proc. of the 39th IEEE Conference on Decision and Control, pp. 3891-3895, Sydney, Australia, Dec. 2000. R. Karlsson and F. Gustafsson, “Range estimation using angle-only target tracking with particle filters,” in Proc. of the American Control Conference, pp. 3743-3748, Arlington, VA, Jun. 2001. T.S. Rappaport, Wireless Communications: Principles and Practice. 2nd Edition, Prentice Hall, 2002. 3GPP TS 05.08 V8.19.0 (2003-11), pp. 15-23 and pp. 28-41, 2004. B. Sklar, Digital Communications: Fundamentals and Applications. 2nd Edition, Prentice Hall, 2001. T. Kailath, Lectures on Linear Least-Squares Estimation. Springer, New York, 1976. C. Wong, M. Lee, R. Chan, “GSM-Based Mobile Positioning Using WAP,” in Proc. of the WCNC, vol. 2, pp. 874-878, Sep. 2000. A. Doucet, S. Godsill and C. Andrieu, “On sequential Monte Carlo sampling methods for Bayesian filtering,” Statistics and Computing, vol. 10, pp. 197-208, 2000. N. Gordon, D. Salmond, and A. Smith, “Novel approach to nonlinear/non-Gaussian Bayesian state estimation,” IEE Proc.-F, vol. 140, no. 2, pp. 107-113, 1993. G. Kitagawa, “Monte Carlo filter and smoother for non-Gaussian nonlinear state space models,” J. Comput. Graph. Statist., vol. 5, no. 1, pp. 1-25, 1996. J. Liu and R. Chen, “Sequential Monte Carlo methods for dynamical systems,” J. Amer. Statist. Assic., vol. 93, pp. 1032-1044, 1998. ETSI TR 101 115 V8.2.0 (2000-04), Annex V.A. 9