Through-Wall Tracking with Radio Tomography Networks Using Foreground Detection

Transcription

1 IEEE Wireless Communications and Networking Conference: Services, Applications, and Business Through-Wall Tracking with Radio Tomography Networks Using Foreground Detection Yi Zheng and Aidong Men Multimedia Technology Center, Beijing University of Posts and Telecommunications Beijing, China {yizheng, Abstract This paper presents a novel method for tracking a moving person or object through walls using wireless networks. The method takes advantage of the motion-induced variation of received signal strength (RSS) measurements in a radio tomography network. Based on real measurements of a deployed network, we show that the RSS distribution on a wireless link can be modeled as a mixture of Gaussians. An online learning algorithm is then proposed to update the model and detect whether the link is affected by the motion. Using spatial locations of the affected links, we apply the sequential Monte Carlo (SMC) methods to track the coordinates of a moving target. Experimental results show that the proposed method achieves high tracking accuracy in time-varying environment without the need for offline training. I. INTRODUCTION Through-wall tracking has emerged as an attractive technology in recent years due to the increasing demand of monitoring human activities in non-line-of-sight (NLOS) environments [] [5]. This technology has wide applications in the areas of security, surveillance, search-and-rescue, and military. For example, a through-wall tracking system can help emergency responders to locate survivors during disaster situations, such as fires and earthquakes. Similarly, showing the positions of people behind walls can increase the chance of successful military operations with minimum casualties. To make these applications practical, the person s position should be estimated when no tag or device is attached. Furthermore, rapid deployment is a critical feature required by the tracking system, since one can not expect to have sufficient time in emergency situations. Radio tomography network [] is an effective technique for through-wall tracking. When an object moves into the network area, wireless links near that object will experience variations in received signal strength (RSS) measurements. These variations provide position information about the object, allowing the network to track motion in real-time. Compared to radar-based through-wall imaging (TWI) systems [], [], radio tomography uses low-cost, low-power radio-frequency (RF) sensor nodes which can be readily deployed around the area of interest. Several existing methods can be used to track motion through walls with radio tomography networks. However, all these methods have practical limitations when applied to emergency situations. Fingerprint-based method needs a manual training process which will take a significant amount of time [4]. Variance-based radio tomographic imaging (VRTI) works only in heavily obstructed areas []. Fade-level-based method requires a calibration period during which the network area is vacant of people [5]. Generally speaking, all these methods rely on the multipath environment of the network area. When the environment changes, the systems may fail to perform tracking, or must be trained offline. As a result, these methods are impractical for emergency applications. In this paper, a foreground detection method is proposed for through-wall tracking using radio tomography networks. The method allows tracking the coordinates of a moving person or object in different NLOS environments without changing system parameters. Furthermore, offline training is not required by the proposed method, so the tracking system can be deployed rapidly. The training, in essence, is performed online in a mixture learning algorithm. Therefore, the throughwall tracking system is able to track motion even when the environment is time-varying. The rest of this paper is organized as follows. Section II lists and evaluates the existing through-wall tracking methods. In Section III, we show that the RSS distribution on each wireless link can be modeled as a mixture of Gaussians. An online learning algorithm is then proposed to update the model and detect whether the link is affected by the motion. In Section IV, the sequential Monte Carlo (SMC) methods are applied to perform tracking and remove noise. Section V presents the testbed setup, the experimental results for the proposed tracking system, and a discussion of system parameters. Finally, conclusions are drawn in Section VI. II. RELATED WORK Various sensor techniques can be used for the purpose of through-wall tracking. The most common choice is to use radar-based TWI systems, such as ultra-wideband (UWB) radar [] and Doppler radar []. Since the cost of radar systems is relatively high, commercial products based on this technology are prohibitively expensive for most applications. In addition, these systems suffer from accuracy and noise issues at long range due to monostatic radar losses. In comparison, in this paper, we discuss radio tomography, which uses the RSS measurements to detect and locate objects through walls. Some initial attempts [4] demonstrate that motion tracking can be formulated as a fingerprint-matching problem. A radio map is constructed manually during the //$. IEEE 78

2 (a) (b) Probability Density.. Measured RSS Mixture Model Background Foreground (c) Probability Density... Measured RSS Mixture Model Background Foreground (d) Fig.. Comparison of RSS measurements and distributions on two wireless links: (a) measurements on the NLOS link; (b) measurements on the LOS link; (c) distribution on the NLOS link; (d) distribution on the LOS link. offline phase by recording RSS with a person standing at known locations. During the online phase, the person s position is estimated by comparing current RSS with training data in the map. In many situations, however, manual training is not possible since it can take a significant amount of time. An alternative scheme for radio tomography is to use measurement models to estimate the coordinates of the target. The measurement model describes the relation between RSS value and position of the target based on radio-propagation characteristics. Since the motion-induced fading phenomenon is determined by the multipath environment of the network area, this relation varies as the environment changes. As a result, existing model-based through-wall tracking systems either work in a specific environment [], or must be trained by RSS measurements while the surveillance area is vacant [5]. Thus the model-based methods are impractical for emergency applications such as hostage rescue and building fire. III. FOREGROUND DETECTION In this section, we discuss modeling the RSS measurements on each wireless link as a mixture of Gaussians and using an online learning algorithm to update the model. The Gaussian distributions of the model are then evaluated to determine whether or not the link is affected by the motion. A. RSS Distribution When RF sensor nodes communicate, radio waves travel through the physical area of the network. People within that area will absorb, reflect, diffract, or scatter some of the waves, causing changes in RSS measurements on wireless links. To track moving objects through walls using a radio tomography network, it is first necessary to examine the motion-induced variation of RSS measurements in NLOS environments. Past studies [5] have shown that, when a wireless link is line-of-sight (LOS), a person crossing the direct path will cause a drop in signal strength due to the shadowing effect of the human body; when the environment is rich in multipath and heavily obstructed, the link is more likely to experience a high variance of RSS when the person shadows the direct path. These fading phenomena represent two extreme situations; that is, when a wireless link is neither heavily obstructed nor LOS, both drops and variances may occur on the link. To illustrate these fading phenomena from real measurements, we set up two wireless links with equivalent distance of 4 m. One link is obstructed by two dense walls and the other is LOS. For each link, a person moves across the direct path repeatedly within minutes and about 5 measurements are collected. Part of the measurements are plotted in Fig. (a), (b), while the RSS distributions are shown in Fig. (c), (d). We see that the experiment shows similar results to [5]. The measured RSS is confined within a range of - dbm when the link is not affected by motion, and when the person moves near the link, RSS varies over a range of to dbm. We define these two types of variations as background and foreground. Then the measurements on a link can be modeled as a two-state Markov chain [6] which alternates between a background and a foreground process. Figure (c), (d) show that, over all time, measurements show a mixture of one high-variance and one low-variance distribution. The RSS data can be modeled as a mixture of two Ricean distributions as in [7], or a mixture of skew-laplace distributions as in [5]. Considering a trade-off between model accuracy and computational complexity, we simply model the RSS distribution on each wireless link as a mixture of Gaussians, which can be expressed in the following form: P(s) = K w i g(s;µ i,σ i ) () i= where s is the RSS measurement in dbm, P(s) is the probability density function of the random variable s, K = is the number of Gaussian distributions, w i is the weight of 79

3 the ith Gaussian G i and K i= w i = ; µ i, σ i represent the mean and standard deviation of G i respectively, and g is a Gaussian probability density function: g(s;µ,σ) = πσ exp ] [ (s µ) σ From Eq. (), we may see that a new measurement will, in general, be represented by one of the Gaussian components in the mixture model. Therefore, to detect whether current measurement results from a foreground process, it is essential to know the values of the model parameters. B. Learning Algorithm This section presents an online learning algorithm to update the parameters of the Gaussian mixture model. The basic concept for various mixture learning algorithms can be understood in terms of the recursive filter formulation with a learning rate schedule [8], [9]: θ(t) = ( η(t)) θ(t )+η(t) (s(t);θ(t )) () where the model at time t, θ(t), is updated by a local estimate (s(t); θ(t )) at a rate controlled by η(t). If η(t) = /t [], the parameter estimates quickly approach the expected value at the initial time steps, but converge to the local-optimal estimation on stationary distributions only. If η(t) = α [8], current parameter estimates reflect the most recent observations within a /α window; however, it takes relatively long time for the parameters to converge. In a through-wall tracking system, the RSS distribution is usually time-varying due to environment changes such as doors opening or moved furniture, so we need a solution that combines fast convergence and temporal adaptability. The basic algorithm in this paper follows the formulation by Lee [9]. A variable c k is introduced to count the number of effective observations for Gaussian G k and calculate the learning rate η k. The value of c k is incremented when parameters of G k are updated, and reset to when G k is reassigned. This variable will improve both convergence speed and learning accuracy. Algorithm shows the proposed online mixture learning algorithm in pseudocode, where V is an initial standard deviation and T σ is a predetermined threshold. For simplicity, the notation is based on a single wireless link. Extension to a higher dimension is straight-forward. Compared to the conventional Gaussian mixture learning algorithms for background subtraction [] or object tracking [] in video sequences, the proposed online learning algorithm has several modifications. The most important one is the selection of Gaussian components for parameter update at each time step. Usually in video applications, one measurement only updates a single best-matching Gaussian component. However, when the distribution consists of two overlapping components, this approach can lead to starvation where one Gaussian stretches with increasingly more weight to overdominate the others. To solve this problem, in this paper, all Gaussians that match current measurement are updated by a () Algorithm : Online mixture learning algorithm Control variables: K, V, α, T σ Initialization: i=...k,w i =,µ i = inf,σ i = V,c i = while new RSS measurement s t do w i g(s;µ i,σ i) if s µi < T σ i=...k,p i = σ i otherwise if K i=pi > then // at least one match is found for (k=; k<k; k++) do q k = p k / K i= pi // expected posterior of G k w k (t) = ( α) w k (t )+α q k if q k > then // for matched ( Gaussians ) α c k = c k +q k, η k = q k +α c k µ k (t) = ( η k ) µ k (t )+η k s σk(t) = ( η k ) σk(t )+η k (s µ k (t )) else // no match is found i=..k, w i(t) = ( α) w i(t ) k = argmin i{w i/σ i},w k = α,µ k = s,σ k = V,c k = normalize w variableq k proportional to their estimated posterior probability P(G k s): P(G k s) = P(G k)p(s G k ) K i= P(G i)p(s G i ) = p k K i= p = q k (5) i where s is the RSS measurement at time t; p k can be obtained from Eq. (4) and represents the joint probability of s and G k. This soft-partition approach [9] improves robustness of the mixture learning algorithm for links which are heavily obstructed by walls or other obstacles. C. Detection Strategy As the parameters of the Gaussian mixture model are constantly updated by new RSS measurements, we would like to determine whether or not the current measurement is produced by a foreground process; that is, we need to detect the wireless links affected by the motion at each time step, since the target s position can be estimated from spatial locations of the affected links. First, foreground detection involves incorporating domain knowledge of the foreground process into the estimation of P(F G k ), where F is the set of foreground components. In contrast, previous works in mixture learning are normally interested in the background components in pixels, P(B G k ), where B represents the background class and P(B G k ) = P(F G k ). In [8], P(B G k ) equals to for Gaussians with relatively high w/σ and for others. In [9], P(B G k ) is approximated by a sigmoid function on w/σ using logistic regression. The value of w/σ increases as a distribution gains more evidence and as the variance decreases, so w/σ is an (4) 8

4 appropriate variable for modeling the foreground processes as well. Therefore, we define P(F G k ) as an exponential function on w k /σ k, which can be expressed as: P(F G k ) = f(w k /σ k ;a,b) = e a(w k/σ k ) b (6) where w k /σ k [,+ ), f(w k /σ k ;a,b) (,], and parameters a, b control the mapping from w k /σ k to P(F G k ). Then we formulate foreground detection as a classification problem based on P(F s). This posterior probability can be expressed in terms of the mixture components P(G k ) and P(s G k ), and the density estimate P(F G k ) as follows: P(F s) = K P(F G k )P(G k s) (7) k= where P(G k s) is defined in Eq. (5) and P(F G k ) in Eq. (6). The affected links are determined by the mixture models where P(F s) > T s ; T s is a predefined threshold. Since the Gaussian components overlap for most NLOS links, our calculation of P(G k s) may be inaccurate when a foreground measurement falls near the mean of the background distribution. As a result, errors may occur in the estimation of affected links based on P(F s). To eliminate this type of error, an intuitive solution is to average the P(F s) estimates within a certain time interval to obtain the local means, as expressed in the following equation: P l (A s t ) = N s + N s n= N s P(F s t+n ) (8) where N s + represents the length of the interval, A is the set of affected links, and s t is the RSS measurement at time t; the term P l (A s) is obtained from the time averaging process, and defined as the posterior probability of the link l being affected given the RSS measurement s. Then the affected links can be estimated from the mixture models where P l (A s) > T s. We find that the performance of P l (A s) is better than that of P(F s) in real systems. IV. MOTION TRACKING The algorithm described above allows us to detect the affected links in a radio tomography network by updating the Gaussian mixture model on each link. Since the affected links t to cross through space containing moving objects, the coordinates of the person can be tracked from spatial locations of the links. There are many frameworks for motion tracking in radio tomography networks, such as the Kalman filter used in [] and the best-cover algorithm proposed in []. In this section, we introduce a tracking algorithm based on the sequential Monte Carlo (SMC) methods. SMC methods are a class of numerical methods for the solution of optimal estimation problems in non-linear non- Gaussian scenarios. In this paper we provide a brief overview of the SMC-based motion tracking algorithm in radio tomography networks. See our previous work [4] for more details. Let x n be the coordinate of the target at time n. Then the target moves according to the following linear Gaussian dynamics: x n = x n +v n (9) where the process noisev n is zero-mean Gaussian white noise. We are interested in estimating {x n } n, but only have access to the observations {y n } n, which is given by y n = [P :L (A s n )] T = [ψ :L (x n )] T +w n () where for any sequence {z n } n and any i j, z i:j = (z i,z i+,...,z j ); the measurement noise w n is zero-mean Gaussian white noise and L is the number of wireless links. Based on our experiments in previous work [4], the observation model ψ l (x n ) for link l can be defined as follows: ψ l (x n ) = φe ( xn xt + xn xr xt xr )/σ () where x t, x r are the coordinates of RF sensor nodes for link l; φ is the mean of P l (A s) when the link is shadowed and σ controls the rate of decay. With these assumptions and equations, the tracking algorithm can be summarized below in Algorithm, with each step carried out for i =,...,N, where N is the number of particles. Note that µ(x ) is an uniform distribution within the network area, the function f(x n x n ) is based on Eq. (), and the weights of particles are calculated from Eq. (). See [5] for a tutorial introduction and overview of the SMC methods. Algorithm : SMC-based motion tracking algorithm At time n = : Sample X i µ(x ) Compute the weights w (X i ) and set W i w (X i ), N i= Wi = Resample {W i,x i } to obtain N equally-weighted particles { N,Xi } Compute the state of the target X = N N i= Xi At time n : Sample Xn i f(x n x n ) Compute the weights w n(xn :n) i and set Wn i w n(xn :n), i N i= Wi n = Resample {Wn,X i n} i to obtain N equally-weighted particles { N,Xi n} Compute the state of the target X n = N N i= Xi n A. Testbed Setup V. EXPERIMENTAL RESULTS To evaluate the performance of the proposed through-wall tracking system, we choose a radio tomography network that contains 4 RF sensor nodes, as illustrated in Fig.. The network was deployed around a room with four walls, two doors and a window. The walls are constructed of bricks, steel and concrete, while the doors are made of wood and steel. The nodes were fixed on stands to keep them m off 8

5 y coordinate (m) Fig...4m.4m.9m.m 5.7m Wall Door Window Wireless Sensor Node Person Ground Truth Path The system layout of a through-wall tracking experiment..m Probability (a) RSS measurement Mean of Gaussian Mean of Gaussian (b) P(F s) P(A s) the ground, and these stands were placed in a rectangular perimeter. Markers were placed inside the room, so that the person would be able to follow the predetermined ground truth path. The nodes utilize the IEEE protocol, and transmit in the.4ghz frequency band. A token-ring protocol is developed to avoid transmission collisions. Data packet broadcast by each node contains the node ID, the time of transmission, and the measured RSS. A base-station node that receives all broadcasts is used to gather RSS measurements and save it to a laptop. The time interval between each transmission is 5 ms, so that the RSS value on each link is updated every ms; note that RSS value on a link is calculated by averaging the measurements on both forward and reverse links. At time n =, a person at coordinate (,.9) starts to walk counter-clockwise along the ground truth path at a constant speed. At time n = 56, the person moves back to the starting point, and meanwhile we open the door at coordinate(.5,.4) to change the NLOS environment. Then, the person continues to move and returns at time n =. RSS measurements are taken while the person moves, and the ground truth is interpolated by using time records and marker positions. B. Results for Foreground Detection We test our online learning algorithm with the RSS measurements on a single link, (.9,) to (.9,.6). Over trials are performed with different parameter settings, and the contrast in performance shows that the values in Table I best fit the measurement data. The corresponding results are plotted in Fig.. We may see that the Gaussian component TABLE I PARAMETERS FOR THE PROPOSED SYSTEM Parameter Value Parameter Value K V α. T σ. a b 5 N s 5 N φ.9 σ.5 Fig.. y coordinate (m) Experimental results for the online learning algorithm. Affected Wireless Link Person Wireless Sensor Node Fig. 4. Affected wireless links when the person is located at (,.9). G, which is updated by foreground measurements before time n = 56, quickly converge to the background process as the NLOS environment changes. On the other hand, the noise in P(F s) estimates is smoothed by the time-averaging process of P(A s). To show the effect of the online learning algorithm on through-wall tracking, let the threshold T s be.8. Then we plot the links where P l (A s) > T s, when the person is located at (,.9), in Fig. 4. It can be seen that the affected links t to cluster around the person rather than only cross through the human body. Obviously, spacial locations of the links can be used to estimate the person s coordinate. C. Results for Motion Tracking In this section, we present experimental results for the proposed through-wall tracking system and discuss effect of system parameters on tracking accuracy. The coordinate of the person is estimated using the SMC-based motion tracking algorithm in Section IV with parameters in Table I. Figure 5 shows these position estimates superimposed on the true tracks over time steps. Figure 6 plots the individual x, y coordinates of the estimates at each time step. We see that the estimated positions successfully agree with the ground truth with a time delay of 5 time steps (.6 s). The root MSE (RMSE) for the tracking results is. m. 8

6 y coordinate (m) Ground Truth Estimated Positions RMSE (m) Parameter φ (a) σ =. σ =.5 σ =.9 Fig Position estimates superimposed on ground truth over time steps. RMSE (m).5..5 φ =.5 φ =.7 φ =.9 y coordinate (m) Ground Truth Estimated Positions Fig. 6. Plots of x and y components of position estimates against time, superimposed on ground truth. The parameters φ and σ in Eq. () determine the observation model in the SMC methods and play an important role in improving tracking accuracy. Figure 7(a) shows the RMSE for three σ values over a range of φ values, while, by contrast, Fig. 7(a) plots the RMSE for three φ values over a range of σ values. We can find that the most accurate tracking is accomplished with φ =.9 and σ =.5. VI. CONCLUSION This paper presents a novel method for tracking a moving person or object through walls using radio tomography networks. An online learning algorithm is proposed to detect links affected by the motion, and the SMC methods are applied to track the coordinates of the target. Experimental results show that the proposed method achieves high tracking accuracy in time-varying environment without the need for offline training. Tracking multiple targets through walls can be a challenging and open topic for future research. ACKNOWLEDGMENT The author would like to thank Xiang Xu and Yuzhu Qin for conducting the through-wall tracking experiments and proposing constructive suggestions. REFERENCES [] J. Wilson and N. Patwari, See-through walls: Motion tracking using variance-based radio tomography networks, Mobile Computing, IEEE Transactions on, vol., no. 5, pp. 6 6, Parameter σ Fig. 7. (b) RMSE for different system parameters. [] L. Ma, Z. Zhang, and X. Tan, A novel through-wall imaging method using ultra wideband pulse system, in Intelligent Information Hiding and Multimedia Signal Processing, 6. IIH-MSP 6. International Conference on. IEEE, 6, pp [] Y. Kim and H. Ling, Through-wall human tracking with multiple doppler sensors using an artificial neural network, Antennas and Propagation, IEEE Transactions on, vol. 57, no. 7, pp. 6, 9. [4] M. Seifeldin and M. Youssef, Nuzzer: A large-scale device-free passive localization system for wireless environments, Arxiv preprint arxiv:98.89, 9. [5] J. Wilson and N. Patwari, A fade level skew-laplace signal strength model for device-free localization with wireless networks, Mobile Computing, IEEE Transactions on,. [6] J. Roberts and J. Abeysinghe, A two-state rician model for predicting indoor wireless communication performance, in Communications, 995. ICC 95 Seattle, Gateway to Globalization, 995 IEEE International Conference on, vol.. IEEE, 995, pp [7] M. Ghaddar, L. Talbi, T. Denidni, and A. Sebak, A conducting cylinder for modeling human body presence in indoor propagation channel, Antennas and Propagation, IEEE Transactions on, vol. 55, no., pp. 99, 7. [8] C. Stauffer and W. Grimson, Adaptive background mixture models for real-time tracking, in Computer Vision and Pattern Recognition, 999. IEEE Computer Society Conference on., vol.. IEEE, 999. [9] D. Lee, Effective gaussian mixture learning for video background subtraction, Pattern Analysis and Machine Intelligence, IEEE Transactions on, vol. 7, no. 5, pp. 87 8, 5. [] M. Sato and S. Ishii, On-line em algorithm for the normalized gaussian network, Neural Computation, vol., no., pp. 47 4,. [] L. Cheng, M. Gong, D. Schuurmans, and T. Caelli, Real-time discriminative background subtraction, Image Processing, IEEE Transactions on, vol., no. 5, pp. 4 44,. [] M. Heikkila and M. Pietikainen, A texture-based method for modeling the background and detecting moving objects, Pattern Analysis and Machine Intelligence, IEEE Transactions on, vol. 8, no. 4, pp , 6. [] D. Zhang, J. Ma, Q. Chen, and L. Ni, An rf-based system for tracking transceiver-free objects, in Pervasive Computing and Communications, 7. PerCom 7. Fifth Annual IEEE International Conference on. IEEE, 7, pp [4] X. Chen, A. Edelstein, Y. Li, M. Coates, M. Rabbat, and A. Men, Sequential monte carlo for simultaneous passive device-free tracking and sensor localization using received signal strength measurements, in Information Processing in Sensor Networks (IPSN), th International Conference on,, pp [5] A. Doucet and A. Johansen, A tutorial on particle filtering and smoothing: fifteen years later, in OXFORD HANDBOOK OF NONLINEAR FILTERING. Citeseer,. 8