Chapter 4 Smart Device Localization 4.1 Introduction Localization algorithm estimates the position, coordinates, or location of a smart device. A simple solution for localization and tracking is to use Global Positioning System (GPS). However, the cost and energy consumption of GPS is high. Additionally, GPS works well in line-of-sight scenarios like open space, not in underground or complex apartments. The applications of localization methods include sensing, monitoring, surveillance, and tracking. In a large agriculture field especially in hilly terrain, it is not possible to monitor farm- ers and animals manually. The location and time stamp are tagged with the sensed data. In another application, a camera can be used to locate elderly people in a smart home, however, it breaches privacy. Thus, the location of the person within a house is computed using opportunistic signals like Wi-Fi, 3 omagnetic, and visible light. If the user is located in the washroom frequently, then something is wrong with the person’s health. In a large hospital equipped with several laboratories and departments, the patients, labs, or departments need to be tracked Further, the location of the intruder is important in order to locate or track and combat them. There is a need to keep track of our personnel and vehicles. ‘This is required during stealth operation and surgical strikes for coordination among own personnel and resources. The location of the robot in a large in- door environment is also needed. The location information helps the robot in igation in a constricted space. This also assumes importance in a multi- robot system so that two robots should not collide. Next, the environmental na sensed data such as humidity, temperature, and pollution-causing particles levels in addition to the location and time stamp are essential. Lastly, pedes- trian tracking and vehicle-to-vehicle communication also received significant attention recently. The chapter presents various distance-based and distance- free localization techniques along with their performance evaluation, 844. 121 DOI: 10.1201/9781003:122 Fundamentals of Internet of Things [Localization Algorithms| Wultilateration [Approximate Point] based in Triangle Test J] | Time-of-Arrival Centroid based J«— based lrime-Difference-of “Arrival based lance Vector Distance-based| Fngle-atAnval [Distance-free Localization oe) Localization Closest Point based |] t—| Received Signal | Strength based | >| [Assumption based| Multidimensional Coordinates |<] Scaling based L__>[ Visual based Inertial based J<— FIGURE 4.1: Classification of Localization Algorithms 4.2 Taxonomy of Localization Algorithms In general, localization methods utilize the knowledge of the locations of a small subset of nodes. The node that is aware of its location is known as an anchor, beacon, source, or landmark. Localization is a one-time process for a static smart device, whereas tracking is the continuous localization of the mobile smart device over time. Localization methods can be classified as distance-free and distance-based methods in the literature as shown in Fig- ure 4.1. Distance-based localization methods utilize the distance between the anchor and unknown node. On the other hand, distance-free methods uti- lize only connectivity information. Distance-based methods are fine-grained localization methods, whereas, distance-free methods are coarse-grained. A localization method can be chosen based on whether we need coarse-grained resolution, say, (+ 5 m) or fine-grained resolution, say, (+ 10 cm). The smart device refers to an unknown node. We are interested in estimating the loca- tion of imknown nodes given the measurements from anchors or other smart devicesSmart Device Localization 123 Antenna \Y <> Tag or Reader Transponder FIGURE 4.2: RFID based system for localization 4.3 Distance-based Localization Methods Distance-based methods are based on multilateration, Time-of-Arvival (TOA), Time-Difference-of- Arrival (TDOA), Angle-of-Arrival (AOA)/Direetion- of-Arrival (DOA), Received Signal Strength (RSS), Multidimensional Scaling (MDS), and visual information as shown in Figure 4.1. Localization algorithms using TOA and TDOA require precise synchronization. AOA and DOA-based algorithms need special hardware to estimate the angle at which the signal arrives at the antenna array. The time of arrival (TOA) method performs lo- calization using the information about the time of arrival of the signal from different anchors. However, TOA requires a reference timestamp to synchro- nize the local clock of the nodes, ‘To overcome this issue of reference offset, a time difference of arrival (TDOA) method is used. TDOA is based on the principle of the time difference of arrival of the anchor signals at a pair of nodes. In this method, one of the anchors is generally taken as a reference node. The TDOA method does not require synchronization between anchors and the unknown node, However, synchronization among different anchors is still needed. The visual-based localization algorithm is generally based on the distance between pixels in a spatio-temporal video. Notably, the algorithm requires a camera and breaches the privacy of the users. Smart device localization using RSS provides a cost-effective solution, ‘The RSS-based approach is also leveraged in Radio Frequency Identification (RFID) based positioning, which is suitable for IoT networks. The RFID based technique uses tag or transponder and reader or interrogator to locate the objects as shown in Figure 4.2. The RFID reader acts as a transceiver that de- tects and receives the RF wave originating from the RFID tag. The energy of RF signals decreases exponentially with the distance. The RFID reader is then connected to computing devices. The computing devices estimate the location of the objects using RSS signals. The RSS-based algorithm is marginally ina curate, however, RSS measurements are easily accessible and low-cost. Hence, an RSS-based localization algorithm is suitable for IoT networks.124 Fundamentals of Internet of Things FIGURE 4.3: Illustration of a simple localization method 4.3.1 Multilateration Based Localization Let the locations of the anchors be Ay = (x1, y1), Az = (x2, yz), and Ag (3,43) as shown in Figure 4.3. The location of the unknown node is denoted by (x,y). The Enclidean distance between Ath anchor and unknown node: (1-2) +(m-yP = qj 4.) This can be further rewritten as Baye — yy + (2? +?) = df ~ x} — yt (4.2) Similarly, we can write for Ay and Ay anchors as ager — yoy + (a? +y") = d3~ 2} —y5 (43) —2urgar — 2ysy + (2? + y?) = B- 23-8 We can write Equations 4.2 and 4.3 in matrix form as 20, —2y, 1 Qxy —2yp 1 (44) 2x3 —2yg3 1 ar This can be succinctly written as AX=B (45) ‘The solution of this system of equations is the location of the unknown node. This is obtained by X=A'B (46)Smart Device Localization 125 Now, assume the number of anchors is greater than three, that is, N > 3. In this case, matrices A and B are rectangular of sizes N x 3 and N x 1, respec tively. In this case, the solution is obtained using the concept of pseudo-invers for a rectangular matrix. The location of the unknown node for rectangular matrices is given by XK =(ATA)1ATB (4.7) ‘This can easily be extended for a three-dimensional coordinates system, 4.3.2 Time-of-Arrival Based Localization In Time-of-Anvival (TOA) based localization, consider the same variables from previous section on multilateration based localization. We know that speed multiplied with time elapsed gives the distance between transmitter and receiver. The difference between theoretical and observed distances is given by 9(%) = v(t = 7) (i-aP + iru, teN (4.8) where X = [z y 7]. is the unknown parameter vector. v and AV denote the speed of signal and and the set of umknown nodes. The first term on right, hand side is the observed distance between the anchor and the smart device based on measurements. On the other hand, theoretical distance is shown using Euclidean distance in second term of right hand side. The error, g, is the difference between theoretical and measured distances. The cost function can be written as the sum of squared error. The cost function is given by G(X) = Yo wie?) (4.9) ie where w is the confidence parameter between 0 and 1. The expression is the function of the unknown node location. The w; is the weight of the ith anchor and the smart device. The smaller the distance, the higher the weight or confidence parameter. This makes sense becanse nearby distance information is more reliable than larger distance. Hence, the cost function is weighted sum of squared error. The location of unknown node can be estimated as X= argmin G(X) (4.10) x The TOA-based method requires synchronization between the unknown node and the anchor. This is because of the utilization of distance information between unknown nodes and anchors. Example of ToA based Localization: The locations of the all the an- 1 nl = (0 0)", feo we)” = (0 1)", [zs w)" =f 9)”, and [x4 ys” = [1 1)". The actual location of the unknown sensor node is chors are assumed to be [x y]” = [0.5 0.5)"126 Fundamentals of Internet of Things 0.4 x axis 0. 5 y axis 0.2 FIGURE 4.4: Sum of squared error versus network dimensions in TOA based localization method; The locations of the anchors are at the corners of the network Solution: Assume that all the sensor nodes have a radio communication range of 1.2 so that the unknown node can hear all the neighboring anchors The Gaussian noise in distance measurement is considered to be N’(0, 10-8). It can be seen that sum of square error is minimum at [0.5 0.5]. The cor- responding minimum value of the cost function is 3.1302x 10-8. That is, the estimated and actual node locations are the same which shows the efficacy of the TOA-based localization method. This is illustrated in Figure 4.4. Note that, the confidence parameter is given by 1 1+ distance wi = (4.11) is chosen to be inversely proportional to the distance between the anchor and the unknown node. If the distance is zero, w; is one. On the other hand, if the distance is infinity then kj is zero. We choose higher w; closer to unity for a smaller distance. If the unknown node falls in the convex hull formed by the neighboring anchors, localization error is minimal. In that case, the region of the ambiguity is smaller. Refer to the problem of exercise. 4.3.3 Time-Difference-of-Arrival Based Localization We explain Time-difference-of-arrival (‘TDOA) based localization now. The received signals at ith unknown node is given by yilt) = od (t)a(t — 7) + ni(t) for i = {1,2} (4.12)Smart Device Localization 127 “ x10 525 £ B2 £ Bis g 1 Pos ° a5 0 5 0 5 0 4 Lag FIGURE 4.5: Correlation versus lag where of represents the attenuation factor for the anchor and ith node. n; is zero-mean Gaussian noise. 7; denotes the time delay associated with the ith node. The estimated er relation is expressed as Lp Runt = 3 wilt)ya(t— 7) dt (4.13) where 7’ is the observation period. The TDOA can be estimated by maximizing the correlation function as TDOA = argmax Ry yo (7) (4.14) The TDOA information is used to further to locate the smart device. Example of TDoA based localization: Two signals y:(t) and yo(t) are utilized for computing the cross-correlation. We consider 10 samples and 15 samples, respectively, in y1(#) and yo(#). The maximum correlation occurs when the lag is 5 as shown in Figure 4.5. This is expected since the difference between the number of samples in those two signals is The difference between TOA of pair of nodes, that is, TDOA is utilized herein for obtaining the location of unknown sensor node. It requires precise synchronization among anchors. However, it does not require synchronization between the unknown node and anchors. It may be noted that TDOA is an efficient estimator as compared to the TOA-based method. 4.3.4 Angle-of-Arrival Based Localization We first use trigonometry and geometry in Angle-of-Arrival (AOA) based localization. The angle and location relationship between the unknown node and anchor can be expressed as (zi -2) =(yi—y)tandi, ie N (4.15)128 Fundamentals of Internet of Things r Tos where X = [x |” is an unknown node, and Xj = [2; y|” is ith anchor The angle of arrival for ith anchor is denoted by ¥;.The location of an unknown node using least-square method can be estimated a: &=(ATA)ATB (4.16) where 1 —tanda m1 —y tan yy 1 —tanw: 22 — yg tan yo A and B 1 Similar to multilateration based localization, the concept of pseudo-inv« is used herein for rectangular A and B matrices. It may be noted that the AOA-based method uses angle information. This increases cost and complexity because of the use of an antenna array. Example of AoA based localization: Consider the locatious of the zy ~ yy tandy all the anchors as [71 v1] = [0 0)", [2 wJ" = [0 1)", [zs ys] = [2 O]?,and (rg ys)” = [1 1]”. The actual location of the unknown sensor node is assumed to be [x y]” = [0.5 0.5)”. The Ganssian noise in angle measurement is considered to be A’(0, 10~%). Solution: We get yy = 180° +.45°, yo = 180° — 45° 45°. Now, the matriees can be written as 0° — 45°, and i -1 a [ 0.0000 | 1 0.99930 0.99930 7 A= 1 o99970/ sd B= 1 (a7) li 1.0016, [-o ovis] ‘The estimated location of the unknown node is 0.4998 8 [ soo (48) The estimated location is closer to the actual location. Notably, if the unknown node lies in the convex hull formed by the neighboring anchors, the location estimate becomes accurate as expected, 4.3.5 Received Signal Strength Based Localization We discuss next Received Signal Strength (RSS) based localization. The RSS measurements at the unknown node due to an anchor can be expressed as a P,(d) = Po(do) — 10af log (2) +n (4.19) lo where do denotes the reference distance from the transmitter and a/ represents the path loss exponent. 1) is Gaussian noise with zero mean and variance o?Smart Device Localization 129 20 40 60 80 Received Signal Strength (4Bm) 5 10 15 20 25 30 Distance (m) FIGURE 4.6: Received signal power with distance Py is the power at reference distance do. Unbiased estimate of true distance using maximum likelihood estimate is given by Py) (—log(10)o e: .20 as «o(ain) oo( 2052 (420) RSS is a low-cost measurement and it depends upon the wireless channel. It may yield large errors because of the requirement of an accurate propagation model. We can also use the least-square method to estimate distance. Refer to the problem of exercise. Example of RSS based localization: The path loss exponent 7) varies between 2 for free space to 6 for cluttered environment. This can be also esti- mated given the measurements. We consider n = 4.5, dp = 1 m. The received power is the function of transmitted power, distance, antenna parameter, and the state of the channel. Given the received signal power, we estimate the distance and vice-versa, The received signal power decreases wit! the distance between the anchor and the umknown node as shown in Figure 4.6. Received signal strength below some threshold like —80 dBin or —90 dBm is generally considered to be the noise floor. This varies depending upon the specifications of the hardw: environmental conditions. e and 4.3.6 Multidimensional Scaling (MDS) Based Localization ‘The algorithm based on the MDS computes the location of the unknown node, given the set of distances between cach pair of nodes, The squared130 Fundamentals of Internet of Things distance matrix is given by 0 dew. diy a, 0 oe Big p-| aw (4.21) Gy Bye 0 Apply double centering matrix to D Dae = IDI 4.22) de where 1 J =LUyixiny wit (4.23) with 1 is the vector of all ones. Apply Singular Value Decomp ion (SVD) on Dag Dac = QAQ" (4.24) where A is the diagonal matrix having cach clement as eigenvalue and Q a matrix corresponding to the eigenvector. Choosing the first two eigenvalues and the corresponding eigenvectors for two-dimensional coordinates system Dnow = QnewAnew@ how (4.25) ‘Transform the relative coordinates system X = QuewAiiew (4.26) to absolute coordinates system using anchor nodes. Example of MDS based Localization: The locations of the all the sen- sor nodes are [x1 yi)” = [0 0), [22 vl” = (1 0], [zs vs]” = [0 1], {ra mJ" = [1 I, and [x5 ys)” = [0.5 0.5] Solution: The squared distance matrix oO 1 1 2 08 1 0 2 1 05 D 1 2 0 1 O05 (4.27) 2 1 1 0 05 The double centering matrix 08 -0.2 -02 0.2 -0.2 -0.2 08 -0.2 -0.2 -0.2 02 -02 08 -0.2 -0.2 (4.28) -0.2 -0.2 -02 08 —0.2 02 -02 -02 -02 08Smart Device Localization 131 and 10 0 10 Oo -1 1 0 0 Dic=}O 1 00 (4.29) 1 0 0 -10 0 0 0 0 0 ‘Therefore, using singular value decomposition, 0 —0.7071 0 0.7071 0 0.7071 0 0.7071 0 0 0.70710 07071 0 0 (4.30) 0 0.7071 0 0.7071 0 [| o 0 0 0 4 and 20000 02000 A=|00000 (431) 00000 00000 for two dimensional coordinate systems, considering fist two columns of Q and squared root of two dominant eigenvectors, we get 0 -% [ 0 1 0 1 0 v3 v2 0 ] x an) 10 4.32) vo4 [ 0 v2 01 oe) va 0 0 oo which is a translated, rotated, and scaled version of original location of the nodes. The MDS-based localization method only requires initial pairwise dis- tance information. It also provides good initial coordinates for other localiza- tion methods, This method requires the set of anchors as reference nodes for converting the relative coordinate system to absolute coordinate system. 4.4 Distance-free Localization Methods ‘The distance-free based localization has low complexity than the distance- based method. Distance-free methods use connectivity information and yield132 Fundamentals of Internet of Things coarser location estimates. Distance-free localization methods like Approx- imate Point in Triangle Test (APIT), Distance Vector-HOP, Centroid al- gorithm, Closest point-based method, and Assumption based coordinat. method have been extensively dealt with in literature. The inertial based localization algorithm uses accelerometer and gyroscope for estimating the step counts and acceleration, respectively. Distance-free localization methods are based on distance-hop or geometric configuration of sensor nodes. These localization methods are good for coarse location in a large IoT network 4.4.1 Centroid-based Localization The centroid based localization is the most basic localization technique, ‘This is very simple and effective for coarse location. The centroid-based local- ization method for sufficiently large number of anchor nodes is given by (4.33) where [# g]” is the estimated location of unknown node. [2; yi]" denotes the ith anchor and of represents the confidence of ith anchor communication. Example of Centroid Localization: Let the locations of all the an- chors be [er mJ” = [0 0}, [x2 ys)” = [0 1), [es ys)” = [1 O], and {za uJ" = [1 1]. The location of the unknown sensor node is assumed to be [0.2 0.2]. ‘The distances of the unknown node from the anchors are d; = 0.2828, dz = 0.8246, dg = 0.8246, and dy = 1.1314. The radio communication range of the anchor is denoted by R. Assume, the value of R is set to 0.9 without the loss of generality. The smart devices have limited communication range and hence it cannot hear all the anchors deployed in the network. Thus, the node \" |" can hear anchors located at [r1 yi], [z2 yo], and [x3 ys the centroid locations of the these three anchors is [: 231 (4.34) This is the estimated location of the unknown node, Notably, in order to increase the localization accuracy further, the confidence parameter af is also taken into account. The confidence parameter is expressed as of = qataas The anchor which is located closer to the unknown node has larger weight The estimated location is [0.3 0.3)" for af = gL. This is called weighted centroid based localization ‘Therefore,Smart Device Localization 133 FIGURE 4.7: Illustration of the DV-HOP localization 4.4.2 Distance Vector Hop Based Localization ‘The Distance Vector (DV)-HOP localization method is based on the av- crage hop distance between two sensor nodes as shown in Figure 4.7. The method does not utilize the absolute distance between two neighboring sensor node. Therefore, it gives coarse location of node than distance-based methods. The DV-HOP based localization method is given by Dis Ve = 9)? + i = 5? HoPaverage = 4.35) loPaverags Dig hig (488) where (1;,y;) is the coordinates of the ith anchor, and hj, is the hop count from ith node to jth node. Example of DV Hop Based Localization: Let dij = 0.5 and dj, = 0.3. The average hop distance using Equation 4.35 is given by 0540.3 _ 29 36) 7 = 0.29 (4.36) There are three hops between nodes i and k. Hence, the distance between i and k nodes are 3 x 0.29 = 0.8, This is a low-cost localization method for a small sensor network in line-of-sight scenario, However, it may be noted that, computing the accurate distance between two sensor nodes is highly erroneous especially in an indoor scenario or a large IoT network. This is because of the loss due to obstruction such as wall and furniture, This creates non-line-of- sight conditions. 4.4.3. Closest Point Based Localization ‘The closest point localization method is employed for obtaining the car- dinal location information, The coarse location of the unknown seusor node may be decided based on the following conditions: © Which anchor is the closest to node receiving maximum power? « Whether the node is in the communication range of ith anchor?134 Fundamentals of Internet of Things A Ay AL Ay Aa As FIGURE 4.8: An illustration of point inside and outside of a triangle in APIT * Direction at which nth node receives maximum ener * Nearfar information between pair of nodes Example of Closest Point Based Localization: Consider the same set locations of all the anchors as follows: [21 yi]” = [0 0], [22 yo|” = (0 1), [zs ys)” = [1 0], and [4 yaJ” = [1 1]. The location of the unknown sensor node is assumed to be [0.2 0.2) ‘The distances of the unknown node from the anchors 1, 2, 3, and 4 are, respectively, dy = 0.2828, dz = 0.8246, ds = 0.8246, and dy = 1.1314, The radio communication range of the anchor is denoted by R. Assume, the value of R is set to 0.9 without the loss of generality. According to closest. point based localization technique, the unknown node is in the vicinity of anchor 1 Hence, the coarse location of the unknown node is mi] _ fo | fn) bl aan Note that the closest point localization provides a low-cost solution, easily implementable, low energy consumption, and low resolution of the location estimate. 4.4.4 Approximate Point in Triangle Test (APIT) We know that a few nodes that are equipped with high-powered trans- mitters are called anchors. The unknown node can obtain its location based on the information from the anchors. If there exists a direction of movement such that NV is closer or further away to Ay, 4g, and Ag simultaneously, then N is outside of A AjAgAy as shown in Figure 4.8. Otherwise, V is inside A A\A2Ag Example of APIT: First, a grid array is formed and all the grids are initially assigned a zero value as shown in Figure 4.9. It is then scanned using a grid scan algorithm. If the node lies inside the triangle, it is incremented bySmart Device Localization 135 FIGURE 4.9: Illustration of the intersection of overlapping region in APIT one, The grid array is used to choose the grid where the node is most likely to reside. In this example, the size of the grid is 0.2 times the radio commu- nication range of the sensor. It is worth mentioning that the number of grids can be increased by decreasing the size for better accuracy. The number of the possible triangle for our scenario is three for illustration. However, in practice, all the possible combinations of the triangle are taken into account. APIT is a cost-effective approach and it has good localization aceuracy, even in the presence of irregular radio patterns, and with the random node deployment. 4.4.5 Assumption Based Coordinates (ABC) Localization Assumption Based Coordinates (ABC) is a distributed localization method. Initially, it is difficult to get the range measurements from more than three anchors for a 2D coordinate system. Therefore, a map based on the neighboring ranges is generated at the beginning of the network operation. ‘The Assumption Based Coordinates (ABC) localization method is shown, in Fignre 4.10. Here, we consider only positive value after square root op. tion. If the line joining origin and (2, y2,0) subtends an angle € with « axis, then we can write using law of cosine as ds + dip ~ dig cos(é) = 4.38) ©) 2dordoz (4.88) ‘We can write x coordinate using cos(é) = £2 and Equation 4.38 as dy + dip ~ din c: ——_ 4.39) ca ae (4.39) Here, doz = G2 — 0)? + (v2 — 0)? + (0— 0)? = \/73 +f, we can write y coordinate as yo = Gin ~ 3 (4.40)136 Fundamentals of Internet of Things y axis, Or Yay 0 x axis zaxis FIGURE 4.10: An illustration of node placement in ABC algorithm Similarly, the following geometry based expressions for node location is given by dos + diy ~ dis aa 2dor aay) By + dy — dy — 2 yy — Hot dis — 93 (4.42) _— 23 = ig — 23-5 (4.43) ‘The expressions for «rs and 23 are evident. Refer to the problem of exercise for determination of y, coordinate. Example of ABC Localization: Let the actual locations of the sensors be (1,42) = (1, 0), (x2, yz) = (0.5, 1) and (:r3, yz, z3) = (1, 1, 1). Subsequently, distances among nodes are do: = 1, doz = 1.1180, dig = 1.1180, dos = 1.7321, dys = 1414, and dz = 1.1180. We add the noise \’(0,10-*) in the actual distance to get the estimated distances. We get the coordinates as follows: _ 1.007? + 1.1180? — 1.11877 an 21.007) (4.44) ye = V 1.118 0.49! (4.45) 1.007? + 1.7318? — 1.4141? at a= 2(1.0007) (4.46) 1.7318? — 1.1195? + 0.4995? + 1.002? — 2(0.4995) (0.9998) yy = ROHN) Caan) (1.002) 23 = V/L.7318 — 0.9998? — 0.9983? (4.48)Smart Device Localization 137 The estimated locations of the sensors are (#2,i2) = (0.4995, 1.002) and (4s, Gs, 2s) = (0.9998, 0.9983, 1.0015) which is close to the actual locations of the sensors. 4.5 Performance Metrics ‘The performance of the localization algorithm is assessed by different met- rics, namely, average localization error, the variance of localization error, the minimum variance of localization error by Cramer-Rao Lower Bound (CRLB), box and whisker analysis, and cumulative distribution function (CDF) 4.5.1 Average Localization Error ‘The motivation for using normalized localization error instead of absolute localization error is as follows: getting 0.2 m in 1 m x 1m network is poorer than 0.2m in 10 m x 10m, The normalized errors are 20 % and 2% for the former and latter scenarios, respectively. The normalized localization exror is defined as the average of sum of Euclidean distance between the estimated and actual target locations per target and per network dimension. It is given by 1 2 Normalized Localization Eztor = 57 SY ii — xe (4.49) where D and T denote the network dimension and the uumber of targets, respectively. x; and %, are the actual and the estimated target locations, respectively. 4.5.2 Cramer-Rao Lower Bound In this section, Cramer-Rao Lower Bound (CRLB) analysis for the pa- rameter of node location is presented. CRLB gives the lower bound on the variance of any unbiased estimator. It is used as a standard benchmark with spect to which variance of the estimator is compared. The parameters of node location is y = [2 y]"", whose CRLB is to be estimated. Consider p = af + 85 logdg as an RSS-distance model, where af and By are the parameters of Friss transmission formula. This depends on antenna gain, transmitted power, and wavelength of the carrier signal. For Gaussian RSP observations, that is, p ~ N(u(y),C(y)), Fisher information matrix (FIM) is written using the expression of Kay et al. (1993) as FIM(y) = [MA]Porve [22 t je[(cw 22] (4.50)138 Fundamentals of Internet of Things where tr(.) denotes trace of a matrix, which is sum of its diagonal elements. C is the covariance matrix which does not depend on y. It can be shown that, g,e=2—) g, yew BS 8, Oy g, <= 2a) (y= vo) [uo _ [bra Bra ae (4.51) a, Begs), let & Substitute Equation 4.51 into Equation 4.50 to get the following expression, for FIM, FIM(y) = x 0 07) Pea Bae ar aie a 6,7 B4] | 0 a 0) ea BG ae 3 ee taps o 0 sx] [9% qz4 By 4G aad ere Pe where o2 represents variance corresponding to cach of the ath anchor, Va € A. The FIM for parameters of the node location is finally given as Since CRLB is the sum of diagonal elements of inverse of FIM CRLB(y) = [hia + a2 (4.54) If the variance of the location estimator is close to the CRLB, the estimator is said to be efficient as shown in Figure 4.11, 4.5.3 Box and Whisker Plot G Some of the observations follow a certain pattern and treated as normal, while some of them may not follow that distribution and are deemed as outliers. In order to detect the outliers, a box-and-whisker analysis plot is used. The plot is used for two purposes. First, to evaluate the effectiveness of the location estimator based on localization error. The location estimator is said to be enerally, observations are affected because of noise and multipath fading, efficient if the number of outliers is minimal. Second, it can be utilized for cleansing the raw signal strength data first and subsequently feeding it to the location estimator.Smart Device Localization 139 014. 0.12 - on CRLB 0.08 -, 0.06 1 08S _— “02 yaxis xaxis FIGURE 4.11: Cramer-Rao lower bound of the position estimator Median error is shown by horizontal segment at the center of the box in Figure 4.12. Note that 25% and 75% percentile are first and third quartiles, respectively. These are denoted using lower and upper hinges of the box plot respectively, Minimum and maximun location errors are represented by lower and upper fences. The observation which is greater than one and half times the height of the box from the median is treated as an outlier. Hence, location error above the outer fence is considered to be an outlier as shown by cross symbols. Although, the number of extreme localization errors is more for method 1, however, the median localization error is small for method 1. Hence, it can be concluded that method 1 is good in average sense and method 2 in the outlier sense. 4.5.4 Cumulative Distribution Function ‘The cumulative distribution function (CDF) plot is utilized to investigate the distribution of localization error. For illustration purposes, we show the CDF plot for two methods in Figure 4.13. We mark the localization error and probability of error on x and y axes, respectively. The CDF curve should attain a unity probability rapidly. Ideally, it should resemble a unit step function. In this scenario, the probability of a large localization error is small. It ean be seen from the figure that the performance of method 1 is better as compared to method 2.140 Fundamentals of Internet of Things Method 1 Method 2 FIGURE 4.12:-Box-plot- Cumulative Distribution Function Local FIGURE 4.13:-Cumulative-distribution-function-plotSmart Device Localization 141 4.6 Summary ‘The localization accuracy of the distance-based methods depends upon the reliability of the radio propagation model. Fingerprinting localization yields good accuracy at the expense of extensive offline training. The primary aspects in choosing the method are accuracy, computational complexity, scalability, cost, and energy consumption. There is a trade-off between accuracy and avail- able computational resources. For a one-shot process, one may choose the algo- rithm which has high accuracy. On the other hand, for real-time applications, the accuracy is tolerated up to some extent for low complexity algorithms. We can also utilize the available opportunistic signals in the environment such as light signal, acoustic signal, radio signal, geomagnetic disturbances, and crowd-sourcing for localization tasks. 4.7 Exercises in an indoor 1. What is the average localization exor using GPS nal 2. Consider multilateration based localization using pseudo-inverse. There are three clements in unknown matrix A. However, only two unknown variables x and y in the two-dimensional coordinates system are of in- terest. Explain. 3. Consider the location of the anchors [21 yi” = [0.2.0.2], [ea ye = [0.2 0.84], (x3 ys]” = [0.25 0, and [24 ys]” = [0.82 0.82 and unknown node [xy] 0.12 0.14] are randomly deployed. De- termine the location of the unknown node using ToA based technique. 4. What is the physical significance of lag in TDoA based localization? 5. Ifwe use sound signal having speed 331 m/s in timing-based localization, how inuch would be an error in the distance for clock synchronization exror of 1 s? Draw some inference. 6. How do we measure the angle in AoA/DoA based localization method? When does the centroid-based localization give very poor accuracy? 8. If we consider a two-dimensional network [0 1] x [0 1]. There are four anchors at the four corners of the network. If the node hears all four an- chors, what is the estimated location using centroid-based localization?142 10. 1. 12. 13, 14. 16, Fundamentals of Internet of Things Why do we take average hop size not absolute hop siz localization? :n DV-Hop-based Determine the expression for y coordinate of Assumption Based Coor- dinates (ABC) localization technique as mentioned in the theory section, Let us assume that the average RSS at distance of 0.1 m, 2 m, and 3m from the access point are, respectively, —30 dBm, —45 dBm, and —55 dBm. Estimate the distances between access points and smart de- vices, Let the coordinates of three access points be (x1,y1) = (0, 0), (x2, y2) = (2, 1) and (x3, ys) = (6, 6). Refer to Problem 11, perform multilateration based localization using the estimated distances How do we convert to the absolute coordinate system after using MDS- based localization? ind C) ving F 1 0 12 Find CRLB for following FIMs: [ ‘| and Ci 4 In performance evaluation, we calculate the localization error as the Euclidean distance between actual and estimate node locations. Mention other metries that can be used. Do It Yourself: Explore accelerometer and gyroscope and make a project on pedestrian tracking using measurements from these devices.