M.R.BASHEER ET AL.: LOCALIZATION AND TRACKING OF OBJECTS USING CROSS-CORRELATION OF SHADOW FADING NOISE Appendix Computing Area of Overlap between Ellipses The overlapping area |π12 | in (9) can be calculated by first computing the point of intersection ππ = (π₯ππ , π¦ππ ); π ∈ {1,2,3,4} between elliptical regions π1 and π2 . Since the maximum RSSI path detection delay τm is same for both receivers, and the ellipses π1 and π2 share a common focus at the transmitter, it can be easily shown that the number of intersection points is only two. Let π1 = (π₯π1 , π¦π1 ) and π2 = (π₯π2 , π¦π2 ) be the point of intersection between π1 and π2 then to find π1 and π2 , simultaneously solve the implicit polynomial equations of ellipse π1 and π2 by eliminating one variable, for e.g. x, leading to a quartic equation in y. The intersection points are then the real solutions of this quartic equation. The generalized implicit equation of an ellipse with semi-major and semi-minor axis given by a i and bi , oriented at an angle Οi w.r.t x-axis (ππ₯π , ππ¦π ) with center at is given by [(π₯−ππ₯π ) πππ ππ +(π¦−ππ¦π ) π ππ ππ ]2 ππ2 + [−(π₯−ππ₯π ) π ππ ππ +(π¦−ππ¦π ) πππ ππ ]2 ππ2 = 1. For π1 , π1 = 0 and (ππ₯1 , ππ¦1 ) = (0,0) while for ellipse S2 , π2 = π = πππ −1 ( π2 π1 π2 2 2 ( πππ π − , 2 2 π12 +π22 −π12 2π1 π2 (ππ₯2 , ππ¦2 ) = )and π ππ π). Subsequently, the area can be computed from pl using Gauss-Green theorem as 1 π |π12 | = |π1 | + |π2 | − ∫0 [π₯1 (π1 ) 2 1 2π ππ¦1 (π1 ) ππ − π¦1 (π1 ) − ∫π [π₯2 (π2 ) 2 where π1 = ψ11 = (π12 −π11 ) π x cos−1 ( ap1) 1 , ππ¦2 (π2 ) π12 = πππ ππ (π22 −π21) π −1 ( ] ππ − π¦2 (π2 ) ππ π + π11 , π2 = ππ₯1 (π1 ) π₯π1 π1 ππ₯2 (π2 ) π2 [(π¦ππ −ππ¦2 ) π ππ(π)+(π₯ππ −ππ₯2 ) πππ (π)] ) , π21 = ) π22 = π2 [(π¦π2 −ππ¦2 ) πππ (π)−(π₯π2 −ππ₯2 ) π ππ(π)] π‘ππ−1 ( π2 [(π¦π2 −ππ¦2 ) π ππ(π)+(π₯π2 −ππ₯2 ) πππ (π)] π1 πππ (π1 ) − (π₯π1 +π₯π2) 2 , ), π¦1 (π1 ) = π1 π ππ(π1 ) − π2 πππ (π) πππ (π2 ) − π2 π ππ(π) π ππ(π2 ) + ππ₯2 − ] ππ π + 2π21 − π22, π2 [(π¦ππ −ππ¦2 ) πππ (π)−(π₯ππ −ππ₯2 ) π ππ(π)] π‘ππ−1 ( ππ π₯1 (π1 ) = (π¦π1 +π¦π2 ) 2 , π₯2 (π2 ) = (π₯π1+π₯π2 ) and 2 π¦2 (π2 ) = π2 π ππ(π) πππ (π2 ) −π2 πππ (π) π ππ(π2 ) + ππ¦2 − (π¦π1 +π¦π2 ) 2 . Proof of Theorem 1 (Shadow Fading Correlation Coefficient Between IEEE 802.15.4 Receivers) Figure 2 shows the elliptical scatterer regions π1 and π2 surrounding receivers π 1 and π 2 respectively. Let the number of obstacles in π1 and π2 at any communication instance between the transmitter and receiver is given by the Poisson distribution (2). If πΌππ ; π ∈ {1,2}, π ∈ {1,2, β― , π(ππ )} represents the attenuation of a radio signal reaching receiver R i due to jth obstacle in scatterer region ππ , then the net reduction in signal strength (in dBm) due π(π ) to π(ππ ) obstacles in region ππ is given by ππ π = ∑π=1 π πΌππ π where ππ is the signal strength attenuation due to shadow fading. For log-normally distributed shadow fading noise under stationary conditions, if π(ππ ) is given then πππ is normally distributed i.e. if μs and σ2s corresponds to the mean and variance of πΌππ , then π(ππ π |π(ππ )) = π(ππ )π(ππ , ππ 2 ) where π(β) is the normal distribution PDF 1 with conditional mean and variance given by πΈ[ππ π |π(ππ )] = ππ π(ππ ) and πππ[ππ π |π(ππ )] = ππ 2 π(ππ ) respectively. The correlation coefficient between shadow fading random variables ππ 1 and ππ 2 is given by π= πππ£(ππ 1 ,ππ 2 ) √πππ(ππ 1 )πππ(ππ 2 ) (A1) which require the derivation of πππ(ππ π ) and πππ£(ππ 1 , ππ 2 ). Since πΈ{π(ππ )} = πππ{π(ππ )} = π|ππ |, πππ(ππ π ) can be derived from law of total of variance as πππ(ππ π ) = πππ{πΈ[ππ π |π(ππ )]} + πΈ{πππ[ππ π |π(ππ )]} = πππ{π(ππ )ππ } + πΈ{π(ππ )ππ 2 } = π(ππ 2 + ππ 2 )|ππ | (A2) whereas, πππ£(ππ 1 , ππ 2 ) can be derived from the law of total covariance as πππ£(ππ 1 , ππ 2 ) = πππ£{πΈ[ππ 1 |π(π1 )], πΈ[ππ 2 |π(π2 )]} +πΈ{πππ£[ππ 1 , ππ 2 |π(π1 ), π(π2 )]}. (A3) Since conditional mean is given by πΈ[ππ π |π(ππ )] = ππ π(ππ ), the covariance of the conditional mean in (A2), can be simplified as πππ£{πΈ[ππ 1 |π(π1 )], πΈ[ππ 2 |π(π2 )]} = ππ 2 πππ£[π(π1 ), π(π2 )]. Since for a spatial Poisson processes, random variables corresponding to the Poisson count for disjoint areas are independent, the radio obstacle count for Si can be written as the sum of two independent Poisson random variables by splitting the region ππ into two π )βπ disjoint areas as ππ = (ππ βπ12 12 resulting in π(ππ ) = π )βπ π ) π((ππ βπ12 ) = π(π βπ + π(π Let π΄π = 12 π 12 ). 12 π ) π |) π(ππ βπ12 and π΅ = π(π12 ) then, π΄π ~ππππ π ππ(π|ππ βπ12 and π΅~ππππ π ππ(π|π12 |) resulting in πππ£[π(π1 ), π(π2 )] = πππ£(π΄1 + π΅, π΄2 + π΅) = πππ(π΅) = π|π12 |. Therefore, the covariance of conditional mean in (A3) is given by πππ£{πΈ[ππ 1 |π(π1 )], πΈ[ππ 2 |π(π2 )]} = ππ 2 π|π12 |. To compute the expectation of conditional covariance πΈ{πππ£[ππ 1 , ππ 2 |π(π1 ), π(π2 )]} in (A3), we split the shadow π(π ) π΄π +π΅ π fading noise at each receiver as ππ π = ∑π=1 π πΌππ = ∑π=1 πΌπ = πΆπ + π· where πΆπ is the shadow fading attenuation due to π region ππ βπ12 and π· is the shadow fading due to the overlapping region π12 . Therefore, the conditional mean is given by πΈ{πππ£[ππ 1 , ππ 2 |π(π1 ), π(π2 )]} = πΈ{πππ£[πΆ1 + π·, πΆ2 + π·|π(π1 ), π(π2 )]} = πΈ{πππ[π·|π(π12 )]} = ππ 2 π|π12 | resulting in (A3) being simplified as πππ£(ππ 1 , ππ 2 ) = π(ππ 2 + ππ 2 )|π12 |. (A4) Finally applying (A2) and (A4) on (A1) results in (9). β Proof of Theorem 2 (Shadow Fading Cross-Correlation Likelihood Function) The cost function for the maximum likelihood estimate of a parameter is the joint distribution of the multiple observations of a random variable when the value of the parameter is assumed to be known. For our application the Cartesian coordinates of the transmitter is the parameter to be estimated while the random variables that are being observed are the shadow fading residuals at each receiver. Therefore, to derive the joint distribution of shadow fading residuals from semi-parametric marginal CDF giv- 2 IEEE TRANSACTIONS ON MOBILE COMPUTING, MANUSCRIPT ID en by (11) and pair-wise correlation coefficient given by (9) we will use the elliptical copula function since the dependency between the shadow fading residuals that is being modeled is the correlation coefficient which is a linear dependency. In addition, t-copulas capture the linear dependency between extreme values of the random variable [17]. Hence for π receivers, the student-t copula density with π degree of freedom (DoF) and π × π correlation coefficient matrix βΆ is given by [17] as ππ,βΆ (π’1 , π’2 , … , π’π ) = (A5) −1 ∏π π=1 ππ (π‘π (π’π )) π+π ππ,βΆ (π₯1 , π₯2 , … , π₯π ) = π€( 2 ) 1 π π |π²|2 (ππ) 2 π€( ) 2 [1 + π© = [π₯1 , π₯2 , … , π₯π ππ (π₯) = π (π+π) 2 − ] , (π+1) ππ = ππ−1 √1 − 2π₯ππ πππ ππ−1 ππ−1 , and Μπ (π§π π ) π€(β) is the Gamma function. Finally, setting π’π = πΉ +[ written as |ππ | = |ππ−1 | (1 − π½π−1 = √1 − βn−1 rn−1 rm +rn−1 2π₯ππ πππ ππ−1 ππ−1 ) √1 − +[ π₯ππ ππ−1 βn−1 rn−1 rm +2rn−1 2 +2π π π(ππ +ππ )√ππ π π 4 . Since 2 ] the area ππ can be π½π−1 ππ−1 ππ +ππ−1 ) √1 − π₯ππ 2 ππ−1 ] − 1. 2π½π−1 ππ−1 ππ +2ππ−1 Setting where γn−1 = − 1 results in the elliptical area for ππ being represented by the area of ππ−1 as |Sn | = |Sn−1 |(1 + γn−1 ). Therefore, (A6) can be written as Dα (n − 1 β₯ n) = ω|Sn−1 |[(1 − α)γn−1 + 1] k − log {∑∞ k=0 [ω|Sn−1 |(1+γn−1 )(1−α) ] k! 1 }. Since ∑∞ k=0 [ω|Sn−1 |(1 + k! k − 2 π€( 2 ) π₯2 ] π£ [1 + π √πππ€(2) π+1 ]π π©π βΆ−1 π© and |ππ | = 4 (1 − ππ,βΆ (π‘π−1 (π’1 ),π‘π−1 (π’2 ),…,π‘π−1 (π’π )) where π’π ∈ [0,1] is the standard uniform random variable, 2 +2π π π(ππ +ππ−1 )√ππ π π−1 γn−1 )(1−α) ] = exp{ω|Sn−1 |(1 + γn−1 )(1−α) }, (1−α) k [ω|Sn−1 |(1+γn−1 ) log {∑∞ k=0 k! ] resulting } = ω|Sn−1 |(1 + γn−1 )(1−α) in . in (A5) gives the likelihood function for transmitter Hence, the πΌ-divergence between RSSI values collected at localization using shadow fading cross-correlation as time instants (π − 1) and π is given by (13). (12) β Proof of Theorem 3 (πΆ-Divergence of Shadow Fading Residuals From An IEEE 802.15.4 Transmitter) Figure A1 shows the tracking of an IEEE 802.15.4 mobile transmitter by a stationary receiver. At each sampling instance, receiver collects a sequence of π RSSI values from the transmitter. Assume that at time instant π − 1 , the mobile transmitter is at position ππ−1 and in the subsequent instance π , the receiver moved by radial distance π₯ππ to reach location ππ . In addition, assume that during this time period, the heading of the mobile transmitter changed by π₯ππ−1 while the bearing between the mobile transmitter and stationary receiver at the origin changed by π₯ππ−1 . Shadow fading noise at positions ππ−1 and ππ arise from the movement of pedestrians or machinery within elliptical fading region ππ−1 and ππ respectively formed between the transmitter and receiver. Therefore, the αdivergence between π RSSI values collected at positions ππ−1 and ππ can be derived by substituting the shadow fading PDF given by (1) at positions ππ−1 and ππ in αdivergence equation (2) resulting in π·πΌ (π − 1 β₯ π) = ∞ πΌ 1−πΌ − πππ{∫−∞ ∑∞ ππ₯ } π=0[π(π₯|π)π(π|ππ−1 )] [π(π₯|π)π(π|ππ )] ∞ πΌ 1−πΌ = − πππ{∫−∞ ∑∞ ππ₯} π=0 π(π₯|π)[π(π|ππ−1 )] [π(π|ππ )] ∞ πΌ 1−πΌ = − πππ{∑∞ ∫−∞ π(π₯|π)ππ₯ } π=0[π(π|ππ−1 )] [π(π|ππ )] πΌ 1−πΌ } = − πππ{∑∞ π=0[π(π|ππ−1 )] [π(π|ππ )] |ππ | = π(1 − πΌ)|ππ−1 | (|π π−1 | πππ {∑∞ π=0 − 1) − ππ₯π{−π|ππ−1 |}(π|ππ−1 |)π π! |ππ| (|π π−1 ) | (1−πΌ)π } (A6) where |ππ−1 | and |ππ | are the area for elliptical regions ππ−1 and ππ respectively and are given by |ππ−1 | = Figure A1. Continuous tracking of a mobile receiver