2.1 Locata Navigation Component

LocataLite devices transmit signals at two S-band frequencies (S1 – 2·41428 GHz and S6 – 2·46543 GHz), and each LocataLite setup consists of two transmitting (Tx1 and Tx2) antennae and one receiving (Rx) antenna. Hence a cluster of four distinct signals are transmitted from each LocataLite. The frequency diversity provided by the cluster of four signals could help mitigate the cycle slip problem caused by signal interference.

2.1.1 Locata multipath-mitigation technology

Locata uses signals chipped at 10·23 MHz and the chip length is reduced to about 30 metres. This can help to eliminate the long wave multipath delays of more than two chips. However, the challenge in an indoor environment is that the short wave multipath causes less than one chip delay. To overcome the multipath problem in such challenging environments the Locata “V-Ray” antenna is used – a design based on Correlator Beam-Forming (CBF). This technique uses only one RF front-end to generate beams by sequentially switching through elements of an antenna array. Then the beam is formed with phase and gain corrections in the receiver's channel correlators. The V-Ray antenna can provide multiple beams pointing in different directions almost simultaneously (at microsecond-level switching speed).

To mitigate multipath, the formed beams must be pointed to the LocataLite transmitters. The LocataLites’ coordinates are known; thus the beam's pointing direction is dependent on the orientation of the antenna and the position of the Locata receiver. The angle-of-arrival measurements can be used to estimate the approximate rover antenna orientation and position. The receiver sweeps the beams looking for beam settings that could maximise the signal power from each LocataLite; in this way the angle-of-arrival measurements can be determined once the optimal beam settings are found.

The “V-Ray” antenna used in these tests is a basketball-sized antenna array consisting of 20 panels with 80 elements. Each panel has four elements arranged with one central and the other three on three corners to form a triangle. The space of the element between two adjacent panels is a half wavelength of the Locata signals (LaMance and Small, Reference LaMance and Small2011). The V-Ray antenna is shown in Figure 1.

Figure 1. Locata V-Ray antenna which is based on the correlator beam-forming technique to overcome the multipath effect indoors.

2.1.2 Locata mathematical model

Similar to GNSS, Locata has two types of measurements: pseudorange and carrier phase. For LocataLite channel i, the basic Locata carrier phase observation equation can be written as:

(1) $$\phi^i \cdot \lambda =\rho^i +\tau _{trop}^i + c \cdot \delta T + N^i \cdot \lambda +\varepsilon_{\phi}^i$$

where ϕ ⁱ is the Locata carrier phase measurement (in cycles), λ is the carrier signal wavelength, ρ ⁱ is the geometric distance between the Locata receiver and the transmitting antenna i, $\tau_{trop}^{i}$ represents the tropospheric delay, δ T is the receiver clock bias, c is the speed of light, N ⁱ is the Locata float carrier ambiguity and $\varepsilon_{\phi}^{i}$ is the unmodelled residual errors. Due to the tight time synchronisation of the LocataLites, there is no transmitter clock error item in the observation equation.

The receiver clock error can be estimated via system filtering, or eliminated by differencing the measurement. In this research, the receiver clock error is eliminated by Single-Differencing (SD) the carrier phase measurements of the same frequency between two LocataLites. The tropospheric error (if present) in $\Delta \tau _{trop}^{ij}$ could be calculated by tropospheric model and many tropospheric models have been proposed in published work (Choudhury et al., Reference Choudhury, Harvey and Rizos2009b).

The SD observation equation can be written as:

(2) $$\Delta \phi^{ij} \cdot \lambda = (\phi^i - \phi^j) \cdot \lambda = \Delta \rho^{ij} + \Delta \tau_{trop}^{ij} + N^{ij} \cdot \lambda + v$$

where j is the reference Locata signal and i is a SD signal of the same frequency as the reference.

The unknown parameters are the Locata carrier ambiguities and the receiver's three position coordinates in the SD geometric range.

2.1.3 Ambiguity resolution

GNSS carrier ambiguities can be resolved by three methods: Known Point Initialisation (KPI), Static Initialisation (SI), or On-The-Fly (OTF) resolution. The SI needs the rover receiver to start at a known point, operating for several minutes until the estimated ambiguities converge to their likely values. As the GNSS satellites move in their orbits, the line-of-sight vectors will have obvious changes after many observation epochs, and this spatial diversity aids Ambiguity Resolution (AR). However, this approach is not suitable for Locata. In Locata applications, the LocataLites are fixed on the ground. For a stationary receiver, the ambiguities cannot be computed by solving these equations directly.

On-The-Fly-Ambiguity Resolution (OTF-AR) is also applied with enough transmitter-receiver geometry change. As the LocataLites are not moving, the OTF-AR method could be realised “kinematically” to provide the relative geometry change. The spatial diversity provided by the moving rover receiver makes Locata OTF-AR implementation possible (Jiang et al., Reference Jiang, Li and Rizos2013).

However, OTF-AR also has a limitation as the spatial diversity should be sufficiently large. Our indoor experiment was conducted in a warehouse. LocataLites were installed at the corners of the building. The distances between the Locata rover and LocataLites were short and the speed of the rover was low, hence the experimental set-up could not satisfy the OTF-AR requirements. Thus the KPI method of AR was used, based on the following expression:

(3) $$N^{ij} \cdot \lambda =\Delta \phi ^{ij} \cdot \lambda - \Delta \rho^{ij} - \Delta \tau_{trop}^{ij} - v$$

To obtain the accurate SD geometric distance $\Delta \rho ^{ij}$ , the initial position of the user receiver is measured by a precise survey. Standard survey techniques (Total Station if indoors and a survey-grade GNSS if outdoors) can be used to compute the initial rover coordinates.

Once the floating ambiguities are computed, they can be treated as a known item in Equation (2).

2.1.4 Locata estimation process

The Kalman Filter (KF) is adapted to process the Locata parameter estimation. The state vector of the system consists of nine parameters (three for position, three for velocity and three for acceleration). Typically, the acceleration is modelled as a white noise process – the white noise process is “isolated” in time since its value at one time is uncorrelated with that at any other time. However, the acceleration of the host vehicle (the moving platform in the proposed research) is dominated by the thrust, which has a temporal correlation. In such a case, it is appropriate to model the acceleration as a first-order Gauss Markov (GM) process rather than as a white noise process. A GM process is “local” in time as its value at one time depends on values at other time instants only through its nearest neighbours. The acceleration model is given by:

(4) $$\dot{a}=-\lpar {1/\tau} \rpar \cdot a + \omega _a$$

where τ is the constant correlation-time and ω _a is white noise with zero-mean.

The state vector can be given as:

(5) $${\bf x}\lpar t \rpar = \left[ {\bf r} \quad \dot{\bf r} \quad \ddot{\bf r} \right]^{\rm T}$$

where r, $\dot{\bf {r}}$ and $\ddot{\bf {r}}$ vectors are the position, velocity and acceleration of the platform and each vector contains three components in three directions respectively.

The Locata system dynamic model is:

(6) $${\bf x} \lpar t \rpar = {\bf F} \lpar t \rpar \cdot {\bf x} \lpar t \rpar + {\bf \omega} \lpar t \rpar $$

where ${\bf F} \lpar t \rpar = \left[\matrix{ {\bf 0}_{{\bf 3} \times {\bf 3}} \quad {\bf I}_{{\bf 3} \times {\bf 3}} \quad \quad {\bf 0}_{{\bf 3} \times {\bf 3}} \hfill \cr {\bf 0}_{{\bf 3} \times {\bf 3}} \quad {\bf 0}_{{\bf 3} \times {\bf 3}} \quad \quad {{\bf I}}_{{\bf 3} \times {\bf 3}} \hfill \cr {\bf 0}_{{\bf 3} \times {\bf 3}} \quad {\bf 0}_{{\bf 3} \times {\bf 3}} \quad -{{\bf 1} \over {\bf \rtau}} \cdot {{\bf I}}_{{\bf 3} \times {\bf 3}} } \right], {\bf \omega} \lpar t \rpar = \left[ {\bf 0}_{{\bf 1} \times {\bf 3}} \quad {\bf 0}_{{\bf 1} \times {\bf 3}} \quad \omega_{a} \quad \omega_{a} \quad \omega_{a} \right]^{\rm T}$

The measurement model is non-linear, the Extended Kalman Filter (EKF) (Kappl, Reference Kappl1971) is thus used to linearize the (non-linear) function around the prediction of the current state.

The Locata system measurement model is:

(7) $${\bf z}\lpar t \rpar ={{\bf H}} \lpar t \rpar \cdot {\bf x}\lpar t \rpar +{\bf \rupsilon} \lpar t \rpar $$

where the measurement vector z (t) is formed by the SD carrier phase measurements:

(8) $${\bf z}\left( t \right)=\left[ {\Delta \phi ^{1j}} \quad {\Delta \phi ^{2j}} \quad \cdots \hfill \quad {\Delta \phi ^{nj}} \right]^{\rm T}$$

H (t) is the observation matrix, which can be written as:

(9) $${\rm {\bf H}}\lpar t \rpar = \left[ \left. \matrix{ {e^1-e^j} \cr {e^2-e^j} \cr \vdots \cr {e^n-e^j}} \right| {0_{n \times 6}} \right]$$

where e* is the line-of-sight vector between receiver and satellite.

${\bf \rupsilon} \lpar t \rpar {\bf \sim} {\bf N} \lpar {\bf 0,R} \rpar$ represents the observation noise. The variance for the carrier phase measurements can be treated as the combination of transmitter location error, multipath, and measurement noise. The errors are assumed to be uncorrelated between measurements. However, the covariances of two different carrier measurements should be set as half of the carrier variance, because half of the measurements are common due to the single differencing (Amt, Reference Amt2006). The measurement covariance matrix can be written as,

(10) $${\bf R}_{\Delta \phi} = {\sigma _{\Delta \phi}^2 \over {2}} \left[\matrix{ 2 \quad 1 \quad \cdots \quad 1 \cr 1 \quad 2 \quad \cdots \quad 1 \cr \vdots \quad \vdots \quad \ddots \quad \,\, 1 \cr 1 \quad 1 \quad \,\, 1 \quad \,\,\,\, 2 \hfill} \right]$$

where $\sigma_{\Delta \phi}^{2}$ is the noise variance of the SD carrier phase measurements.

2.1.5 Cycle slip processing

Similar to GNSS, Locata signals also experience the problem of cycle slips and loss-of-lock due to signal interference or internal receiver tracking issues. The presence of cycle slips could cause a change of the ambiguity parameter value, which would create inconsistencies in the coordinate solution, resulting in a biased solution.

Low Signal-Noise-Ratio (SNR) values can be a useful indicator of an increased risk of cycle slips in the associated measurement, however the SNR indicator is not always reliable, because cycle slips can also occur during periods of good SNR. Another detection strategy has been developed that makes use of the typical Locata signal setup, where a group of four signals are transmitted from the same LocataLite. This method uses the Double-Differenced (DD) carrier phase measurements to detect cycle slips. The first difference is made on the raw carrier phase measurements from the Locata signal i between two subsequent epochs t ₁ and t ₂, to eliminate the ambiguities and any unchanged error:

(11) $$\Delta \phi ^i =\phi ^i (t_1 )-\phi ^i (t_2 )$$

Each LocataLite has four signals and these four come from the same or a very close location (since there are two transmit antennas, and each antenna transmits two signals), thus the second difference is made between the four signals from the same LocataLite, see Equation (12).

(12) $$\nabla \Delta \phi^{ij} = \Delta \phi^i - \Delta \phi^j$$

The original carrier measurements, the first differenced carrier phase measurement and the DD carrier phase measurements from one LocataLite are listed in Table 1. If there are no cycle slips, the DD carrier phase measurements would only have the receiver clock drift and other unmodelled errors, which can be assumed to have very small values (Bertsch et al., Reference Bertsch, Choudhury, Rizos and Kahle2009).

Table 1. The original, the first differenced and double-differenced carrier phase measurements for cycle slip detection.

By comparing the six DD carrier phase measurements from each LocataLite, the channel that contains the cycle slips will exceed a pre-defined threshold, here assumed to be 0·32 cycles. If all the DD carrier phase measurements that are larger than the threshold contain a specific signal, then this signal is assumed to suffer from cycle slips.

After the cycle slip is detected there are two ways to proceed: either to repair the cycle slip in the raw carrier phase measurement(s) or re-fix the ambiguities. Since Locata's cycle slips may be multiples of half cycles, which makes them harder to repair, the second approach is used. The navigation algorithm first calculates the rover's location from the assumed correct carrier phase measurements without cycle slips, then re-fixes the ambiguities with the calculated position, and finally uses the re-fixed ambiguities as known parameters in the subsequent EKF positioning estimations.

2.2 Inertial Navigation Component

Inertial navigation is a self-contained technique, which uses accelerometers and gyroscopes to measure the acceleration and rate-of-angle change, so as to determine the point displacement. The measured accelerations are integrated to obtain the user velocity and position, whereas the gyroscopes are used to maintain a reference frame and measure the angles that the coordinate axes have rotated. Three orthogonal accelerometers and three gyroscopes form an Inertial Measurement Unit (IMU). An Inertial Navigation System (INS) consists of IMU and a navigation computer which provide the user position, velocity, acceleration and attitude information by using the IMU output.

An INS can provide a Position, Velocity and Attitude (PVA) solution by integrating the measured acceleration and angular velocity. The inertial navigation mechanisation is implemented in the navigation coordinate system (n-frame) for PVA updating. The velocity and position updating is implemented by integrating the measured acceleration, whereas the attitude updates are achieved using the quaternion approach. The equations are:

(13) $$\eqalign{ \quad \dot{{\bf r}}^{\bf {n}} &= {\bf v}^{\bf {n}} -{\bf \romega}_{\bf {en}}^{\bf {n}} \times {\bf r}^{\bf {n}} \cr \dot{{\bf {v}}}^{{\bf n}} &= {\bf f}^{\bf {n}} +{\bf {g}}^{\bf {n}}-\lpar {\bf \romega}_{{\bf {en}}}^{\bf {n}} +{\bf 2 \romega}_{\bf {ie}}^{\bf {n}} \rpar \times {\bf v}^{\bf {n}} \cr \dot{{\bf {q}}} &= {{\bf 1} \over {\bf 2}}\left[ {\bf \romega}_{\bf nbq}^{\bf b} \right] {\bf q}}$$

where the subscripts “n”, “e”, and “i” represent the n-frame, the earth-centred-earth-fixed frame (e-frame) and Earth-centred inertial frame (i-frame), rⁿ and vⁿ refer to the position and velocity vectors respectively, $\dot{\bf {r}}^{\bf {n}}$ and $\dot{{\bf {v}}}^{\bf {n}}$ to the position-rate and velocity-rate vectors respectively, gⁿ is the gravity and fⁿ represents the specific force sensed by the accelerometer. ${\bf \romega}_{{{\bf en}}}^{\bf {n}}$ is the angular rate of the n-frame with respect to the e-frame, ${\bf \romega}_{{{\bf ie}}}^{\bf {n}}$ is the angular rate vector of the e-frame with respect to the i-frame, q and $\dot{{\bf {q}}}$ refer to the attitude quaternion and its rate, respectively. ${\bf \romega}_{{\bf nb}}^{{\bf b}}$ and its skew-symmetric matrix $\lsqb {{\bf \romega}_{{\bf nb}}^{{\bf b}}} \rsqb $ can be written as:

(14) $${\bf \romega}_{{\bf nb}}^{{\bf b}} = 0 + \omega _x i + \omega_y j + \omega _z k$$

Then,

(15) $$\lsqb {{\bf \romega}_{{\bf nb}}^{{\bf b}} } \rsqb =\left[ \matrix{ 0 \quad {-\omega_x } \quad {-\omega_y } \quad {-\omega_z} \cr {\omega_x } \quad 0 \quad \quad {\omega_z } \quad \,\,\, {-\omega_y } \hfill \cr {\omega_y } \,\,\, {-\omega_z } \quad \quad 0 \quad \quad {\omega_x } \hfill \cr {\omega_z } \quad {\omega_y } \quad \, {-\omega_x } \quad \quad 0 \hfill } \right]$$

The problem of inertial navigation is that its errors accumulate with time. Even if the INS initialised with proper conditions, the accuracy of the calculated position, velocity and attitude will degrade with time due to the inertial systematic errors and the non-linear nature.

2.3 Indoor Navigation Mode Based on Terrestrial and Inertial Technology

For the indoor navigation configuration of the proposed system, the Locata and INS systems are integrated with a loosely-coupled approach. The computed Locata Position and Velocity (PV) solutions are used to correct the INS PVA solutions within the system integration filtering. In addition to reducing PVA errors, the Locata PV updates are also used to converge the lever arm between IMU centre and the Locata antenna and the gyroscope and accelerometer errors. The estimated lever arm information and sensor errors are fed back to the integration filter to achieve a closed-loop update correction.

To implement the integration estimation filter, the INS error model is defined. Many INS error models have been proposed (Arshal, Reference Arshal1987; Goshen-Meskin and Bar-Itzhack, Reference Goshen-Meskin and Bar-Itzhack1992; Lee et al., Reference Lee, Lee, Rho and Park1998). The psi-angle error model is applied in this research (Goshen-Meskin and Bar-Itzhack, Reference Goshen-Meskin and Bar-Itzhack1992):

(16) $$\eqalign{ {\bf \rdelta} \dot{{\bf r}}^{\bf {n}} & ={{\bf \rdelta}{\bf v}}^{\bf {n}} -{\bf \romega}_{{{\bf en}}}^{\bf {n}} \times {{\bf \rdelta}{\bf r}}^{\bf {n}} \cr {{\bf \rdelta} \dot{{\bf v}}}^{\bf {n}}& = -\lpar {\bf \romega}_{{\bf ie}}^{\bf {n}} +{\bf \romega}_{{\bf in}}^{\bf {n}} \rpar \times {\bf \rdelta} {\bf v}^{\bf {n}} +{\bf \rpsi} \times {\bf f}^{\bf {n}} + \nabla^{\bf {n}} + {\bf \rdelta} {\bf g}^{\bf {n}} \cr \dot{{\bf \rpsi}} &= -{\bf \romega}_{{\bf in}}^{\bf {n}} \times {\bf \rpsi} +{\bf \rvarepsilon}^{\bf {n}} }$$

where ${\bf {\rdelta}{\bf r}}^{\bf {n}}$ , ${\bf {\rdelta}{\bf v}}^{\bf {n}}$ and ${\bf {\rpsi}}$ represent the position, velocity and attitude error respectively. ${\bf \romega}_{{{\bf in}}}^{\bf {n}}$ is the angular rate of the n-frame with respect to the i-frame, $\nabla^{\bf {n}}$ is the accelerometer error, ${\bf {\rdelta}} {\bf g}^{\bf {n}}$ is the error of the computed gravity and ${\bf {\rvarepsilon}}^{\bf {n}}$ is the gyroscopes’ drift error.

The estimated state vector consists of 18 navigation error states, which can be written as three sub-vectors: ${\bf x}_{{\bf {\rpsi}{\bf rv}}}$ , ${\bf x}_{\rvarepsilon \nabla}$ and ${\bf x}_{\bf {\reta}}$ representing the motion, sensor errors and lever arm components, respectively:

(17) $${\bf x} \lpar t \rpar = \left[ {\bf x}_{{\bf {\rpsi rv}}} \quad {\bf x}_{{\rvarepsilon \nabla}} \quad {\bf x}_{\bf {\reta}} \right]$$

Specifically:

(18) $${\bf x}_{\bf {\rpsi rv}} = \left[ \psi_N \quad \psi_E \quad \psi_D \quad \delta r_N \quad \delta r_E \quad \delta r_D \quad \delta v_N \quad \delta vE \quad \delta v_D \right]$$

(19) $${\bf x}_{{\bf {\rvarepsilon}}\nabla} = \left[ \varepsilon_x \quad \varepsilon_y \quad \varepsilon_z \quad \nabla_x \quad \nabla_y \quad \nabla_z \right]$$

(20) $${\bf x}_{\bf {\reta }} = \left[ \delta \eta_{Lx} \quad \delta \eta _{Ly} \quad \delta \eta_{Lz} \right]$$

where the subscripts “N”, “E”, “D” represent the north, east, and down axes in the n-frame, respectively; the subscripts “x”, “y”, “z” denote the direction of the roll, pitch, and yaw axes in the body frame, respectively; and the last six symbols $\delta \eta _{L\ast }$ represent the unknown lever arm components. The accelerometer and gyroscope errors are modelled as random biases.

The system dynamic model is:

(21) $$\left[ \matrix{ \dot{{\bf x}}_{\bf {\rpsi}{\bf rv}} \cr \dot{{\bf x}}_{\bf {\rvarepsilon} \nabla} \cr \dot{{\bf x}}_{\bf {\reta}} } \right] = \left[\matrix{ {\bf F}_{{\bf 11}} \quad \,\, {\bf F}_{{\bf 12}} \quad \,\, {\bf 0}_{{\bf 9} \times 3} \cr {\bf 0}_{{\bf 6} \times {\bf 9}} \quad{\bf 0}_{{\bf 6} \times {\bf 6}} \quad{\bf 0}_{{\bf 6} \times 3} \cr {\bf 0}_{3\times {\bf 9}} \quad{\bf 0}_{3\times {\bf 6}} \quad{\bf 0}_{3\times 3} \cr } \right] \left[\matrix{ {\bf x}_{\bf {\rpsi}{\bf rv}} \cr {\bf x}_{\bf {\rvarepsilon}\nabla}\cr {\bf x}_{\bf {\reta}} \cr } \right]+\left[\matrix{ {\bf 0}_{{\bf 9} \times {\bf 1}} \cr {\bf \romega}_{\bf {\rvarepsilon} \nabla} \cr {\bf 0}_{3\times {\bf 1}} } \right]$$

where the sub-matrices F₁₁ and F₁₂ of system matrix F are calculated based on the INS error model (Arshal, Reference Arshal1987; Goshen-Meskin and Bar-Itzhack, Reference Goshen-Meskin and Bar-Itzhack1992), ${\bf \romega}_{\bf {\rvarepsilon} \nabla}$ is the noise of accelerometer and gyroscope errors, which is modelled as random bias.

Because of the lever arm uncertainty, the Locata position and velocity should be converted to the IMU centre by using the lever arm between the two points (Li et al., Reference Li, Li, Rizos and Xu2012). The relationship between the Locata antenna and IMU centre can be represented as:

(22) $${\bf r}_{{{\bf Locata}}}^{\bf {n}} ={\bf r}_{{{\bf INS}}}^{\bf {n}} +{{\bf C}}_{{\bf b}}^{\bf {n}} {\bf {\reta }}^{{\bf b}}$$

(23) $${{\bf v}}_{{{\bf Locata}}}^{\bf {n}} ={{\bf v}}_{{{\bf INS}}}^{\bf {n}} +\lpar {{\bf \romega}_{{{\bf ie}}}^{\bf {n}} \times +{\bf \romega}_{{{\bf en}}}^{\bf {n}} \times } \rpar {{\bf C}}_{{\bf b}}^{\bf {n}} {\bf {\reta }}^{{\bf b}}-{{\bf C}}_{{\bf b}}^{\bf {n}} \lpar {{\bf {\reta }}^{{\bf b}}\times } \rpar {\bf \romega}_{{{\bf ib}}}^{{\bf b}}$$

where ${{\bf C}}_{{\bf b}}^{\bf {n}}$ is the direction cosine matrix between the body frame and the navigation frame and ${\bf \romega}_{{{\bf ib}}}^{{\bf b}}$ is the angular-rate measurement obtained from the IMU.

Then the computed Locata position and velocity at the Locata antenna centre is given by:

(24) $$\hat{{\bf r}}_{{\bf Locata}}^{\bf {n}} = \hat{{\bf r}}_{{\bf INS}}^{\bf {n}} + \hat{{\bf C}}_{{\bf b}}^{\bf {n}} {\bf \reta}_{{\bf 0}}^{{\bf b}} - \hat{{\bf C}}_{{\bf b}}^{\bf {n}} \lpar {\bf \reta}_{{\bf 0}}^{{\bf b}} \times \rpar {\bf \rpsi} + \hat{{\bf C}}_{{\bf b}}^{\bf {n}} {\bf \rdelta \reta}^{{\bf b}}$$

(25) $$\hat{{\bf v}}_{{\bf Locata}}^{\bf {n}} = \hat{{\bf v}}_{{\bf INS}}^{\bf {n}} + \hat{{\bf C}}_{{\bf b}}^{\bf {n}} \left( {\bf \romega}_{{\bf ib}}^{{\bf b}} \times \right) {\bf \reta}_{{\bf 0}}^{{\bf b}} + \hat{{\bf C}}_{{\bf b}}^{\bf {n}} \left( {\bf \romega}_{{\bf ib}}^{{\bf b}} \times \right) {\bf \rdelta \reta}^{{\bf b}} - \left[ \left( \hat{{\bf C}}_{{\bf b}}^{\bf {n}} \left( {\bf \romega}_{{\bf ib}}^{{\bf b}} \times \right) {\bf \reta}_{{\bf 0}}^{{\bf b}} \right) \times \right]$$

where ${\bf {\reta}}_{\bf 0}^{{\bf b}}$ is the initial estimate of the lever arm.

Therefore, the measurement model can be written as:

(26) $$\matrix{ \left[ \matrix{ {\bf \rdelta}{\bf r}_{{\bf Locata}}^{\bf {n}} - {\bf \rdelta}{\bf r}_{{\bf INS}}^{\bf {n}} -{{\bf C}}_{{\bf b}}^{\bf {n}} {\bf \reta}_{{\bf 0}}^{{\bf b}} \cr {\bf \rdelta}{\bf v}_{{\bf Locata}}^{\bf {n}} - {\bf \rdelta}{\bf v}_{{\bf INS}}^{\bf {n}} -{{\bf C}}_{{\bf b}}^{\bf {n}} \left( {\bf \romega}_{{\bf ib}}^{{\bf b}} \times \right) {\bf \reta}_{{\bf 0}}^{{\bf b}} } \right] \hfill \cr \cr \quad =\left[\matrix{ \hat{{\bf C}}_{{\bf b}}^{\bf {n}} \left( {\bf \reta}_{{\bf 0}}^{{\bf b}} \times \right) \quad \quad \quad \quad {\bf I}_{{\bf 3} \times {\bf 3}} \quad {\bf 0}_{{\bf 3} \times {\bf 3}} \quad {\bf 0}_{{\bf 3} \times {\bf 6}} \quad \quad \hat{{\bf C}}_{{\bf b}}^{\bf {n}} \cr -\left[ \left( \hat{{\bf C}}_{{\bf b}}^{\bf {n}} \left( {\bf \romega}_{{\bf ib}}^{{\bf b}} \times \right) {\bf \reta}_{{\bf 0}}^{{\bf b}} \right) \times \right] \quad {\bf 0}_{{\bf 3} \times {\bf 3}} \quad {\bf I}_{{\bf 3} \times {\bf 3}} \quad {\bf 0}_{{\bf 3} \times {\bf 6}} \quad \hat{{\bf C}}_{{\bf b}}^{\bf {n}} \left( {\bf \romega}_{{\bf ib}}^{{\bf b}} \times \right) } \right] {\bf x} \left( {\bf t} \right) + {\bf \rupsilon} \left( {\bf t} \right)}$$

The measurement noise covariance in the system filtering is defined by the general accuracy of the computed Locata PV solution, which is described as:

(27) $${\bf R}=diag\left( {\lpar {0{\cdot}05} \rpar^2} \quad {\lpar {0{\cdot}05} \rpar^2} \quad {\lpar {0{\cdot}05} \rpar^2} \quad {\lpar {0{\cdot}1} \rpar^2} \quad {\lpar {0{\cdot}1} \rpar^2} \quad {\lpar {0{\cdot}1} \rpar^2} \right)$$

where the first three terms are the covariances of position (m ²), and the last three terms are the covariances of the velocity (m/s)².

3.1 Federated Kalman Filtering

The FKF was developed by Carlson and Berarducci (Reference Carlson and Berarducci1994) and Carlson (Reference Carlson1996). It is typically a two-stage data processing algorithm, comprised of one master filter and a set of local filters, as described in Figure 2. The local filters obtain the measurements from independent sensors and process in parallel, then the master filter is used to fuse the obtained local estimates and yield a global solution. This decentralisation characteristic of the FKF ensures good fault detection and isolating ability and a low computation load. However, the accuracy cannot compete with conventional centralised filtering (Li, Reference Li2014; Jiang et al., Reference Jiang, Li and Rizos2015). To improve the accuracy, the information sharing approach is applied in the FKF algorithm. This approach can divide the global state prediction and the covariance across the local filters to help ensure an optimal global estimation.

Figure 2. System configuration of FKF which contains one master filter and several parallel local filters.

Suppose the dynamic equation and observation equation of the local system are:

(28) $${\bf x}_i \lpar k \rpar ={{\bf F}}_i \lpar k \rpar {\bf x}_i \lpar {k-1} \rpar +{\bf \romega}_i \lpar k \rpar \!, i=1,2\cdots n$$

(29) $${\bf z}_i \lpar k \rpar ={{\bf H}}_i \lpar k \rpar {\bf x}_i \lpar k \rpar +{{\bf v}}_i \lpar k \rpar \!, i=1,2\cdots n$$

where both ${\bf \romega}_{i}\!\lpar k\rpar $ and v _i (k) are assumed to be zero-mean white Gaussian noises, and the covariance matrix are Q _i (k) and R _i (k), respectively.

If the local filter and the master filter are statistically independent, they can be combined to yield the global estimates using the following information equations, where inverse covariance is the “information matrix”:

(30) $$\hat{{\bf {P}}}_f^{-1} =\sum\limits_{i=1}^n {\hat{{\bf {P}}}_i^{-1} } + \hat{{\bf P}}_m^{-1}$$

(31) $$\hat{{\bf {P}}}_f^{-1} \hat{{\bf {x}}}_f =\sum\limits_{i=1}^n {\hat{{\bf {P}}}_i^{-1} \hat{{\bf {x}}}_i} + \hat{{\bf {P}}}_m^{-1} \hat{{\bf {x}}}_m$$

where the subscript “i”, “m”, and “f\!” denote the local filter, master filter and full filter, respectively.

For the typical FKF, the local filters and the master filter share the same system model (Carlson and Berarducci, Reference Carlson and Berarducci1994; Carlson, Reference Carlson1996). Thus the master filter and local filters’ estimates are correlated because they use a common dynamic system. To eliminate the correlation, the covariance of state error and process noise are set to their upper bounds:

(32) $$\left\{ \matrix{ \hat{{\bf x}}_i \lpar k \rpar = \hat{{\bf {x}}}_f \lpar k \rpar \cr {\bf F}_i \lpar k \rpar = {\bf F} \lpar k \rpar } \right. \quad i=1,2\cdots n,m $$

(33) $$\left\{ \matrix{ {\hat{{\bf {P}}}_i^{-1} \lpar k \rpar =\hat{{\bf {P}}}_f^{-1} \beta _i } \cr {{{\bf Q}}_i^{-1} \lpar k \rpar ={{\bf Q}}_f^{-1} \beta _i }} \right. \quad i=1,2\cdots n,m $$

where $\beta _{i} \lpar >0 \rpar $ is the information-sharing factor and it satisfies the condition:

(34) $$\sum\limits_{i=1}^{n,m} {\beta _i} = 1$$

According to the general KF updating process, the time updating is executed for both local filters and master filter:

(35) $$\hat{{\bf {P}}}_i \lpar {k\vert k-1} \rpar = {{\bf F}}_i \lpar k \rpar \hat{{\bf {P}}}_i \lpar {k-1} \rpar {{\bf F}}_i^T \lpar k \rpar +{{\bf Q}}_i \lpar k \rpar ; \quad i=1,2\cdots n,m$$

(36) $$\hat{{\bf {x}}}_i \lpar {k\vert k-1} \rpar = {{\bf F}}_i \lpar k \rpar \hat{{\bf {x}}}_i \lpar {k-1} \rpar ; \quad i=1,2\cdots n,m $$

As for the measurement propagation, local filters and the master filter are different:

(37) $${{\bf K}}_i \lpar k \rpar =\hat{{\bf {P}}}_i \lpar {k\vert k-1} \rpar {{\bf H}}_i^T \lpar k \rpar \left[ {{{\bf H}}_i \lpar k \rpar \hat{{\bf {P}}}_i \lpar {k\vert k-1} \rpar {{\bf H}}_i^T \lpar k \rpar +{\bf R}_i \lpar k \rpar } \right] ^{-{{\bf 1}}}; \quad i=1,2\cdots n$$

(38) $$\hat{{\bf {P}}}_i \lpar k \rpar = \left[ {{{\bf I}}-{{\bf K}}_i \lpar k \rpar {{\bf H}}_i \lpar k \rpar } \right] \hat{{\bf {P}}}_i \lpar {k\vert k-1} \rpar ; \quad i=1,2\cdots n $$

(39) $$\hat{{\bf {x}}}_i \lpar k \rpar =\hat{{\bf {x}}}_i \lpar {k\vert k-1} \rpar +{{\bf K}}_i \lpar k \rpar \lsqb {{\bf z}_i \lpar k \rpar -{{\bf H}}_i \lpar k \rpar \hat{{\bf {x}}}_i \lpar {k\vert k-1} \rpar } \rsqb ; \quad i=1,2\cdots n $$

(40) $$\hat{{\bf {P}}}_m \lpar k \rpar =\hat{{\bf {P}}}_m \lpar {k\vert k-1} \rpar $$

(41) $$\hat{{\bf {x}}}_m \lpar k \rpar = \hat{{\bf {x}}}_m \lpar {k\vert k-1} \rpar $$

The fusion algorithm can be applied to combine the above local and master results:

(42) $$\hat{{\bf {P}}}_f^{-1} \lpar k \rpar =\hat{{\bf {P}}}_m^{-1} \lpar k \rpar +\sum\limits_{i=1}^n {\hat{{\bf {P}}}_i^{-1} } \lpar k \rpar$$

(43) $$\hat{{\bf {x}}}\lpar k \rpar =\hat{{\bf {P}}}_f \lpar k \rpar \left[ {\hat{{\bf {P}}}_m^{-1} \lpar k \rpar \hat{{\bf {x}}}_m \lpar k \rpar +\sum\limits_{i=1}^n {\hat{{\bf {P}}}_i^{-1} \lpar k \rpar \hat{{\bf {x}}}_i \lpar k \rpar } } \right]$$

Then the correct global estimates are obtained.

3.2 FKF-based GNSS/Locata/INS Hybrid System

For a GNSS/Locata/INS hybrid system, both Locata and GNSS can provide independent position and velocity solutions. The two sets of position and velocity solutions are combined with INS sensor output separately to construct the measurements of the two local systems. As the measurements of the two local filters include INS information, and the two local filters apply the INS navigation model, the two local filters are correlated. The information-sharing factor-based FKF algorithm is thus applied. Both Locata/INS and GNSS/INS local filters estimate the nine motion parameters, INS sensor error and corresponding lever arms use their own observations. As both the Locata and GNSS local filters apply the INS navigation model, the dynamic equations of the two filter systems are the same. The ith ( $i=\hbox{Locata}$ , GNSS) local filter is described by the following equations:

(44) $${\bf x}\lpar k \rpar ={{\bf F}}\lpar k \rpar {\bf x}\lpar {k-1} \rpar +{\bf \romega}\lpar k \rpar$$

(45) $${\bf z}_i \lpar k \rpar ={{\bf H}}_i \lpar k \rpar {\bf x}_i \lpar k \rpar +{{\bf v}}_i \lpar k \rpar $$

where ${\bf z}_{i} \lpar k \rpar \in {\bf R}_{i}^{m}$ is the output of the ith sub-system, H _i (k) and v _i (k) are the observation matrix and the observation noise of the local systems.

The estimated state vector could be written as three sub-vectors: motion component, inertial sensor errors, and lever arm components, as in Equation (17). Since the Locata and GNSS antennae are installed in different positions, the lever arm correction must be applied. As in the indoor navigation mode, both Locata and GNSS antennae positions are corrected to the IMU centre using estimated corresponding lever arms. Therefore, the lever arm component consists of six states, and the estimated state vector has 21 navigation error states in total. The lever arm component in Equation (17) can be written as:

(46) $${\bf x}_{\bf {\reta }} =\left[\matrix{ {\delta \eta_{Lx} } \quad {\delta \eta_{Ly} } \quad {\delta \eta_{Lz} } \quad {\delta \eta_{Gx} } \quad {\delta \eta_{Gy} } \quad {\delta \eta_{Gz} } } \right]$$

Then the system dynamic model is of the same form as Equation (21), rewritten as below:

(47) $$\left[\matrix{ \dot{{\bf {x}}}_{\bf {\rpsi}{\bf rv}} \cr \dot{{\bf {x}}}_{\bf {\rvarepsilon}\nabla} \cr \dot{{\bf {x}}}_{\bf {\reta}} \cr } \right] =\left[\matrix{ {{{\bf F}}_{{{\bf 11}}} } \quad \, {{{\bf F}}_{{{\bf 12}}}} \quad \,\, {{\bf 0}_{{{\bf 9}}\times {{\bf 6}}} } \cr {{\bf 0}_{{{\bf 6}}\times {{\bf 9}}} } \quad {{\bf 0}_{{{\bf 6}}\times {{\bf 6}}} } \quad {{\bf 0}_{{{\bf 6}}\times {{\bf 6}}} } \cr {{\bf 0}_{{{\bf 6}}\times {{\bf 9}}} } \quad {{\bf 0}_{{{\bf 6}}\times {{\bf 6}}} } \quad {{\bf 0}_{{{\bf 6}}\times {{\bf 6}}} } \cr } \right]\left[\matrix{ {{\bf x}_{{\bf {\rpsi}{\bf rv}}}} \cr {{\bf x}_{{\bf {\rvarepsilon}}\nabla}} \cr {{\bf x}_{\bf {\reta}}} \cr } \right] + \left[\matrix{ {{\bf 0}_{{{\bf 9}}\times {{\bf 1}}}} \cr {{\bf \romega}_{{\bf {\rvarepsilon}} \nabla}} \cr {{\bf 0}_{{{\bf 6}}\times {{\bf 1}}}} } \right]$$

The linearized measurement equations of the Locata and GNSS local filters are given by:

(48) $$\matrix{ \left[\matrix{ {{\bf \rdelta}{\bf r}}_{{\bf L}}^{\bf {n}} - {{\bf \rdelta}{\bf r}}_{{{\bf INS}}}^{\bf {n}} -{{\bf C}}_{{\bf b}}^{\bf {n}} {{\bf \reta }}_{{{\bf L0}}}^{{\bf b}} \cr {\bf \rdelta}{\bf v}_{{\bf L}}^{\bf {n}} - {{\bf \rdelta}{\bf v}}_{{\bf INS}}^{\bf {n}} -{{\bf C}}_{{\bf b}}^{\bf {n}} \left( {\bf \romega}_{{\bf ib}}^{{\bf b}} \times \right) {\bf \reta}_{{\bf L0}}^{{\bf b}} } \right] \hfill \cr \cr \quad = \left[ \matrix{ \hat{{{\bf C}}}_{{\bf b}}^{\bf {n}} \left( {\bf \reta}_{{\bf L0}}^{{\bf b}} \times \right) \cr -\left[ {\left( {\hat{{{\bf C}}}_{{\bf b}}^{\bf {n}} \left( {{\bf \romega}_{{{\bf ib}}}^{{\bf b}} \times } \right) {{\bf \reta }}_{{{\bf L0}}}^{{\bf b}} } \right) \times } \right] } \quad \matrix{ {\bf I}_{{\bf 3} \times {\bf 3}} \cr {\bf 0}_{{\bf 3} \times {\bf 3}}} \quad \matrix{ {\bf 0}_{{\bf 3} \times {\bf 3}} \cr {\bf I}_{{\bf 3} \times {\bf 3}}} \quad \matrix{ {\bf 0}_{{\bf 3} \times {\bf 6}} \cr {\bf 0}_{{\bf 3} \times {\bf 6}}} \quad \matrix{ \hat{{\bf C}}_{{\bf b}}^{\bf {n}} \cr \hat{{\bf C}}_{{\bf b}}^{\bf {n}} \left( {{\bf \romega}_{{{\bf ib}}}^{{\bf b}} \times } \right)} \quad \matrix{ {\bf 0}_{{\bf 3} \times {\bf 3}} \cr {\bf 0}_{{\bf 3} \times {\bf 3}}} \right] \cr \cr \quad \quad \times {\bf x}\left( t \right) +{{\bf v}}\left( t \right) \hfill}$$

(49) $$\matrix{ \left[\matrix{ {\bf {\rdelta r}}_{{\bf G}}^{\bf {n}} -{\bf {\rdelta r}}_{{{\bf INS}}}^{\bf {n}} -{{\bf C}}_{{\bf b}}^{\bf {n}} {\bf {\reta }}_{{{\bf G0}}}^{{\bf b}} \cr {{\bf {\rdelta} {\bf v}}_{{\bf G}}^{\bf {n}} -{\bf {\rdelta} {\bf v}}_{{{\bf INS}}}^{\bf {n}} -{{\bf C}}_{{\bf b}}^{\bf {n}} \lpar {{\bf \romega}_{{{\bf ib}}}^{{\bf b}} \times } \rpar {\bf {\reta }}_{{{\bf G0}}}^{{\bf b}}} } \right] \hfill \cr \cr \quad = \left[ \matrix{ \hat{{{\bf C}}}_{{\bf b}}^{\bf {n}} \left( {\bf \reta}_{{\bf G0}}^{{\bf b}} \times \right) \cr -\left[ {\left( {\hat{{{\bf C}}}_{{\bf b}}^{\bf {n}} \left( {{\bf \romega}_{{{\bf ib}}}^{{\bf b}} \times } \right) {{\bf \reta }}_{{{\bf G0}}}^{{\bf b}} } \right) \times } \right] } \quad \matrix{ {\bf I}_{{\bf 3} \times {\bf 3}} \cr {\bf 0}_{{\bf 3} \times {\bf 3}}} \quad \matrix{ {\bf 0}_{{\bf 3} \times {\bf 3}} \cr {\bf I}_{{\bf 3} \times {\bf 3}}} \quad \matrix{ {\bf 0}_{{\bf 3} \times {\bf 6}} \cr {\bf 0}_{{\bf 3} \times {\bf 6}}} \quad \matrix{ {\bf 0}_{{\bf 3} \times {\bf 3}} \cr {\bf 0}_{{\bf 3} \times {\bf 3}}} \quad \matrix{ \hat{{\bf C}}_{{\bf b}}^{\bf {n}} \cr \hat{{\bf C}}_{{\bf b}}^{\bf {n}} \left( {{\bf \romega}_{{{\bf ib}}}^{{\bf b}} \times } \right)} \right] \cr \cr \quad \quad \times {\bf x}\lpar t \rpar + {\bf v} \lpar t\rpar \hfill}$$

As the measurements of each local filter are the position and velocity corrections from GNSS or Locata in three directions, this means the six identical components in the matrix represent the uncertainty of position and velocity solutions from GNSS and Locata. Thus, they are defined by the local Locata and GNSS position and velocity estimation errors, the corresponding covariance is:

(50) $${\bf R}_{{\bf L}} = diag\left( {\lpar {0{\cdot}05} \rpar^2} \quad {\lpar {0{\cdot}05} \rpar^2} \quad {\lpar {0{\cdot}05} \rpar^2} \quad {\lpar {0{\cdot}1} \rpar^2} \quad {\lpar {0{\cdot}1} \rpar^2} \quad {\lpar {0{\cdot}1} \rpar^2} \right)$$

(51) $${\bf R}_{{\bf G}} =diag\left( {\lpar {0{\cdot}15} \rpar^2} \quad {\lpar {0{\cdot}15} \rpar^2} \quad {\lpar {0{\cdot}05} \rpar^2} \quad {\lpar {0{\cdot}2} \rpar^2} \quad {\lpar {0{\cdot}2} \rpar^2} \quad {\lpar {0{\cdot}2} \rpar^2} \right)$$

where the first three terms are the covariances of position (m ²), and the last three terms are the covariances of the velocity m/s ².