APPARATUS AND METHOD FOR DATA-BASED REFERENCED NAVIGATION

Abstract

A method is provided. The method comprises: initializing a point mass filter; initializing the one or more Bayesian filters; obtaining measurement data associated with a horizontal position on a surface; obtaining measurement data of a horizontal position and a velocity; obtaining geo-mapping data; estimating, with the point mass filter, the horizontal position on the surface based upon the geo-mapping data and the measurement data; estimating, with the one or more Bayesian filters, remaining state parameters based upon an output of the point mass filter and the measurement data; predicting, with the point mass filter, an a priori horizontal position, on a surface for a future time when the next measurement data will be obtained; and predicting, with the one or more Bayesian filters, an a priori remaining state parameters for a future time when the next measurement data will be obtained.

Description

BACKGROUND

When position data from a global navigation satellite system is unavailable, a parameter, associated with a horizontal position (e.g. latitude and longitude) of a moveable object over a known surface such as terrain map, may be measured. Deterministic models may be used to determine the horizontal position.

A deterministic approach does not take into account model and sensor inaccuracies and noise. To address these shortcomings, non-linear Bayesian state estimators of stochastic dynamic systems may be used to estimate the horizontal position (and related parameters) of the moveable object on a known surface based upon a parameter associated with a horizontal position. However, a non-linear state estimator, based on a numerical solution of the Bayesian relations, i.e., a point mass filter, requires significant computation resources. The state estimator estimating three or more variables cannot be solved with typical data processing systems in a period of time consistent with application requirements. Therefore, there is a need to reduce the computational requirements for these non-linear estimators used to estimate the horizontal position (and related parameters) on a known surface based upon a measured parameter associated with the horizontal position.

SUMMARY

A method is provided. The method comprises: initializing a point mass filter; initializing the one or more Bayesian filters; obtaining measurement data associated with a horizontal position on a surface; obtaining measurement data of a horizontal position and a velocity; obtaining geo-mapping data; estimating, with the point mass filter, the horizontal position on the surface based upon the geo-mapping data and the measurement data; estimating, with the one or more Bayesian filters, remaining state parameters based upon an output of the point mass filter and the measurement data; predicting, with the point mass filter, an a priori horizontal position, on a surface for a future time when the next measurement data will be obtained; and predicting, with the one or more Bayesian filters, an a priori remaining state parameters for a future time when the next measurement data will be obtained.

DRAWINGS

Understanding that the drawings depict only exemplary embodiments and are not therefore to be considered limiting in scope, the exemplary embodiments will be described with additional specificity and detail through the use of the accompanying drawings, in which:

FIG. 1 illustrates an exemplary block diagram of a moveable object that includes a data-based referenced navigation system comprising a point mass filter and one or more Bayesian filters;

FIG. 2 illustrates an exemplary continuous probability density function and an exemplary discretized probability density function;

FIG. 3 illustrates an exemplary grid point selection function; and

FIG. 4 illustrates an exemplary method of operation of the data-based referenced navigation system comprising a point mass filter and one or more Bayesian filters.

In accordance with common practice, the various described features are not drawn to scale but are drawn to emphasize specific features relevant to the exemplary embodiments. Reference characters denote like elements throughout figures and text.

DETAILED DESCRIPTION

In the following detailed description, reference is made to the accompanying drawings that form a part hereof, and in which is shown by way of illustration specific illustrative embodiments. However, it is to be understood that other embodiments may be utilized and that structural, mechanical, and electrical changes may be made. Furthermore, the method presented in the drawing figures and the specification is not to be construed as limiting the order in which the individual steps may be performed. The following detailed description is, therefore, not to be taken in a limiting sense.

A Rao-Blackwellized point mass filter (RBPMF) is used to overcome the computational complexity related limitations of the point mass filter (PMF). The point mass filter is a non-linear Bayesian state estimator based on a numerical solution to the integral Bayesian recursive relations. RBPMF is an estimator that is an extension of a point mass filter, and permits a state variable model to be decomposed into non-linear state equation(s) and linear state equation(s). The non-linear state equation(s) can be solved by a standard point mass filter. The linear state equation(s) can be solved by one or more linear estimators, i.e. one or more Bayesian filters (BFs), which are much more computationally efficient. In one embodiment, the Bayesian filters are Kalman filters (KF). The term Kalman filters shall include all variations of Kalman filters, including for example extended Kalman filters.

This decomposition, therefore, significantly reduces the computational requirements of the point mass filter as the corresponding non-linear state equation(s) have fewer variables. As a result, the Rao-Blackwellized point mass filter can estimate, with typical data processing systems, the horizontal position (and other related parameters) of a moveable object on a known surface based upon a measured parameter associated with a horizontal position and geo-mapping data with much lower computational load than the point mass filter. Therefore, processing systems can perform the necessary computations in a period of time consistent with application requirements.

The Rao-Blackwellized point mass filter can be employed with any type of known surface, e.g. three dimensional bathymetric maps, barometric pressure maps, stellar maps, gravity maps, gravity gradient maps, magnetic maps, and terrain maps. In one embodiment, the map's horizontal position is parameterized by latitude and longitude. In another embodiment, such maps may be digitized. In a further embodiment, horizontal position and geo-mapping data are in reference to mean sea level.

FIG. 1 illustrates an exemplary block diagram of a moveable object 100 that includes, or to which is attached, a data-based referenced navigation system 102 comprising a point mass filter 106 and one or more Bayesian filters (BF) 108. The moveable object 100 can be a living creature (such as a human being or animal), a vehicle (such as an aircraft, a helicopter, a missile, a space craft, a rocket, a ship, a submarine, an automobile, and a truck), or anything for which position tracking is desirable.

The data-based referenced navigation system 102 comprises a processing system 109 coupled to an inertial navigation system (INS) 107 and one or more additional sensors (sensors) 105. The inertial navigation system 107 provides estimates of vector velocity and position. In one embodiment, the inertial navigation system 107 comprises one or more inertial measurement units (IMUs) coupled to a processor. In another embodiment, each of the one or more inertial measurement units includes one or more gyroscopes and one or more accelerometers.

The one or more sensors 105 measure data associated with a horizontal position on a known surface. In one embodiment, the one or more sensors 105 may include one or more of altimeters (e.g. barometric and/or radar), pressure sensors, accelerometers (e.g. measuring gravity), magnetometers, gravity gradiometers, gravimeters, water depth sensors, a bathymetric echo-sounder, a camera-type sensor, star trackers, and/or IMUs including accelerometer(s) and/or gyroscope(s).

The processing system 109 comprises a memory 110 coupled to a processor 112. An estimator 104, geo-mapping data 114, measurement data 115, and models 116 are stored in the memory 110. The estimator 104 includes the point mass filter 106 and the one or more Bayesian filters 108. The processing system 109 is configured to estimate the horizontal position on the surface based upon the geo-mapping data 114, the models 116, the measurement data 115, and the estimator 104. Horizontal position error may also be estimated as further described below.

The geo-mapping data 114 is a representation of a distribution of a parameter over the surface. In one embodiment, geo-mapping data 114 includes terrain geo-mapping data, bathymetric geo-mapping data, gravity field geo-mapping data, gravity gradient field geo-mapping data, magnetic field geo-mapping data, and image-based (e.g. stellar and/or terrestrial image-based) geo-mapping data. In another embodiment, the geo-mapping data 144 is stored in the form of a database, e.g. with records indexed based upon horizontal position such as latitude and longitude, and which store a corresponding surface parameter, such as terrain height, at that horizontal position. The measurement data 115 includes data measured by the inertial navigation system 107, and by the one or more sensors 105.

In one embodiment, the models 116, which may be in the form of executable software, include without limitation a state space model for direct and/or indirect estimation of the horizontal position. In another embodiment, the state space model comprises a state equation that is linear and a measurement equation that is non-linear. In a further embodiment, the models include models of the one or more sensors 105, the inertial navigation system 107, and the dynamics of the moveable object 100. In yet another embodiment, the state space model comprises a state equation and a measurement equation.

Typically, the state equation describes time-evolution of state parameters, e.g., of the horizontal position of the moveable object 100; of the error (e.g. bias, noise, or correlated noise) of the one or more sensors 105 and/or the inertial navigation system 107; and/or error (e.g. bias, noise, correlated noise) of the geo-mapping data 114. The measurement equation relates directly unmeasurable state parameters with available measurements of the one or more sensors 105, and the inertial navigation system 107. Definition of the state variables, selection of the sensor measurements, and design of the state-space model is performed by the filter designer.

One embodiment of the state space model, and the model for a direct estimation of the vehicle horizontal position, is illustrated below. The ‘state equation’ identified below is the linear state equation. The measurement equation is the non-linear equation.

$\begin{array}{cc}{x}_{k+1}=\left[\begin{array}{c}{x}_{k+1}^n\\ x_{k+1}^l\end{array}\right]=\left[\begin{array}{cc}F& \beta \\ \gamma & \alpha \end{array}\right]x_k+\left[\begin{array}{c}{u}_k\\ 0\end{array}\right]+w_k\left(\mathrm{state}\mathrm{equation}\right)z_k=h\left(x_k^n\right)+{\mathrm{Hx}}_k^l+v_k\left(\mathrm{measurement}\mathrm{equation}\right)& \left(\mathrm{Equation}1\right)\end{array}$

• • x_k+1 is the state for which time behavior is modelled by the state equation,
  • x_k+1ⁿ is a non-linear component of a state, the horizontal position of the moveable object 100, which is estimated with the point mass filter 106,
  • x_k+1^l is a linear component of the state, including parameters related to the horizontal position (e.g., horizontal velocity or acceleration), bias(es) (i.e. errors correlated in time) of the one or more sensors 105, the inertial navigation system 107 and/or the geo-mapping data 114, which are estimated with the one or more Bayesian filters 108. The correlated errors may be, for example, due to variations in weather conditions, such as temperature, local pressure, and due to the variations in altitude,
  • u_k is a known input vector, that is part of the state equation, and represents a change of horizontal position based upon the output of the inertial navigation system 107. In one embodiment, the change of horizontal position is determined by determining a vector velocity between two sampling time instants. In another embodiment, the change of the horizontal position is not assumed to be known and is considered to be zero,
  • z_k is part of the measurement equation, and is a parameter measured by the one or more sensors 105, e.g. altitude,

$F=\left[\begin{array}{cc}1& 0\\ 0& 1\end{array}\right]$

is a known system matrix of the non-linear part of the model of the state equation. This matrix is specified by the filter designer and corresponds to the expected type of motion of the moveable object 100. In one embodiment, the matrix can vary over time,

• • α is a known system matrix of the linear model of the state equation, and has elements determined based upon the properties of the states. This matrix is specified by the filter designer and corresponds to the expected type of motion of the moveable object 100, or to the properties of the one or more sensors 105 and/or the inertial navigation system 107. In one embodiment, the matrix can be vary over time,
  • β,γ are known system matrices describing relations between linear and non-linear parts of the model. These matrices are specified by the filter designer and correspond to the expected type of motion of the moveable object 100, or to the properties of the one or more sensors 105 and/or the inertial navigation system. In one embodiment,

$\beta =\left[\begin{array}{cc}0& 0\\ 0& 0\end{array}\right]\mathrm{and}\gamma =\left[\begin{array}{cc}0& 0\\ 0& 0\end{array}\right]$

In another embodiment, the matrices can vary over time,

• • h(x_kⁿ) is a known non-linear function of the non-linear component of the state vector x_kⁿ, which is typically represented by a map. In one embodiment, the function can vary over time,
  • H is a known matrix in the measurement equation relating the linear component of the state vector x_k^l and the measurement. In one embodiment, the matrix can vary over time,

$w_k=\left[\begin{array}{c}{w}_k^n\\ w_k^l\end{array}\right]$

is unknown noise in the state equation which quantifies uncertainty in the state equation and input vector. The probability density function (PDF) p_wk(w_k) is known. For example, it may be a Gaussian variable with zero mean and a known covariance matrix,

$Q=\left[\begin{array}{cc}{Q}^n& 0\\ 0& Q^l\end{array}\right],$

and

• • v_k is unknown noise in the measurement equation that quantifies all uncertainty in the measurement equation, measurements of the one or more sensors 105, and possibly uncertainty in the geo-mapping data. The noise is described by the known PDF p_vk(v_k). For example, it may have a Gaussian distribution with zero mean and known covariance matrix R.

The models 116 may alternatively include a state space model for indirect estimation of the horizontal position in lieu of the state space model for direct estimation of the horizontal position (described above). In this case, the point mass filter 106 does not directly estimate the horizontal position of the moveable object 100, but rather estimates the difference between the output of the one or more IMUS 107 and a true position. In one embodiment, the state space model comprises a state equation that is linear and a measurement equation that is non-linear. The state space model is illustrated below. The ‘state equation’ identified below is the linear state equation. The measurement equation is the non-linear equation.

$\begin{array}{cc}{x}_{k+1}=\left[\begin{array}{c}{x}_{k+1}^n\\ x_{k+1}^l\end{array}\right]=\left[\begin{array}{cc}F& \beta \\ \gamma & \alpha \end{array}\right]x_k+w_k\left(\mathrm{state}\mathrm{equation}\right)\begin{array}{cc}{z}_k=\tilde{h}\left(x_k^n+{\hat{x}}_k^{\mathrm{DR}}\right)+{\mathrm{Hx}}_k^l+v_k& \left(\mathrm{measurement}\mathrm{equation}\right)\\ =h\left(x_k^n\right)+{\mathrm{Hx}}_k^l+v_k& \end{array}& \left(\mathrm{Equation}2\right)\end{array}$

• • x_k+1 is the state for which time behavior is modeled by the state equation,
  • {circumflex over (x)}_k^DR is a horizontal position of the moveable object 100 provided by the inertial navigation system 107,
  • x_k+1ⁿ is a non-linear component of the state, a difference between the horizontal position of the moveable object 100, provided by the inertial navigation system 107, and the true position,
  • x_k+1^l is a linear component of the state bias(es) (or errors correlated in time) of the one or more sensors (105), the inertial navigation system 107, and/or the geo-mapping data 114,
  • z_k is a parameter measured by the one or more sensors 105, e.g. altitude,
  • F is a known matrix with elements determined on the basis of the properties of position estimate bias of the inertial navigation system 107 in longitudinal and latitudinal directions. This matrix is specified by the filter designer, and corresponds to the expected type of motion of the moveable object 100 and properties of the one or more sensors 105 and/or the inertial navigation system 107. In one embodiment, the matrix can vary over time,
  • α is a matrix with elements determined on the basis of the properties of the biases (or errors correlated in time) of the one or more sensors 105, the inertial navigation system 107 and/or geo-mapping data 114. This matrix is specified by the filter designer and corresponds to the expected type of motion of the moveable object 100,
  • β,γ are known system matrices describing relations between linear and non-linear parts of the model. These matrices are specified by the filter designer and correspond to the expected type of motion of the vehicle 100 or to the properties of the one or more sensors 105 and/or the inertial navigation system 107. In one embodiment,

$\beta =\left[\begin{array}{cc}0& 0\\ 0& 0\end{array}\right]\mathrm{and}\gamma =\left[\begin{array}{cc}0& 0\\ 0& 0\end{array}\right].$

In another embodiment, the matrices can vary over time, and

• • h(·) is a non-linear function represents the geo-mapping data 114, and maps the outputs of the inertial navigation system 107, and the bias of those outputs, to horizontal position on the map. In one embodiment, the he function can vary over time.

The estimated model parameters are represented by probability density functions (PDFs) which are typically continuous function. However, because modern computation systems operate discretely, the PDFs must also be discretized. Thus, a continuous PDF p_x(x) having an independent variable x must be approximated by a grid of equally spaced points ξ⁽ⁱ⁾, i=1, . . . , N, to create an approximate discretized PDF of the random scalar variable x:

$\begin{array}{cc}{p}_x\left(x\right)\cong \sum _{i=1}^NP^{\left(i\right)}\left({\xi }^{\left(i\right)}\right)S\left\{x\text{:}{\xi }^{\left(i\right)},\delta \right\},& \left(\mathrm{Equation}3\right)\end{array}$

where P⁽ⁱ⁾(ξ⁽ⁱ⁾)=p_x(ξ⁽ⁱ⁾), δ is the distance between two adjacent points defining resolution of the grid points, i.e., δ=ξ⁽ⁱ⁾−ξ⁽ⁱ⁻¹⁾, and S{x:ξ⁽ⁱ⁾, δ} is a selection function defined as:

$\begin{array}{cc}S\left\{x\text{:}\xi ,\delta \right\}=1\left[x-\left(\xi -\frac{\delta }{2}\right)\right]-1\left[x-\left(\xi +\frac{\delta }{2}\right)\right]& \left(\mathrm{Equation}4\right)\end{array}$

with 1[·] being a unit step function defined for a scalar variable y as:

$\begin{array}{cc}\begin{array}{c}1\left[y\right]=1,y\ge 0\\ =0,\mathrm{otherwise}.\end{array}& \left(\mathrm{Equation}5\right)\end{array}$

The selection function S{x:ξ,δ} has a value 1 in if

$x\in \left(\xi -\frac{\delta }{2},\xi +\frac{\delta }{2}\right),$

otherwise, it has a value 0. The integral of the selection function is:

∫S{x:ξ,δ}dx=δ.

For vector variables, the selection function, the vicinity δ of a grid point ξ⁽ⁱ⁾, and the volume Δ are defined analogously.

FIG. 2 illustrates an exemplary continuous PDF 202 and an exemplary discretized PDF 204. FIG. 3 illustrates an exemplary grid point selection function 300. FIG. 4 illustrates an exemplary method 400 of operation of the data-based referenced navigation system 102 comprising a point mass filter 106 and one or more Bayesian filters (BF) 108 which shall now be described. In block 402, initialize the filters, i.e. the point mass filter 106 and the one or more Bayesian filters 108, with an a priori horizontal position estimate. In block 404, obtain (or load) geo-mapping data 114 for a region proximate to the a priori horizontal position estimate, e.g. of the moveable object 100, and obtain (or load) measurement data 115 that includes at least one of data associated with a horizontal position on a surface, and horizontal position and velocity. In one embodiment, such measurement data 115 is obtained from the at least one of the one or more sensors 105 and/or the inertial navigation system 107.

In one embodiment, in block 405, determine if consistency testing is satisfied. Consistency testing evaluates whether the measurement data 115 obtained or originating from the one or more sensors 105, to be used in block 406, is in an expected range that is specified by the filter designer and the present properties of the actual state estimate. If such measurement data 115 is outside of the expected range, then proceed to block 408. If such measurement data 115 is within the expected range, then proceed to block 406.

In block 406, estimate horizontal position, e.g. of the moveable object 100, based upon measurement data 115 obtained from the one or more sensors 105, the geo-mapping data 114 loaded, and the a priori (or predicted) horizontal position estimate. Also, estimate of the remaining state parameters using the one or more Bayesian filters 108. In one embodiment, in block 408, predict, with the point mass filter 106, a new a priori horizontal position, e.g. of the moveable object 100, for a future time when the next measurement data 115 will be obtained. Also, predict the remaining state parameters using the one or more Bayesian filters 108. In another embodiment, return to block 404 and continue the process.

Each block in FIG. 4 will now be further described and exemplified. For block 402, the initial discretized PDF is a joint probability distribution function:

p(x₀ⁿ,x₀^l|z⁻¹)=p(x₀^l|x₀ⁿ,z⁻¹)p(x₀ⁿ|z^k−1)=p(x₀^l|z⁻¹)p(x₀ⁿ|z^k−1)

where:

p(x₀ⁿ|z⁻¹)=∫p(x₀ⁿ,x₀^l|z⁻¹)dx₀^l=p(x₀ⁿ)=N{x₀ⁿ:{circumflex over (x)}_0|−1ⁿ,P_0|−1ⁿ}

p(x₀^l|z⁻¹)=∫p(x₀ⁿ,x₀^l|z⁻¹)dx₀ⁿ=p(x₀^l)=N{x₀^l:{circumflex over (x)}_0|−1^l,P_0|−1^l}  (Equation 6)

where

• • N{x:{circumflex over (x)},P} denotes the Gaussian distribution of the random variable x with the mean {circumflex over (x)} and covariance matrix P,
  • p(x₀ⁿ|z⁻¹) is representative of a probability density function associated with an initial estimated horizontal position, e.g. of the moveable object 100, and
  • p(x₀^l|z⁻¹) is representative of a probability density function associated with an initial estimated bias (or an initial estimated error correlated in time) of the one or more sensors 105, the inertial navigation system 107, the geo-mapping data 114, and/or other states belonging to linear part of the state vector.

In block 402, the values stored in the grid points for the point mass filter 106 and the one or more Bayesian filters 108 are a priori estimates, e.g. by the processing system 109, of the horizontal position of the moveable object 100. The subsequent estimation, e.g. performed by the processing system 109, determines more accurately the horizontal position of the moveable object 100.

In one embodiment, the point mass filter 106 is initialized by approximating the point mass filter 106 with a discretized PDF corresponding to an initial estimated horizontal position of the moveable object 100:

$\begin{array}{cc}p\left(x_0^nz^{-1}\right)\cong \sum _iP_{0-1}^{\left(i\right)}\left({\xi }_0^{\left(i\right)}\right)S\left\{x_0^n\text{:}{\xi }_0^{\left(i\right)},\delta \right\}& \left(\mathrm{Equation}7\right)\end{array}$

where

• • ξ₀⁽ⁱ⁾ is the i-th grid point defined by a user,
  • P_0|−1⁽ⁱ⁾(ξ₀⁽ⁱ⁾)=p(ξ₀⁽ⁱ⁾|z⁻¹)=N{ξ₀⁽ⁱ⁾:{circumflex over (x)}_0|−1ⁿ,P_0|−1ⁿ} is the value of the initial discretized PDF at the point, (initial, predicted discretized PDF sampled at i-th grid point), and
  • S{x₀ⁿ:ξ₀⁽ⁱ⁾,δ} is the selection function which is a piece-wise constant function of variable x_kⁿ at the point ξ_k⁽ⁱ⁾ within the interval δ.

Each grid point represents a possible horizontal position of the moveable object 100 where probability is evaluated. The probability that the moveable object 100 is proximate to the position of the point ξ₀⁽ⁱ⁾ is equal to P_0|−1⁽ⁱ⁾(ξ₀⁽ⁱ⁾). The vicinity of the point ξ₀⁽ⁱ⁾ is defined by δ.

For block 402, in one embodiment, the one or more Bayesian filters 108 have an initial condition of:

p(x₀^l|z^k−1)=p(x₀^l)=N{x₀^l:{circumflex over (x)}_0|−1^l,P_0|−1^l}  (Equation 8)

Because of the selected factorization of the initial discretized PDF, i.e. p(x₀^l,x₀ⁿ|z⁻¹)=p(x₀^l|x₀ⁿ,z⁻¹)p(x₀ⁿ|z^k−1), the state of the linear part of the model is estimated by a set of one or more Bayesian filters 108, each associated with one grid point ξ₀⁽ⁱ⁾. Thus, N Bayesian filters are initialized by creating N initial discretized PDFs:

p(x₀^l|ξ₀⁽ⁱ⁾,z^k−1)=N{x₀^l:{circumflex over (x)}_0|−1⁽ⁱ⁾,P_0|−1^l(i)}=N{x₀^l:{circumflex over (x)}_0|−1^l,P_0|−1^l}  (Equation 9)

In one embodiment, the initial mean and variance at each grid point is the same. The approximate discretized initial PDFs can be visualized with grids representing horizontal position. Each grid point represents a possible horizontal position of the moveable object 100 and corresponding probability that the moveable object 100 is in the vicinity of that grid point.

Block 404 shall now be further discussed. The geo-mapping data 114 that is loaded in block 404 corresponds to the a priori estimate of the horizontal position of the moveable object 100 determined in block 402. The geo-mapping data 114 is a non-linear function of the horizontal position x₀ⁿ, represented in the state equation by the term.

One embodiment of block 406, position estimation, shall now be described. Description of the filtering step of the RBPMF is presented with respect to the direct model with zero matrices for β,γ. Block 406 comprises estimating horizontal position representing the non-linear part of the state vector and state components representing linear part of the state vector based upon measurement data 115 obtained (originating from the one or more sensors 105), the geo-mapping data 114 loaded, and the a priori horizontal position estimate. The horizontal position estimate is in the form of a discretized PDF p(x_kⁿ,x_k^l|z^k−1) and is determined by applying the point mass filter 106 and the one or more Bayesian filters 108 to the measurement data 115 obtained (originating with the one or more sensors 105), the geo-mapping data 114 loaded, and the a priori state estimate, including the horizontal position estimate.

The PDF p(x_kⁿ,x_k^l|z^k) is determined based upon the following factorisation:

p(x_k|z^k)=p(x_kⁿ,x_k^l|z^k)=p(x_k^l|x_kⁿ,z^k)p(x_kⁿ|z^k)  (Equation 10),

where p(x_kⁿ|z^k) is determined with the point mass filter 106, and p(x_k^l|x_kⁿ,z^k) is determined with the one or more Bayesian filters 108. The determination of non-linear and linear portions of the state space model respectively using the point mass filter 106 and the one or more Bayesian filters 108 shall now be described in further detail.Measurement Update of Non-Linear Portion of State Vector with Point Mass Filter

The output of the point mass filter 106 is updated based upon the measurement data 115 (originating from the one or more sensors 105), and the geo-mapping data 114. The measurement update of the non-linear portion of the state p(x_kⁿ|z^k) with the point mass filter 106 is performed as follows. In one embodiment, a measurement prediction is determined for each grid point ξ_k⁽ⁱ⁾ to determine a probability represented by a discretized PDF of measurement likelihood that the i-th grid point ξ_k⁽ⁱ⁾ maps to the measurement z_k (originating from the one or more sensors 105). The geo-mapping data 114 is evaluated at each ξ_k⁽ⁱ⁾, and the value h(ξ_k⁽ⁱ⁾) is determined for each grid point. The measurement likelihood is determined according to:

P _mLike,k ⁽ⁱ⁾ =N{z _k :h(ξ_k⁽ⁱ⁾)+H{circumflex over (x)}_k|k−1^l(i),HP_k|k−1^l(i)H^T+R},∀i

The estimated position is determined based upon the a priori horizontal position estimate, the measurement data 115 obtained (originating from the one or more sensors 105), and the portion of the geo-mapping data 114:

$\begin{array}{cc}p\left(x_k^nz^k\right)\cong \sum _iP_{kk}^{\left(i\right)}\left({\xi }_k^{\left(i\right)}\right)S\left\{x_k^n\text{:}{\xi }_k^{\left(i\right)},\delta \right\}\mathrm{where}\text{:}P_{kk}^{\left(i\right)}\left({\xi }_k^{\left(i\right)}\right)=c^{-1}P_{kk-1}^{\left(i\right)}\left({\xi }_k^{\left(i\right)}\right)P_{\mathrm{mLike},k}^{\left(i\right)}c=\sum _{j=1}^NP_{kk-1}^{\left(j\right)}\left({\xi }_k^{\left(j\right)}\right)P_{\mathrm{mLike},k}^{\left(j\right)}& \left(\mathrm{Equation}11\right)\end{array}$

c is a normalizing function so that the sum of the discretized PDFs equals one. In this step, each PMF grid point is weighted according its measurement likelihood P_mLike,k⁽ⁱ⁾ and predictive probability. Thus, the grid points remain unchanged, but their weights (in other words, the associated probability) are altered based upon the measurement data 115.

Once the discretized PDF p(x_kⁿ|z^k) is determined, various conditional statistics of the discretized PDF p(x_kⁿ|z^k) can be evaluated. The conditional mean and covariance matrices are determined according to:

$\begin{array}{cc}{\hat{x}}_{kk}^n=\sum _i{\xi }_k^{\left(i\right)}P_{kk}^{\left(i\right)}\left({\xi }_k^{\left(i\right)}\right)\Delta =\sum _i{\varpi }_{k|k}^{\left(i\right)}{\xi }_k^{\left(i\right)}P_{kk}^n=\sum _i{\varpi }_{kk}^{\left(i\right)}\left({\xi }_k^{\left(i\right)}-{\hat{x}}_{kk}^n\right){\left({\xi }_k^{\left(i\right)}-{\hat{x}}_{kk}^n\right)}^T\mathrm{where}\text{:}{\varpi }_{kk}^{\left(i\right)}=P_{kk}^{\left(i\right)}\left({\xi }_k^{\left(i\right)}\right)\Delta & \left(\mathrm{Equation}12\right)\end{array}$

The conditional mean {circumflex over (x)}_k|kⁿ in is an optimal estimate applying, e.g. a minimum mean square error analysis. A maximum a posteriori probability estimate can also be determined:

$\begin{array}{cc}{\hat{x}}_{kk\left(\mathrm{Max}\mathrm{AP}\right)}^n=\arg \underset{i}{\max }\left(P_{kk}^{\left(i\right)}\left({\xi }_k^{\left(i\right)}\right)\right)& \left(\mathrm{Equation}13\right)\end{array}$

Horizontal position error can be estimated based upon the covariance matrix.Measurement Update of Linear Portion of State Vector with Bayesian Filters

Additionally, for block 406, errors of at least one of one or more parameter sensors, the geo-mapping data, the one or more IMUs, and/or remaining elements of the linear state are estimated using the one or more Bayesian filters 108 based upon the output of the point mass filter 106 and the obtained measurement data 115 (originating from the one or more sensors 105). The measurement update of the linear portion of the state p(x_k^l|x_kⁿ,z^k) with the one or more Bayesian filters 108 is performed as follows. A Bayesian filter is applied for each grid point. Thus, the predictive estimates of the linear part of the state associated with each grid point are updated with the available measurement according to:

{circumflex over (x)} _k|k ^l(i) ={circumflex over (x)} _k|k−1 ^l(i) +K _k ⁽ⁱ⁾(z_k−h(ξ_k⁽ⁱ⁾)−{circumflex over (x)}_k|k−1^l(i))

P _k|k ^l(i) =P _k|k−1 ^l(i) −K _k ⁽ⁱ⁾ P _k|k−1 ^l(i) ,∀i,

where:

K _k ⁽ⁱ⁾ =P _k|k−1 ^l(i)(P_k|k−1^l(i)+R)⁻¹  (Equation 14)

Having the filtering estimates determined for each grid point, the global estimate of the linear part of the state space model can be determined according to:

$\begin{array}{cc}{\hat{x}}_{kk}^l=\sum _i{\varpi }_{kk}^{\left(i\right)}{\hat{x}}_{kk}^{l\left(i\right)}& \left(\mathrm{Equation}15\right)\\ P_{kk}^l=\sum _i{\varpi }_{kk}^{\left(i\right)}\left[P_{kk}^{l\left(i\right)}+\left({\hat{x}}_{kk}^{l\left(i\right)}-{\hat{x}}_{kk}^l\right){\left({\hat{x}}_{kk}^{l\left(i\right)}-{\hat{x}}_{kk}^l\right)}^T\right]\mathrm{where}\text{:}{\varpi }_{kk}^{\left(i\right)}=P_{kk}^{\left(i\right)}\left({\xi }_k^{\left(i\right)}\right)\Delta ,& \end{array}$

The relations can be understood as the relations for computation of the moments of a Gaussian mixture PDF.

Block 408 shall now be further discussed. In block 408, a predictive estimate is determined for a new a priori horizontal position, e.g. of the moveable object 100, for a future time when the next measurement data 115 from the one or more sensors 105 will be next obtained. The new a priori horizontal position is predicatively estimated based upon measurement data 115 from the inertial navigation system 107 and the state equation.

In one embodiment, a new grid is first constructed. Then, the point mass filter 106 is updated with the new grid. Finally, the one or more Bayesian filters 108 are updated with the new grid.

In one embodiment, a predictive discretized PDF p(x_k+1ⁿ,x_k+1^l|z^k) is determined according to the Chapman-Kolmogorov equation based upon the filtered discretized PDF p(x_kⁿ,x_k^l|z^k) based upon the following factorisation:

p(x_k+1|z^k)=p(x_k+1ⁿ,x_k+1^l|z^k)=p(x_k+1|x_kⁿ,z^k)p(x_k+1ⁿ|z^k)  (Equation 16),

where p(x_k+1ⁿ|z^k) is determined by the point mass filter 106, and p(x_k+1^l|x_k+1ⁿ,z^k) is determined by the one or more Bayesian filters 108. In this block 408, not only is the probability density function of the state estimate at each grid points modified, but so are the locations of the grid points (e.g. assuming that there is predicted estimated horizontal movement of the moveable object 100).

New Grid Construction

Predictively estimating a new a priori horizontal position (for the next time instant) begins with definition of new grid having grid points denoted as ξ_k+1^(j), j=1, . . . , M. The new grid may have the same or different number of grid points, and the same or different resolution δ_p, as the grid from time instant k (the old grid). Three typical situations may occur in design of the new grid depending on the number of significant grid points. The significant grid points are defined as those grid points which have the associated probability greater than a certain threshold. Number of significant grid points directly affects the quality of approximation of the continuous PDF by its discretized version, therefore, it affects the expected performance of the filter.

If the number of significant points is at a reasonable, threshold level, the number and resolution of the grid points may remain constant between the old and new grids. Then, the old grid points are just shifted based upon known input vector u_k (which is a change of horizontal position based upon the measurement data 115 from the inertial navigation system 107):

ξ_k+1^(j)=ξ_k^(j(+u_k  (Equation 17)

If the number of the significant points is below a reasonable, threshold level, then the number of grid points in the new grid is increased in comparison to the number of grid points in the old grid. Correspondingly, the resolution of the new grid points δ_p is smaller than the resolution of the old grid points δ_p to affect an increase in the number of the grid points and significant points in the new grid. Then, the new grid is shifted based upon known input vector u_k (which is a change of horizontal position based upon the output of the inertial navigation system 107):

ξ_k+1^(j)=ξ_k^(j)+u_k  (Equation 18)

If the number of the significant points is above a reasonable, threshold level, then then number of grid points in the new grid is decreased in comparison to the number of grid points in the old grid. Correspondingly, the resolution of the new grid points δ_p is larger than the resolution of the old grid points δ_p to affect a decrease in the number of the grid points and significant points in the new grid. Then, the new grid is shifted based upon known input vector u_k (which is a change of horizontal position based upon measurement data 115 from the inertial navigation system 107).

Determine Predictive Discretized PDF of Non-Linear Portion of State Vector with the Point Mass Filter

The predictive discretized PDF of the non-linear part of the state space model is determined by the following convolution:

$\begin{array}{cc}\begin{array}{c}{P}_{k+1k}^{\left(j\right)}\left({\xi }_{k+1}^{\left(j\right)}\right)=\sum _{i=1}^Np_{w^n}\left({\xi }_{k+1}^{\left(j\right)}-{\xi }_k^{\left(i\right)}-u_k\right)P_{kk}^{\left(i\right)}\left({\xi }_k^{\left(i\right)}\right)\Delta \\ =\sum _{i=1}^NN\left({\xi }_{k+1}^{\left(j\right)}\text{:}{\xi }_k^{\left(i\right)}+u_k,Q^n\right)P_{kk}^{\left(i\right)}\left({\xi }_k^{\left(i\right)}\right)\Delta \end{array}& \left(\mathrm{Equation}19\right)\end{array}$

where p_wn(w_kⁿ)≅p_wn(ξ_k+1^(j)−ξ_k⁽ⁱ⁾−u_k). Thus, the predictive discretized PDF is:

$\begin{array}{cc}p\left(x_{k+1}^nz^k\right)\cong \sum _iP_{k+1k}^{\left(j\right)}\left({\xi }_k^{\left(j\right)}\right)S\left\{x_{k+1}^n\text{:}{\xi }_{k+1}^{\left(i+1\right)},{\delta }_p\right\}& \left(\mathrm{Equation}20\right)\end{array}$

Determine Predictive Discretized PDF of Linear Portion of State Vector with the Bayesian Filters

The predictive discretized PDF of the linear part of the state space model is determined as follows. The predictive estimates, based upon the measurement data 115 from the inertial navigation system 107, of the linear part of the state space model and which correspond to the old grid points ξ_k^(j), are shifted forward in time according to:

{circumflex over (x)} _k+1|k ^l(i) =α{circumflex over (x)} _k|k ^l(i) ,∀i,

P _k+1|k ^l(i) =αP _k|k ^l(i)α^T+Q^l,∀i,  (Equation 21)

These predictive estimates are then transformed into the space of the new grid points ξ_k+1^(j). The transformation, which is based upon the relationship for moments of Gaussian mixture PDF s, is:

$\left(\mathrm{Equation}22\right)$ ${\hat{x}}_{k+1k}^{l\left(j\right)}=\sum _{i=1}^N{\varpi }_{k+1k}^{\left(i,j\right)}{\hat{x}}_{k+1k}^{l\left(i\right)},P_{k+1k}^{l\left(j\right)}=\sum _{i=1}^N{\varpi }_{k+1k}^{\left(i,j\right)}\left(P_{k+1k}^{l\left(i\right)}+\left({\hat{x}}_{k+1k}^{l\left(i\right)}-{\hat{x}}_{k+1k}^{l\left(j\right)}\right){\left({\hat{x}}_{k+1k}^{l\left(i\right)}-{\hat{x}}_{k+1k}^{l\left(j\right)}\right)}^T\right).\mathrm{where}\text{:}$ ${\hat{x}}_{k+1k}^{l\left(i\right)},P_{k+1k}^{l\left(i\right)}=\mathrm{estimates}\mathrm{related}\mathrm{to}\mathrm{old}\mathrm{grid}\mathrm{points}$ ${\hat{x}}_{k+1k}^{l\left(j\right)},P_{k+1k}^{l\left(j\right)}=\mathrm{estimates}\mathrm{related}\mathrm{to}\mathrm{new}\mathrm{grid}\mathrm{points}$ ${\varpi }_{k+1k}^{\left(i,j\right)}=\frac{{\omega }_{k+1k}^{\left(i,j\right)}}{\sum _{i=1}^N{\omega }_{k+1k}^{\left(i,j\right)}}=\mathrm{normalized}\mathrm{transition}\mathrm{weight}\mathrm{matrix}\mathrm{from}\mathrm{old}\mathrm{to}\mathrm{new}\mathrm{grid}$ $\begin{array}{c}{\omega }_{k+1k}^{\left(j\right)}=\sum _{i=1}^N{\omega }_{k+1k}^{\left(i,j\right)}\\ =\mathrm{probability}\mathrm{that}\mathrm{new}j\text{-}\mathrm{th}\mathrm{grid}\mathrm{point}\mathrm{is}\mathrm{reached}\\ \mathrm{from}\mathrm{all}\mathrm{old}\mathrm{grid}\mathrm{points}\end{array}$ $\begin{array}{c}{\omega }_{k+1k}^{\left(i,j\right)}=p_{w^n}\left({\xi }_{k+1}^{\left(j\right)}-{\xi }_k^{\left(i\right)}-u_k\right)\\ =N\left({\xi }_{k+1}^{\left(j\right)}\text{:}{\xi }_k^{\left(i\right)}+u_k,Q^n\right)=\mathrm{transition}\mathrm{weight}\mathrm{from}\mathrm{old}i\text{-}\mathrm{th}\mathrm{grid}\\ \mathrm{point}\mathrm{to}j\text{-}\mathrm{th}\mathrm{new}\mathrm{grid}\mathrm{point}\end{array}$

Although specific embodiments have been illustrated and described herein, it will be appreciated by those of ordinary skill in the art that any arrangement, which is calculated to achieve the same purpose, may be substituted for the specific embodiments shown. Therefore, it is manifestly intended that this invention be limited only by the claims and the equivalents thereof.

Claims

1. A data-based referenced navigation system, comprising: a processing system comprising a memory coupled to a processor;an inertial navigation system coupled to the processing system;at least one sensor coupled to the processing system;wherein the memory comprises geo-mapping data, models, measurement data, and an estimator;wherein the measurement data comprises data measured by at least one of the inertial navigation system, and the at least one sensor;wherein the estimator is a Rao-Blackwellized point mass filter which comprises a point mass filter, and at least one Bayesian filter; andwherein the processing system is configured to estimate a horizontal position on a surface based upon the geo-mapping data, the models, the measurement data, and the estimator.

2. The data-based referenced navigation system of claim 1, wherein the models comprise a state equation and a measurement equation.

3. The data-based referenced navigation system of claim 1, wherein the at least one Bayesian filter comprises at least one Kalman filter.

4. The data-based referenced navigation system of claim 1, wherein the inertial navigation system comprises at least one at inertial measurement unit coupled to a processor.

5. The data-based referenced navigation system of claim 1, wherein the at least one sensor comprises at least one of a barometric altimeter, a radar altimeter, an inertial measurement unit including at least one of an accelerometer and a gyroscope, a pressure sensor, a magnetometer, an accelerometer, a gravity gradiometer, a gravimeter, a water depth sensor, a bathymetric echo-sounder, a camera-type sensor, and a star tracker.

6. The data-based referenced navigation system of claim 1, wherein the geo-mapping data comprises at least one of terrain geo-mapping data, bathymetric geo-mapping data, gravity field geo-mapping data, gravity gradient field geo-mapping data, magnetic field geo-mapping data, stellar image-based geo-mapping data, and terrestrial image-based geo-mapping data.

7. The data-based referenced navigation system of claim 1, wherein the models comprise:

8. The data-based referenced navigation system of claim 7, wherein

9. The data-based referenced navigation system of claim 1, wherein the estimator is configured to: initialize the point mass filter;initialize the one or more Bayesian filters;obtain measurement data from at least one of the one or more sensors, and

10. A method of estimating horizontal position, comprising: initializing a point mass filter;initializing the one or more Bayesian filters;obtaining measurement data associated with at least one of a horizontal position on a surface, and a horizontal position and a velocity;obtaining geo-mapping data;estimating, with the point mass filter, the horizontal position on the surface based upon the geo-mapping data and the measurement data;estimating, with the one or more Bayesian filters, remaining state parameters based upon an output of the point mass filter and the measurement data;predicting, with the point mass filter, an a priori horizontal position, on a surface for a future time when the next measurement data will be obtained; andpredicting, with the one or more Bayesian filters, an a priori remaining state parameters for a future time when the next measurement data will be obtained.

11. The method of claim 10, further comprising determine if consistency testing is satisfied; if consistency testing is satisfied, then performing the estimating with the point mass filter, the estimating with the one or more Bayesian filters, the predicting with the point mass filter, and the predicting with the one or more Bayesian filters; andif consistency testing is not satisfied, then performing the predicting with the point mass filter, and the predicting with the one or more Bayesian filters.

12. The method of claim 10, wherein estimating, with the point mass filter, the horizontal position on the surface based upon the geo-mapping data and the measurement data comprises estimating, with the point mass filter, the horizontal position on the surface based upon the geo-mapping data and measurement data associated with the horizontal position on the surface; and wherein predicting, with the point mass filter, the a priori horizontal position, on a surface for the future time when the next measurement data will be obtained comprises predicting, with the point mass filter, the a priori horizontal position, on the surface based upon measured data from an inertial navigation system for the future time when the next measurement data will be obtained.

13. The method of claim 10, wherein predicting, with the one or more Bayesian filters, the a priori remaining state parameters for the future time when the next measurement data will be obtained comprises: constructing a new grid;updating the point mass filter with the new grid; andupdating the one or more Bayesian filters with the new grid.

14. The method of claim 10, wherein obtaining measurement data associated with at least one of the horizontal position on the surface, and the horizontal position and the velocity comprises obtaining measurement data that is measured altitude over the surface.

15. The method of claim 10, wherein initializing the one or more Bayesian filters comprises initializing mean and variance grids of one or more Kalman filters; and wherein estimating, with the one or more Bayesian filters, the remaining state parameters based upon the output of the point mass filter and the measurement data comprises estimating, with the one or more Kalman filters, the remaining state parameters based upon the output of the point mass filter and the measurement data; andpredicting, with the one or more Bayesian filters, the a priori remaining state parameters for the future time when the next measurement data will be obtained comprises predicting, with the one or more Kalman filters, the a priori remaining state parameters for the future time when the next measurement data will be obtained.

16. The method of claim 10, wherein initializing the point mass filter comprises initializing a probability density function of a point mass function grid; and wherein the probability density function of the point mass filter is represented by:

17. The method of claim 10, estimating, with the point mass filter, the horizontal position on the surface based upon the geo-mapping data and the measurement data comprises determining:

18. The method of claim 10, wherein estimating, with the one or more Bayesian filters, remaining state parameters based upon the output of the point mass filter and the measurement data comprises estimating, with the one or more Bayesian filters, errors of at least one of the one or more sensors, the map data, and the inertial navigation system based upon the output of the point mass filter and the measurement data; and wherein predicting, with the one or more Bayesian filters, the a priori remaining state parameters for the future time when the next measurement data will be obtained comprises predicting, with the one or more Bayesian filters, the a priori errors of at least one of the one or more sensors, the map data, and the inertial navigation system for the future time when the next measurement data will be obtained.

19. A system, comprising: a moveable object comprising a data-based reference navigation system;wherein the data-based reference navigation system comprises: a processing system comprising a memory coupled to a processor;at least one inertial measurement unit coupled to the processing system;at least one sensor coupled to the processing system;wherein the memory comprises geo-mapping data, models, measurement data, and an estimator;wherein the measurement data comprises data measured by the at least one inertial measurement unit, and the at least one sensor;wherein the estimator is a Rao-Blackwellized point mass filter which comprises a point mass filter, and at least one Bayesian filter; andwherein the processing system is configured to estimate a horizontal position.

20. The moveable object of claim 19, wherein the one or more sensors comprises at least one of a barometric altimeter, a radar altimeter, an inertial measurement unit including at least one of an accelerometer and a gyroscope, a pressure sensor, a magnetometer, a gravity gradiometer, an accelerometer, gravimeter, a water depth sensor, bathymetric echo-sounder, a camera-type sensor, and a star tracker; and wherein the geo-mapping data comprises at least one of terrain geo-mapping data, bathymetric geo-mapping data, gravity field geo-mapping data, gravity gradient field geo-mapping data, magnetic field geo-mapping data, stellar image-based geo-mapping data, and terrestrial image-based geo-mapping data.