Sunteți pe pagina 1din 39

Parametric Error Equations for Dead Reckoning

Navigators used in Ground Vehicle Guidance and


Control
David M. Bevly, Auburn University, Auburn, AL,
Demoz Gebre-Egziabher, University of Minnesota, MN
Bradford Parkinson, Stanford University, Palo Alto, CA

Received March 2006; Revised September 2006

Abstract

The performance of a GPS aided dead reckoning navigator for ground vehicle
guidance and control is evaluated. The dead reckoning system consists of a single axis
rate gyro for heading determination and a Doppler radar for speed measurements.
Heading drift due to rate gyro errors is kept in check using a multi-antenna carrier phase
GPS attitude determination system. Position drift due to heading and Doppler radar
speed measurement errors is bounded using position and velocity estimate from a
differential carrier phase GPS navigation system. The effects of the various sensor
errors and sensor sampling rate on position estimation error during momentary GPS
outages are evaluated. The analysis provides equations for the dead reckoning error
growth as a function of the sensor noise and sampling rate. It is shown that with
inexpensive rate gyros, sampling rate can have a significant effect on position error drift.
The results of the error analysis are confirmed experimentally by testing the dead
reckoning system on an automated farm tractor. It is shown that the system can provide
heading to within +1 deg and position estimation to within 0.3 m for control of the farm
tractor during 20-40 s GPS outages. It is also shown that at low sampling rates, the
primary source of position errors is random walk due to integration of wide band noise
from the rate gyros.

1
Introduction
In the past decade there has been an effort to develop systems which allow
autonomous operation of ground vehicles [1, 2]. This innovation is expected to enhance
productivity in applications such as farming [3, 4], construction [5, 6] and mining [7 8].
In some instances, autonomous operations are envisioned as an enhancement to safety [9,
10]. An indispensable part of any of these autonomous systems is precise knowledge of
the vehicle’s location at all times [11] and an enabling technology for such accurate
vehicle localization or navigation has been the Global Positioning System (GPS) [12, 13].
In the environments where ground vehicles operate, however, partial or complete loss of
the GPS signal is not uncommon. This is due to, in part, buildings and other tall
structures periodically obstructing the line-of-sight between a GPS antenna on a vehicle
and the GPS satellites. Electrical interference can also cause signal attenuation which, in
some instances, can lead to a loss in the GPS signal altogether. Therefore, successful
implementation of automated ground vehicles will be facilitated by techniques and
systems which make GPS robust to short outages or provide an alternate navigation
system during these outages. One possible solution is to fuse the information from GPS
with an inexpensive dead reckoning system which will provide a continuous navigation
solution through temporary GPS outages.
Dead reckoning systems are autonomous navigators which do not require external
signals to derive a position estimate. They can be mechanized by fusing heading
information derived from a rate gyro with speed measurements from a Doppler radar
[17], integrated accelerometer [18], or wheel speed sensors [19]. A short coming of dead
reckoning systems is the fact that sensor errors lead to position drift errors which grow
with time. It would be useful, therefore, to understand the position drift growth rate and
identify which types of sensor errors affect it most strongly. This, in turn, will allow
placing an upper limit on how long of a GPS outage can be tolerated for a given set of
sensors. This is the objective of the work discussed in this paper. Specifically, in this
paper we analyze and experimentally validate the performance of a GPS aided dead
reckoning navigator for use in an automated farm tractor. The effect of sensor errors and

2
sensor sampling rate on the length of the GPS outage that can be bridged by these dead
reckoning systems is quantified.

Prior Work
The idea of fusing GPS with inertial sensors is not new [20, 21]; augmentation of
GPS with inertial sensors for land navigation [14] and aviation [15, 16] has been studied
extensively. Additionally, Fiber Optic Gyros (FOGs) have been evaluated for dead
reckoning navigation of an indoor robot [22] and land vehicle navigation [23]. Fusion of
GPS and dead reckoning systems for autonomous land vehicles has also been studied [24,
25]. More recently, researchers have empirically compared the dead-reckoning accuracy
of two different tactical grade IMUs [36]. However, in each of these cases analytical
predictions of the dead reckoning accuracy as a function of the sensors being used has yet
to be developed.

Paper Contributions
In this paper we analyze the performance of an automotive dead reckoning navigator
during short GPS outages. The objective of the analysis is to characterize, qualitatively
and quantitatively, the effects of sensor errors, sensor sampling rate and vehicle dynamics
on the dead reckoning system’s navigation solution accuracy. To this end, one of the
primary objectives of the paper is to answer the following question: “Given a set of
inexpensive (or low accuracy) sensors, what is best navigation performance that can be
expected during short GPS outages when these sensors are used to mechanize a dead
reckoning navigator?” However, since metrics used to assess performance can be
different and depend on the specific application on hand, this paper develops and
experimentally validates a set of parametric error equations that can be easily tailored to a
given application and, thus, help answer the question about “best” performance. We
believe the significance of this work is that these parametric error equations can be used
as quick “rules of thumb” to assess the performance of GPS aided dead reckoning
navigators in automotive applications.

3
Paper Organization
The remainder of this paper is organized as follows: First, the problem statement we
are dealing with is defined mathematically. Next, we develop and discuss sensor error
models, which are used to qualitatively and quantitatively characterize the effect of the
sensor errors on the navigation solution. We then present experimental results validating
the simulation and analysis results discussed earlier in the paper. The paper ends with
our concluding remarks.

Dynamic Model and Problem Statement


Figure 1 shows the coordinate system used in the analysis presented in this paper
where we are interested in quantifying the lateral or y-position errors, y ε , and heading

errors,ψ ε , as a function of the various sensor errors and sensor sampling rate. The
sensors used are a Doppler radar and a Fiber Optic Gyro (FOG). The Doppler radar
provides speed measurements and the FOG provides angular rate measurements. These
rate measurements have to be numerically integrated to obtain an estimate of heading and
lateral position.

Figure 1. Vehicle and Coordinate System.

In this paper’s analysis we are only interested in the lateral position and heading
errors. As seen in Figure 1, the lateral dynamics (lateral velocity) can be calculated from
the vehicle heading (ψ) and velocity (V) at the vehicle’s center of mass as shown below.

y = V sin(ψ ) (1)
The estimate of vehicle heading (ψˆ ), which is the actual heading plus the heading error,
is calculated by integrating the yaw rate gyroscope on the tractor as shown below.

ψ = ψ +ψ ε = ∫ ψ gyro dt = ∫ (ψ +ν gyro )dt


 t t (2)
t0 t0

4
The term vgyro refers to the gyroscope sensor process noise. Similarly, the estimated
lateral position ( ŷ ) is determined by integrating the lateral velocity, defined in Equation
(1), which can be calculated using the estimated heading and vehicle velocity
measurements as shown below.
t t
( ) ( )
yˆ = y + y ε = ∫ V radar sin(ψˆ )dt = ∫ V + v radar sin ψ + ψ ε dt
t0 t0
(3)

The term vradar refers to the Doppler radar sensor process noise. Numerical values for
sensor noises as well as a justification for the structure of the sensor error models
assumed in Equation (2-3) will be discussed in the next section.

Sensor Error Models


The dead reckoning system discussed in this paper fuses information from a
Doppler radar and a rate gyro to derive position estimates when GPS is unavailable. In
this section we describe how mathematical error models for the sensors used in this
system can be developed. The objective of developing mathematical error models is so
that they can subsequently be used in trade-off studies evaluating the performance of the
dead reckoning system. Since, as noted in the introduction, our objective is to provide a
lower bound on the expected navigation performance, the error models we develop will
not be detailed nor will they allow predicting the actual sensor errors. Rather, they are
models that represent the best performance that can be expected from a sensor of a given
type of class. This is motivated, in part, by the authors’ experience that developing high
fidelity error models for low cost sensors such as the ones used in the navigator discussed
in this paper is not very practical; for a given brand of low cost sensors, let alone a class
of sensors, the output error model can be very different from sensor to sensor [27]. A
more meaningful approach, therefore, is to develop simple error models representing the
best performance that can be expected from a class of sensors.

5
Rate Gyro Error Model

A general model that can be used to describe the output of inexpensive rate
gyros is given by [27]:

ψ m
= ( 1 + k )ψ t
+ b( t ) (4)

In Equation (4), ψ m is the sensor’s measured output (measured angular rate) and

ψ t is the true value of the quantity that the sensor is measuring. This true value is
corrupted by a scale factor error, k, and a bias, b(t). The error term k represents the scale
factor errors. This is a deviation of the slope for the input to output mapping. As was
shown in [27], the scale factors are relatively constant over short time intervals. Once
they are estimated (either in real-time or in calibration) their change in response to
outside factors (such as temperature) over the short GPS outages time we are interested is
minimal. Thus, in this work we assumed them to be constants.

The bias term, b(t) ,has the following form:

b ( t ) = b0 + b1 ( t ) + b w ( t ) (5)

The term b0 represents a constant null-shift and is sometimes referred to as the


“turn-on” bias. The term b1(t) represents a time varying component of the bias and is
normally modeled as a random process. A model normally used for this term is a first
order Gauss-Markov process. The term bw(t) represents the sampling or output noise on
the sensor.

One quick way to determine numerical values for the various components of b(t)
is to use an Allan variance plot of the sensor output [27]. . The gyro used in this work is
a KVH FOG [26] and an Allan variance plot for its output when it is subjected to a zero
rate input is shown in Figure 2. The -1/2 slope of the Allan variance is an indication that
bw (wide band noise) is the dominant error source over short time intervals. Thus, for the
short GPS outage times of interest it is reasonable to model this gyros output error as:

b( t ) = bw ( t ) = v gyro (6)

6
Figure 2. Allan Variance Plot for the KVH Fiber Optic Gyroscope

Slopes other than -1/2 at larger averaging times would be indicative of the
presence of an error with low-frequency content. The b1(t) term captures these effects.
Since there are no such slopes for the KVH FOG gyro, we do not include b1(t) in the
error model. It should be noted that the b1(t) term can be large for many low cost inertial
sensors [27]. The above model used for gyro output error, therefore, should be
considered an ideal or best case error model.

The b0 term is excluded from the model for two reasons. First, the b0 term is a
constant null-shift that can be easily estimated from calibration data (like we did for the
radar) or using an observer. For the results that will be discussed later in the paper we
used the observer approach. Secondly, our interest in this paper is in the effect of
residual and stochastic errors on the navigation quality during short GPS outages. Since
a large part of b0 can be estimated (either in calibration or with an on-line observer), it is
not of interest in the error analysis that is discussed later in the paper.

The standard deviation or magnitude of this wide band noise term bw can be
determined from an Allan variance chart at an averaging time of 1 sec. Alternately, since
the band-limit for bw(t) is very high relative to the frequency content of b1(t), a numerical
value for bw(t) can be obtained by looking at the standard deviation of the sensor output
when it is subjected to a zero input and sampled at a rate much higher than the maximum
frequency content of b1(t). That is, the standard deviation can be computed from a time
series of the gyro outputs. The standard deviation for bw for the KVH FOG is given in
Table 1.

It is important to note that computing bw by the latter of the two methods


described above implies that bw is a white noise sequence (i.e., band limited) and not a
white noise process [28]. This assumption is valid because the output of many low cost
inertial sensors such as the KVH FOG used here is digital and has gone through
significant processing that the end user cannot control. Figure 3 is a schematic showing
the internals of a typical low cost rate sensor. From Figure 3, it can be seen that even

7
though the power spectrum of the sensor noise, Φψψ , may be white noise process at the

sensing element, after sampling and filtering the output power spectrum is clearly band
limited Because of this, as we will show later, unless we use special sensor pre-filtering
algorithms which adjust the amount of filtering as a function of sample rate, the sample
rate will have an important influence on the accuracy of the dead reckoning navigator.

Figure 3. Sensing Element and Noise Power Spectrum for a Typical Low Cost Rate
Gyro.

Doppler Radar Error Model

Doppler radars have been used for velocity measurements for some time now and
detailed error models for them can be found in references such as [36]. Using the
information in [36], one can write an error model for the Doppler radar output analogous
to Equation (4):

V measured
= (1 + k 1 + k 2 + " + k n )V True
+ b (t ) (7)

One of the differences between this general error model and the one given for the
rate gyro earlier is that the radar has more scale-factor-like errors. While this sensor error
model is general and appropriate to use when analyzing radars used in dead reckoning
systems which operate unaided for a long period of time, our interest here is in using a
dead reckoning system to bridge short GPS outages. Thus, consistent with what was
done for the rate gyros above, we can simply model the radar output errors as:

b ( t ) = b w ( t ) = v radar (8)

where v radar is assumed to be a white noise sequence. Even though this is a


simplification, it is a reasonable approximation for the work presented in this paper. To
see why this is the case, consider the plot of Doppler radar outputs as a function of time
shown in Figure 4. This data was collected from the Dickey-John Doppler radar, which
is used in the dead reckoning navigator analyzed in this paper, while the vehicle was
stationary. The figure shows that in addition to the sampling noise, there is a null-shift as

8
well as a time varying bias. The null-shift would be the largest error source over the
short propagation times on interest here. However, we do not include the null-shift in the
error model because it is calibrated out using an estimator like the one discussed in [31].
This leaves sampling noise, residual time varying bias (i.e., b1(t)) and other errors not
captured in Figure 4 such as, for example, errors due to scattering characteristics of the
ground below. Over the short propagation times of interest here (< ½ min.) these errors
will not be as significant as the null-shift. Furthermore, if all errors were modeled and
accounted for, there would still be sampling noise (a random process). Since the
overarching objective of this paper is to assess the best possible performance which can
be expected, this would be the case when the radar errors were only due to wide band
noise on the outputs.

Figure 4. Static Data from the KVH Fiber Optic Gyroscope and Dickey-John
Doppler Radar on the John-Deere Tractor.

Table I below summarizes the sensor error models developed above. Next, these
error models will be used in the performance analysis of the dead reckoning navigator.

Table I. Sensor Process Noise.

Error State Sensor Variance (σ2)

v gyro FOG 6.0 × 10−5 (rad/s)2


v radar Radar 2.5 × 10 −4 (m/s)2

9
Navigation Error Analysis
The error analysis can be formulated in the form of a covariance analysis relatively
easily. For completeness, we give the covariance analysis formulation here. However,
this does not give as much insight into the error growth as the algebraic/stochastic error
equations that will be discussed later. From Equation (2) it can be seen that the yaw error
dynamics are simply a function of the yaw gyroscope sensor noise as shown below.

ψ ε = ν gyro (9)

However, the lateral error dynamics in Equation (3) must be approximated in order to use
the covariance analysis. Assuming small yaw errors and that all of the radar sensor noise
is in the lateral direction (which is a conservative assumption) results in the following
approximation for the lateral error dynamics.
y ε ≈ V sin (ψ ε ) + v radar ≈ Vψ ε + v radar (10)

Since these errors are decoupled from the other system errors shown in Figure 1, we
can form a reduced order model that includes only these two error states. This model is a
kinematic relationship between speed, heading and position given by,

x = Ax + Bv sensor (11)
where,
 yε  0 V  1 0 sensor v radar 
x= ε , A=  , B = 0 1  , v =  gyro 
ψ  0 0    v 
In Equation (11), V is the speed of the vehicle’s center of mass, ψε is the heading
angle error, yε is the cross-track position error and vsensor is the sensor process noise
vector. Numerical values for the entries in vsensor and justification for the structure of the
sensor error models assumed in Equation (11) are from the previous section. Before we
processed with developing sensor error models, however, we should clarify the primary
objective of this paper in terms of Equation (11). Our goal is to develop an
understanding of the nature of the reduced set of navigation errors (yε, ψε). In particular
we are interested in knowing how vsensor, V and sensor sampling rate (or the rate at which
Equation (11) is integrated) affects the growth of yε and ψε.

10
The above reduced order error equation can be used to form the following state
covariance propagation (approximating the error growth due to integration of the yaw
gyroscope noise) [28]:
P = ΦPk Φ T + Qd (12)

where:

P
P =  11 =
[ ]
P12   E ( y ε )
2
[ ]
E y εψ ε 
 P21 P22   E y εψ ε
 [ ] [ ]
E (ψ ε ) 
2

Φ = discretized state matrix ( A)


For small sample rates (relative to the dynamics of the system) the discrete process noise
matrix (Qd) can be approximated by [29].
Qd ≈ ΓQc Γ T ≈ Ts2 BQc B T (13)

where:
[
Qc = E (v gyro )
2
]
Γ = discretized input vector ( B)
An exact solution to Qd can be obtained using the methods discussed in [30].
The covariance propagation given by Equation (12) can be used to study the
heading error, P22 , and lateral error, P11 , growth as a function of sensor noise.

However, a deeper insight into how these errors are influenced by sensor noise can be
obtained by expanding out Equation (12) and examining the resulting system of
equations. Alternately, we can arrive at the same covariance equations using an algebraic
and stochastic argument. This is nothing more than expanding out the covariance
equations and examining them as coupled algebraic equations. Even though this is not an
efficient way to derive covariances, we will nonetheless pursue this approach because, as
noted above, it gives insight into the relationship between sampling rate, sensor errors,
lateral errors and heading errors.

11
Heading Error Analysis

Consistent with Table I, a model for the rate gyro’s output is:
ψ gyro = ψ + v gyro (14)

where vgyro is serially uncorrelated white sequence representing sensor noise of the
gyroscope. Integration of the gyroscope to get heading results in:

∫ (ψ + v )dt = ∫ ψ dt + ∫ v
t
gyro
t t
gyro
dt = ψ + ψ ε (15)
t0 t0 to

Assuming perfect integration of the yaw rate, the heading error (ψε) is due only to the
integration of the sensor’s noise.
t
ψ ε = ∫ v gyro dt (16)
t0

Using an Euler integration routine to obtain heading from the yaw gyroscope results in
the following discrete equation for the heading error.
k
ψ kε +1 = ψ kε + Ts v kgyro = ∑ Ts vigyro (17)
i =1

where Ts is equal to the sample rate. Squaring and taking the expected value of the above
equation results in:

[ ( )
Eψ ε ψ ε
T
] = E[(∑ T v
s
gyro
)(∑ T v )
s
gyro T
] = T  ∑ E[(v
s
2
k

i =1
i
gyro
)(v )
i
gyro T
] (18)

Knowing that the expected value of the square of the gyroscope’s sensor noise is
simply the covariance of the gyroscope noise ( σ v2gyro ) results in:

σ ψ2 ε = Ts2 kσ v2 gyro = Ts tσ 2gyro (19)


v

The above equation results in the following equation for the heading error growth from
the integration of sensor noise as a function of time:
σψε = σ v gyro Ts t (20)

Figure 5 shows the integration of a static gyroscope ( σ v gyro = 0.44 deg/s) in the

tractor for 1 minute at 5 Hz and 100 Hz. As seen in the figure, Equation (20) adequately
describes the random walk due to discrete integration of the gyroscope noise.
Additionally, the effect of sampling rate on the heading error can be easily seen assuming
the same noise statistics for each sample rate. Figure 6 shows a Monte-Carlo simulation

12
(consisting of 5,000 iterations with the gyroscope noise values given in Table I) used to
validate the heading error growth.

The sampling rate dependence shown above is a result of the fact that we did not
do optimal pre-filtering based on the sample rate. It should be noted that, however, an
optimal pre-filter or anti-aliasing filter whose bandwidth is reduced with the sample rate
will provide more sensor filtering which would decrease the covariance of the sensor
noise as the sample rate is lowered. If this is the case, then the influence of sampling rate
on the lateral error growth observed above will not be an issue.

Figure 5 Integration of Gyroscope Noise.

Figure 6 Validation of the Heading Error Growth Using a Monte Carlo


Simulation.

Lateral Error Analysis for the Stationary Case

The above error analysis of [27] can be extended to the lateral dynamics of the tractor
to determine the effect of the discrete integration of the yaw rate and velocity noises on
the growth of lateral errors. Consistent with Table I, we assume that the radar output can
be written as:
V radar = V + v radar (21)
where: vradar is serially uncorrelated white sequence representing sensor noise of the
radar. Integration of the radar times sine of the heading angle to get lateral position (y)
results in:
t
( ) ( )
y + y ε = ∫ V + v radar sin ψ + ψ ε dt
t0
(22)

Assuming the mean velocity and heading are equal to zero, the lateral error (yε) is due to
the integration of the sensor’s noise times the heading error.

13
t
y ε = ∫ v radar sin(ψ ε )dt (23)
t0

Note that in this analysis we do not have to assume that all of the radar error is in the
lateral direction as was done in order to develop Equation (11). Again, an Euler
integration of the sensor outputs results in:
k
y kε +1 = y kε + Ts v kradar sin(ψ ε ) ≈ y kε + Ts v kradarψ kε = ∑ Ts viradarψ iε (24)
i =1

Squaring and taking the expected value of the above equation results in:

[
E yε yε ( ) T
] = E[(∑ T v radar
s i ψ iε )(∑ Ts viradarψ iε )
T
] (25)

Because the radar noise and heading errors are uncorrelated:

[
E yε yε ( )T
] = T  ∑ E[(ψ
s
2
k

i =1
ε
i )(ψ ) ε T
i ]E[(v radar
i )(v radar T
i ) ] (26)

Knowing that the expected value of the square of the radar’s noise is simply the
covariance of the radar noise ( σ v2radar ) and using the expected value for heading from

Equation (20):

[
E yε yε ( ) T
] = T  ∑ T  ∑ v
s
2
k

s
2
k
radar
j v radar
j
 2
 σ ν radar




 i =1   j =1   (27)
k k
=T s
4
∑ kσ ν
i =1
2
gyro σν 2
radar = T σ ν gyro σ ν radar ∑ k
s
4 2 2

i =1

Simplifying the above equation results in the covariance of the lateral error growth:
σ y2ε = Ts4σ v2 σ v2
radar gyro ( 1
2
)
k 2 + 12 k ≈ 12 k 2Ts4σ v2radar σ v2gyro = 12 (Ts t ) 2 σ v2radar σ v2gyro (28)

The above equation results in the following approximation for the lateral error growth
due to integration of the yaw gyroscope and radar noises:
σ yε ≈ Ts tσ v radar σv gyro
1 (29)
2

The above estimate of the lateral error growth is due only to integration of the noise
in the radar and gyroscope sensors, and only provides the correct estimation of the lateral
error growth when the vehicle is not moving (average velocity is equal to zero). This is
because the analysis assumes that both processes are zero mean. Figure 7 shows a 5,000
iteration Monte Carlo simulation to validate the lateral error growth at Vx=0 m/s. The
covariance of the radar and gyroscope measurement noises used in the simulation were

14
those given in Table I. Additionally, the plot contains the position error growth of the
tractor using the experimental data shown in Figure 4 from the gyroscope and radar while
the tractor was stationary. This data further validates that Equation (29) predicts the error
growth due to the discrete integration of the gyroscope and radar noise while the mean
velocity is zero.

Figure 7 Validation of the Lateral Error Growth for the Stationary Case
Using a Monte Carlo Simulation.

Lateral Error Analysis for the Non-Stationary Case

While the vehicle is moving, the lateral errors are not dominated by the integration of
radar noise, but rather the propagation of the position from the heading error found in
Equation (20). Therefore, utilizing the lateral error dynamics in Equation (1) and
substituting the heading error from Equation (20) results in a better description of the
lateral error growth. Neglecting the radar noise and assuming the mean heading is equal
to zero, the lateral error (yε) is due to the integration of the vehicle velocity times the
heading error.
t
y ε = ∫ V sin(ψ ε )dt (30)
t0

Again, using an Euler integration of the sensor outputs results in:


k
y kε +1 = y kε + TsV sin(ψ kε +1 ) ≈ y kε + TsVψ kε +1 = ∑ TsVψ iε+1 (31)
i =1

Substituting Equation (17) for the heading error into the above equation results in:
k k  k  k
y k +1 = ∑ TsVψ i +1
ε ε
= ∑ TsV ∑ v j  = Ts V ∑ kv kgyro
gyro 2
−i +1 (32)
i =1 i =1  j =1  i =1

Squaring and taking the expected value of the above equation results in:

[ ( )
E yε yε
T
] = E[T V (∑ kv
s
2 gyro
k − i +1 )(T V ∑ kv )
s
2 gyro T
k − i +1 ]
= T V  ∑ k E [(v )(v ) ] 
(33)

4 2
k
2 gyro gyro T
s i i
 i =1 

15
Assuming that gyroscope measurement noises are not serially correlated with each other:

[ ( )
E yε yε
T
] = T V E ∑ k (v
s
4 2
k

i =1
2
)  = T
gyro 2
k −i +1

s
4  k
 i =1
[(
V 2  ∑ k 2 E vigyro vigyro )( )
T
]  (34)

Substitution the covariance of the gyroscope noise into the above equation results in:

[ T
]
ε y2 = E y ε ( y ε ) = Ts4V 2 ∑ k 2σ ν2
k

i =1
gyro
k
=Ts4V 2σ ν2gyro ∑ k 2
i =1
(35)

Simplifying the above equation results in the covariance of the lateral error growth:
ε y2 = Ts4V 2σ v2 gyro ( 1
3 k 3 + 12 k 2 + 16 k ) ≈ 13 k 3Ts4V 2σ v2gyro = 13 V 2Ts t 3σ v2gyro (36)

The above equation results in the following approximation for the lateral error growth
due to integration of the velocity times the yaw gyroscope noise:

ε y ≈ Vσ ν gyro
1
3 Ts t 3 (37)

In reality, the total error growth is the sum of errors from the integration of sensor
noise (Equation (28)) and heading error (Equation (30)). However, the errors from
Equation (28) are small compared to errors from Equation (30) (for the noise covariance
values of the sensors used in this paper) and can therefore be neglected. Figure 8 shows a
5,000 iteration Monte Carlo simulation to validate the lateral error growth at Vx=2 m/s.
Again, the covariance of the radar and gyroscope measurement noise is given in Table I.

Figure 8 Validation of the Lateral Error Growth for the Non-Stationary


Case Using a Monte Carlo Simulation.

Additionally, a similar analysis can be performed for the longitudinal dynamics.


However, the growth of the lateral errors are a function of sine while the longitudinal
error growth is a function of one minus cosine. Therefore, for small heading errors, the
lateral errors grow as a function of the heading error and the longitudinal errors grow as a
function of the heading error squared. Because the positioning errors associated with the
lateral direction are much more significant than the errors in the longitudinal direction,
the longitudinal error analysis was neglected.

16
Changes in crab angle (or lateral velocity from side-slip) are equivalent to a heading
error that results in a growth of position errors as shown in Equation (30). Therefore, if
the crab angle is changing on the same order as the error in heading, the growth in
position errors can be as much as twice that obtained solely from heading errors.
Unfortunately, increased sample rate cannot decrease the errors associated with changes
in crab angle.
Errors in the gyroscope scale factor are negligible for straight-line tracking. However
for U-turn operation, a scale error as small as ½ percent results in a heading error of 0.9
deg. Additionally integration of the unknown bias will result in a linear growth of the
heading error and a parabolic growth of the lateral errors. Although the bias is being
estimated from the bias estimation discussed in [31, 32], the bias can change over periods
of time. However for short durations, the bias can be assumed constant, such that the
errors associated with the drift of the bias can be neglected. There are also additional
errors associated with the integration routine, which are a function of the routine and the
sample rate [33].

Experimental Dead Reckoning Results


As stated previously, the dead reckoning system was implemented on a John Deere
8400 tractor. The KVH fiber optic gyroscope (FOG) was used for sensing yaw rates and
an onboard Doppler radar was used for measuring velocity. The covariances of the noise
for these sensors are given in Table I. The tractor was given a straight-line trajectory to
follow. The majority of the experimental runs were performed at about 2 m/s (4.5 mph).
The variation on the actual speed and heading on the line trajectories was less than 0.2
m/s and 1 degree respectively. Time was given to allow the radar and gyro biases to be
estimated utilizing a cascaded Kalman Filter to estimate the navigation and control states
for the control the tractor [31, 32]. Once the bias estimates had settled, both GPS
position and attitude measurements were denied to the controller and estimation
algorithms. GPS measurements were still recorded to analyze the performance of the
dead reckoning algorithm.

17
Figure 9 shows a plot of the difference between the actual heading, measured by GPS,
and the estimated heading using the yaw gyroscope during a simulated GPS outage. The
uncertainty in the heading error (Equation (17)) is also shown. The tractor is equipped
with differential braking which gives the ability to add a heading disturbance to the
system by braking one side of the tractor. As seen in the figure, even under yaw
disturbances (from differential braking) during experimental runs, the growth in heading
error developed in Equation (20) describes the error quite well. Therefore, it can be
concluded that for short durations (on the order of a minute), the major error from
gyroscope integration at 5 Hz is due to the discrete integration of the sensor noise. This
also demonstrates that the estimation algorithm has correctly removed the gyroscope bias.
Integrating the gyroscope at a higher sample rate will reduce the heading error and
increase the performance of the dead reckoning system. Again, this is under the
assumption that any pre-filter does not exist or change, such that the covariance of the
sensor noise remains the same regardless of the sample rate. However, as errors from
integration of the gyroscope noise decrease (from increasing the sampling rate) other
errors may become more dominant.

Figure 9 Heading Error Following a Simulated GPS Outage.

Figure 10 shows the lateral error of three runs performing straight line tracking at a
typical tillage speed of 2 m/s. The figure plots the lateral error as a function of time after
GPS measurements were removed. Most of the runs had an estimation error within +0.3
m (1 foot) for over 40 seconds of a “GPS Outage.” This figure shows that for short
durations of GPS outages the estimator is still able to produce state estimates to control
the tractor. The lateral errors remained below 9 cm (half of a GPS Carrier-Phase cycle
length) for only 5-10 seconds as shown in the lower figure. The 9 cm error bound is what
is necessary to identify the Carrier-Phase integers upon reacquisition of the GPS signal.
However, continuous positioning accuracy of 0.3 m is adequate for some tillage farm
operations (since it is comparable to the average overlap of most human operators [34]).
Additionally, search algorithms can be used to produce integer solutions more rapidly by

18
providing a decreased search space from knowledge of the vehicle position to within 0.3
m.

Figure 10 Lateral Error Following a Simulated GPS Outage. The Top


Figure Shows the + 0.3 m (1 foot) Boundaries, While the Bottom Figure is
Zoomed In to Show the + λ2 Boundaries.

Figure 11 shows the difference between the estimated and actual positions for 42 s of
one data run. The inner circle represents a 9 cm error bound (half a GPS wavelength)
while the outer circle represents a 0.3 m error bound. Although the error has grown to 9
cm in only 8 s, the tractor has been controlled for 16 m. Therefore, the dead reckoning
error is on the order of 0.5 percent. Additionally, 40 s of control within 0.3 m also
represents 0.5 percent of dead reckoning error. The tractor is being controlled along a
linear trajectory. As mentioned previously, the lateral errors grow much faster than the
longitudinal errors. This can be seen by the fact that the position errors are along the
lateral axis. Figures 10 and 11 demonstrate the feasibility of being able to recover
integers after 10-15 m of dead reckoning at typical farm speeds of 2 m/s.

Figure 11 Dead Reckoning Position Errors After a Simulated GPS Outage


for a Single Dead Reckoning Experiment.

It is important to reiterate that the error bounds developed in this paper represent
conservative, best-case scenario, performance. The equations assume that the velocity
the Doppler radar measures is in direction the vehicle is headed (x-direction) such that
lateral velocity due to sideslip or crab angle is neglected. Figure 12 illustrates the effects
of three errors on the dead reckoning system: heading error from integration of the
gyroscope, errors due to unknown crab angle, and errors due to integration of the radar
sensor. The figure shows the difference in dead reckoning from using the gyroscope
heading, GPS heading, and GPS velocity heading or course. The direction of travel
(velocity heading) and GPS heading are not available during a GPS outage, but are
simply used to illustrate that the major error in lateral position can occur from changes in

19
crab angle (i.e., sideslip), which are not known during the outage. The crab angle is
being estimated while GPS is available. However, once GPS measurements are lost, the
crab angle can no longer be estimated and must be assumed to remain constant. In this
figure, the system was controlled along a line using heading estimation from the
gyroscope after GPS was turned off. The data was then post-processed using the GPS
and velocity headings for comparison. The velocity heading is obtained by looking at the
direction of two consecutive position measurements as shown below.
V east e − e k −1 (38)
ψ Vel = tan −1 ( ) = tan −1 ( k )
Vnorth n k − n k −1
As seen in the figure, the gyroscope heading is able to estimate the lateral position
nearly as well as using the actual vehicle heading (from GPS). This is because the
gyroscope can accurately estimate vehicle heading to within one degree for short (30-40
second) periods as seen previously. The difference in lateral error from using the
gyroscope estimated heading and the vehicle heading remains less than 0.3 m.

Figure 12 Lateral Estimation Errors due to Integrating Various Headings


(Demonstrating the Effect of Crab Angle on the Errors).

The dead reckoning system using GPS heading as well as the estimated heading from
the gyroscope results in large errors as the crab angle changes. The crab angle in this
experiment was constant for about 20 s, during which time the lateral errors are quite
small. However, after 20 s, the crab angle changed less than 1 deg (violating the
assumption that it remains constant), which led to a linear growth in the lateral error.
This change in crab angle could be attributed to a number of factors such as a tire getting
caught in a furrow. Additionally, crab angle has been shown to change with changes in
terrain side-slope [35]. If the exact vehicle velocity direction is known, the dead
reckoning system can estimate the lateral position very accurately (within 0.1 m during
the 80 s in Figure 12). The growth of errors from using the velocity heading (which
cannot be seen due to the scale of the figure) are due to the integration of the radar noise.
However, as mentioned previously, these errors become insignificant compared to the
lateral errors produced from errors in heading. If the crab angle remains constant, the

20
dead reckoning system can provide good accuracy. However, even on level ground,
disturbances and implements can produce small changes in crab angles that then become
the major cause of errors. Because the crab angle is being estimated in real time, the
variance of the crab angle over time can be monitored to provide a realistic estimate of
the dead reckoning errors during a GPS outage.
Figure 13 represents the use of the entire dead reckoning algorithm. The first lap
around the desired trajectory (shown by a dashed line) was performed clockwise by the
tractor under automatic steering control using GPS measurements. GPS measurements
were then removed and the tractor drove three additional laps around the track (requiring
over 3 min) with no position or attitude measurements. The decrease in dead reckoning
performance over time can be seen in the increased error at the end of the fourth lap
around the track.

Figure 13 Dead Reckoning Performance Around a Track.

The tracking errors for the first lap as well as the errors from the first two dead
reckoning laps are shown in Figure 14. Errors for the last lap grew substantially, due to a
change in the gyroscope bias, and are therefore not shown in the figure. The position and
heading errors listed in the figure are for the two laps with no GPS measurements. The
periodicity of the errors is due to an error in the gyroscope scale factor. As mentioned
previously, this is an avenue of future work.

Figure 14 Dead Reckoning Errors Around the Track.

21
Conclusions
The accuracy of a dead reckoning controller using low cost inertial sensors initialized
with Carrier Phase GPS was discussed in this paper. Experimental results were given
that demonstrate the capabilities of the inertial-aided dead reckoning system to maintain 9
cm accuracy for 5-8 s (or about 10-16 m). Results also show that the dead reckoning
system was able to maintain lateral errors within 0.3 m for about 30-40 s during a GPS
outage. Although 9-cm accuracy is necessary to reacquire integers, 0.3 m provides
adequate accuracy for many farm tillage operations (during a short GPS outage). It was
also shown that the integration of the gyroscope could estimate the vehicle heading to +1
deg during the 20-40 s outages. Therefore, INS can be coupled with GPS to provide
continuous control of a farm tractor with accuracies adequate for tillage operations
through the duration of any reasonably expected GPS outage.
Equations for the growth of heading and lateral errors were developed and shown to
predict the errors in experimental tests. The errors were seen to be a function of the
sample rate and 1σ sensor noise. Therefore increasing the sampling rate of the yaw
gyroscope and velocity sensor could decrease the errors (assuming the covariance of the
sensor noise did not change as a function of sample rate). Dead-reckoning performance
of a farm tractor can be predicted using the presented analysis. This will allow designers
to look at tradeoffs of using various sensors, as well as evaluate the merits of integrating
GPS with inertial sensors for control of farm equipment or other ground vehicles.
However, it was shown that changes in crab angle can become the largest contribution of
lateral errors of the dead reckoning system. Changes in crab angle cannot be estimated
during a GPS outage, and errors associated with crab angle are not a function of sample
rate.

Acknowledgements
22
The authors wish to thank Deere and Co. for their support of this research. Any

opinions, findings, conclusions, or recommendations expressed in this publication are

those of the authors and do not necessarily reflect the views of Deere and Co.

References
[1] Rajamani, R., Tan H. S., Law, B., and Zhang, W. B., Demonstration of Integrated Lateral and
Longitudinal Control for the Operation of Automated Vehicles in Platoons, IEEE Transactions
on Control Systems Technology, Vol. 8, No. 4, July 2000, pp. 695-708.
[2] Nieminen, T., Mononen, J., and Sampo, M., Unmanned Tractors for Agriculture Applications,
XII CIGR World Congress and AgEng ’94 Conference on Agricultural Engineering, Milano,
Italy, September 1994.
[3] O’Connor, M. L., Elkaim, G.H., and Parkinson, B., Carrier Phase DGPS for Closed-Loop
Control of Farm and Construction Vehicles, NAVIGATION, Journal of the Institute of
Navigation, Vol. 43, No. 2, Summer 1996, pp.167-278.
[4] Cordesses, et. al., CP-DGPS Based Combine Harvester Control without Orientation Sensor,
Proceedings of the 1999 ION-GPS Meeting, Nashville, TN, Sept. 1999.
[5] Crane, C. D. III, Armstrong, D. G. II, and Rankin, A. L., Autonomous navigation of heavy
construction equipment, Microcomputers in Civil Engineering, Vol. 10, No. 5, September
1995, pp. 357-370.
[6] Peyret, F., Betaille, D., and Hintzy, G., High-precision application of GPS in the field of real-
time equipment positioning, Automation in Construction, Vol. 9, No. 3, 2000, pp. 299-314.
[7] Zoschke, L. T. and Jackson, M. H., Experiences to Date with Unmanned Haul Trucks in Open
Pit Mines, Proceedings of the 2000 ION-GPS Conference.
[8] Seymour, C., GPS Based Machine Guidance in the Mining Industry, Australasian Institute of
Mining and Metallurgy Publication Series, No. 1, 2003, pp. 389-393.
[9] Gerdes, J. C., and Rossetter, E. J., A Unified Approach to Driver Assistance Systems Based On
Artificial Potential Fields, ASME Journal of Dynamic Systems, Measurement and Control,
Vol. 123, No. 3, September 2001, pp. 431-438.
[10] Bajikar, S., Gorjestani, A., Simpkins, P., and Donath, M., Evaluation of In-Vehicle GPS-Based
Lane Position Sensing for Preventing Road Departure, Proceedings of the IEEE Conference on
Intelligent Transportation System, ITSC 97, Boston, MA, 1997.
[11] Carlson, C., Gerdes, C., and Powell, J. D. Error Sources When Land Vehicle Dead Reckoning
with Differential Wheel Speeds, NAVIGATION, Journal of the Institute of Navigation, Vol.
51, No. 1, 2004. pp. 13-28.
[12] Sergi, M., Newstrom, B., Gorjestani. A., Shankwitz, C., and Donath, M., Dynamic Evaluation
of High Accuracy Differential GPS, Proceeding of the Institute of Navigation National
Technical Meeting, ION-NTM 2003, Anahiem, CA, January 2003.
[13] Misra, P., and Enge, P. Global Positioning System: Signals, Measurements and Performance,
Ganga-Jamuna Press, Lincoln, Massachusetts, USA, 2001.

23
[14] Abbot, E., and Powell, D. Land-Vehicle Navigation Using GPS, Proceedings of the IEEE,
January 1999, Vol. 87, No. 1, pp. 145-162.
[15] Da R., Dedes, G., and Shubert, K., Design and Analysis of a High-Accuracy Airborne GPS/INS
System, Proceedings of the 9th International Technical Meeting of the Satellite Division of The
Institute of Navigation (ION GPS-96), September 1996, pp. 955-964.
[16] Gebre-Egziabher, D., Hayward, R. C., and Powell, J.D., A Low-Cost GPS/Inertial Attitude
Heading Reference System (AHRS) for General Aviation Application, Proceedings of the 1998
IEEE Position Location and Navigation Symposium, April 1998, pp. 518-525.
[17] Rogers, R. M., Integrated DR/DGPS using Low Cost Gyro and Speed Sensor, Proceedings of
the 1999 ION-NTM.
[18] Ibrahim, F., Pilutti, T., Al-Holou, N., and Paulik, M., Accurate Gap Filling Using Properly
Initialized INS During Periods of GPS Signal Blockage, Proceedings of the 1999 ION-GPS
Conference.
[19] Kubo, Y., Kindo, T., Ito, A., and Sugimoto, S., DGPS/INS/Wheel Sensor Integrationfor High
Accuracy Land-Vehicle Positioning, Proceedings of the 1999 ION-GPS Conference.
[20] Berman, Z., and Powell, J. D., The Role of Dead Reckoning and Inertial Sensors in Future
General Aviation Navigation, Proceedings of the 1998 IEEE Position Location and Navigation
Symposium, April 1998, pp. 510-517.
[21] Masson, A., Burtin, D., and Sebe, M., Kinematic DGPS and INS Hybridization for Precise
Trajectory Determination, Proceedings of the 9th International Technical Meeting of The
Satellite Division of the Institute of Navigation (ION GPS-96), September 1996, pp. 965-973.
[22] Borenstein, J., Experimental Evaluation of a Fiber Optics Gyroscope for Improving Dead-
reckoning Accuracy in Mobile Robots, Proceedings of the 1998 IEEE International Conference
on Robotics & Automation, May 1998, pp. 3456-3461.
[23] Allen, D., Bennett, S., Brunner, J., and Dyott, B. R., Low-Cost Fiber Optic Gyro for Land
Navigation, Fiber Optic and Laser Sensors XIII, Vol. 2510, 1994, pp. 28-36.
[24] Schonberg, T., Ojala, M., Suomela, J., Torpo, A., and Halme, A., Positioning an autonomous
off-road vehicle by using fused DGPS and inertial navigation, International Journal of Systems
Science, 1996, Vol. 27, No. 8, pp. 745-752.
[25] Aono, T., Fujii, K., Hatsumoto, S., and Kamiya, T., Positioning of Vehicle on Undulating
Ground Using GPS and Dead Reckoning, Proceedings of the 1998 IEEE International
Conference on Robotics & Automation, May 1998, pp. 3443-3448.
[26] KVH Fiber Optic Group, Specification Sheet for KVH E-Core 1000 Fiber Optic Gyro, KVH
Industries, Inc., Tinley, Park, IL, 2001.
[27] Gebre-Egziabher, D., Design and Performance Analysis of a Low-Cost Aided Dead Reckoning
Navigator, Ph.D. thesis, Department of Aeronautics and Astronautics, Stanford University,
Stanford, CA, December 2001.
[28] Stengel, R., Optimal Control and Estimation, Dover ed, Dover Publications, Meneola, New
York, 1994.
[29] Franklin, G. F, Powell, J. D., and Workman, M.L., Digital Control of Dynamic Systems, 2nd
Edition, Addison Wesley, Menlo Park, CA, 1990.
[30] Bryson, A. E., Control of Spacecraft and Aircraft, Princeton University Press, Princeton, New
Jersey, 1994.
[31] Bevly, D., High Speed, Dead-Reckoning, and Implement Control for Automatically Steered
Farm Tractors Using GPS, Ph.D. thesis, Department of Mechanical Engineering, Stanford
University, Stanford, CA. September 2001.

24
[32] Bevly, D. M., Rekow, A., and Parkinson, B., Evaluation of a Blended Dead-Reckoning and
Carrier Phase Differential GPS System for control of an Off-Road Vehicle, Proceedings of the
1999 ION-GPS Meeting, Nashville, TN, September 1999.
[33] Chapra, S., and Canale, R., Numerical Methods for Engineers, 2nd Ed., McGraw-Hill, Inc,
1989.
[34] Palmer, R. J., and Matheson, S.K., Impact of Navigation on Agriculture, Proceedings of the
American Society of Agricultural Engineers, St. Joseph, Mich., 1988, paper no. 88-1602.
[35] Bell, T., et al., Automated Tractor Row and Contour Control on Sloped Terrain Using Carrier-
Phase Differential GPS, Fourth International Conference on Precision Agriculture, St. Paul,
MN, June 1998.
[36] Zhang, H.T., Petovello, M.G., and Cannon. M.E., Performance Comparison of Kinematic GPS
with Different Tactical Level IMUs, Proceedings of the 2005 ION-NTM, pp. 243-254.

25
Figure 1. Vehicle and Coordinate System.

26
0
10

KVH Autogyro
-1
10
σω ( o /sec)

-2
10

-3
10 0 1 2 3 4
10 10 10 10 10
Time (sec)
Figure 2. Allan Variance for the Fiber Optic Gyro

27
..
Φψψ ..
Φψψ

f f

Analog Digital
ψ
.
Filter Filter

Fiber Optic Analog to


Coil/Sensing Digital
Element Coverter

Figure 3. Sensing Element and Noise Power Spectrum for a Typical Low
Cost Rate Gyro

28
0.07

Yaw Gyroscope (deg/s)


0.06

0.05

0.04

0.03
0 50 100 150 200 250 300
Radar (m/s)

0.55

0.5

0.45
0 50 100 150 200 250 300
Time (s)

Figure 4. Static Data from the KVH Fiber Optic Gyroscope and Dickey-
John Doppler Radar on the John-Deere Tractor.

29
1.5

1 σ (5 Hz)
1.0

Estimated Heading (deg)


0.5
1 σ (100 Hz)

0.0
100 Hz

-0.5
5 Hz
-1.0

-1.5
0 10 20 30 40 50 60
T ime (s)

Figure 5 Integration of Gyroscope Noise.

30
1.6
σψ (Equation 20)
1.4 σ (Monte Carlo Simulation)
µ (Monte Carlo Simulation)
1.2

1
Heading Error (deg)

0.8

0.6

0.4

0.2

−0.2
0 10 20 30 40 50 60
Time (s)

Figure 6 Validation of the Heading Error Growth Using a Monte Carlo


Simulation.

31
1.2
σy (Equation 29)
σ (Monte Carlo Simulation)
1
µ (Monte Carlo Simulation)
Experimental Static Data
0.8
Lateral Error (mm)

0.6

0.4

0.2

−0.2
0 10 20 30 40 50 60
Time (s)

Figure 7 Validation of the Lateral Error Growth for the Stationary Case
Using a Monte Carlo Simulation.

32
2
ε (Equation 37)
y
σ (Monte Carlo Simulation)
µ (Monte Carlo Simulation)
1.5
Lateral Error (m)

0.5

−0.5
0 10 20 30 40 50 60
Time (s)

Figure 8 Validation of the Lateral Error Growth for the Non-Stationary


Case Using a Monte Carlo Simulation.

33
3


2


Heading Error (deg) 1

-1

-2

-3
0 20 40 60 80 100
T ime Following a GPS Outage (s)

Figure 9 Heading Error Following a Simulated GPS Outage.

34
1.5

1.0

Lateral Error (m)


0.5

0.0
0.3 m
-0.5

-1.0

-1.5
0 10 20 30 40 50 60
T ime Following a GPS Outage (s)

0.2

0.1 λ/2
Lateral Error (m)

0.0

-0.1
- λ/2

-0.2
0 5 10 15 20
T ime Following a GPS Outage (s)

Figure 10 Lateral Error Following a Simulated GPS Outage. The Top


Figure Shows the + 0.3 m (1 foot) Boundaries, While the Bottom Figure is
Zoomed In to Show the + λ2 Boundaries.

35
0.4
0-8 s
0.3
8-42 s 0.3 m
0.2

Longitudinal Error (m)


0.1

0.0

-0.1
9 cm = 1/2 λ
-0.2

-0.3

-0.4
-0.4 -0.2 0.0 0.2 0.4
Lateral Error (m)

Figure 11 Dead Reckoning Position Errors After a Simulated GPS Outage


for a Single Dead Reckoning Experiment.

36
0.5
w/ Vx Heading (from GPS)

0.0

Lateral Estimation Error (m)


-0.5
w/ GPS Heading

-1.0
w/ Gyro Heading

-1.5

-2.0
0 10 20 30 40 50 60 70 80
T ime Following a GPS Outage (s)

Figure 12 Lateral Estimation Errors due to Integrating Various Headings


(Demonstrating the Effect of Crab Angle on the Errors).

37
-795
Lap 1
-800

-805

-810
North (m)

1st Lap With GPS


-815
3 Laps DR (NO GPS)
-820

-825

-830
Lap 4
-835
-50 -30 -10 10 30 50
East (m)

Figure 13 Dead Reckoning Performance Around a Track.

38
1
DR East Errors:
Mean = 0.03 m
0.5
East Error (m)
1σ = 0.30 m

−0.5

−1
0 50 100 150 200 250 300

1
DR North Errors:
Mean = −0.04 m
0.5
North Error (m)

1σ = 0.47 m

−0.5

−1
0 50 100 150 200 250 300

2
Heading Error (deg)

−1
DR Heading Errors:
Mean = −0.49 deg
−2
1σ = 0.88 deg
−3
0 50 100 150 200 250 300
Time (s)

Figure 14 Dead Reckoning Errors Around the Track.

39

S-ar putea să vă placă și