Professional Documents
Culture Documents
ABSTRACT: Land-based mobile mapping systems operate in urban environments, where losses of GPS lock occur,
leading to often extended periods of free-navigation mode. The result is degraded navigation accuracy. Performing
periodic ZUPTs (zero velocity updates) enables calibration of velocity error sources, leading to a lower positioning
error rate as compared with unaided navigation mode. This paper presents a study of the impact of ZUPTs on navi-
gation accuracy in urban canyons, with special emphasis on the speed and reliability of ambiguity resolution sup-
ported by periodic ZUPTs. An analysis of the positioning errors after ZUPT events versus unaided and GPS/inertial
navigation system (INS) results is also presented. In addition, the observability aspects of static INS calibration are
briefly addressed. The results discussed here are based on investigation of a high-accuracy, tightly integrated
GPS/INS system providing georeferencing to a fully digital imaging component, designed for near-real-time highway
mapping.
217
Table 1 — Quality of INS Bridging During GPS Loss of Lock
Total Error (m)
Calibration
Period (s) East North Height
Fig. 3 – Errors in Heading, Pitch, and Roll for a 50 s Interval of the Fig. 4 – Error Growth in Free Navigation Mode Supported by a
Free Navigation Mode (0.001 rad 3.4 arcmin) ZUPT Event (time is shown in 10 s intervals)
Vol. 48, No. 4 Grejner-Brzezinska et al.: Bridging GPS Gaps in Urban Canyons: The Benefits of ZUPTs 219
AUTOMATIC ZUPT DETECTION IN g is the gravity anomaly vector; f is the specific
POSTPROCESSING MODE force vector; ie is the rotation vector from the earth-
fixed to the inertial frame; en is the rotation vector
In our current implementation of the GPS/INS fil-
from the navigation to the earth-fixed frame; is
ter, the ZUPT events are detected automatically
the rotation vector from the navigation to the iner-
during the data postprocessing by testing the hori-
tial frame; and a dot denotes the time derivative.
zontal velocity components and the gyro rates with
The dynamics matrix is obtained by a linearization
the tolerance levels shown in equation (1):
of equations (2) and augmentation by the sensor
vN vN0 3 vN dynamics, as shown in [16].
(1) Since the stationary INS calibration is performed
vN vE0 3 vE
at a known location, the position errors are no longer
In equation (1), the empirical mean of velocity unknown, and thus are removed from the state vec-
e s t i m a t e s vN0 vE0 0.0015 m /s,w ith vN vE tor. The Kalman filter measurement update uses
0.002 m /s estimated for a static LN-100. Similar north, east, and down velocity (vN, vE, vD) or acceler-
conditions are applied to the gyro rates, where the ation components. It should be noted that the velo-
empirical mean equal to 0.08 deg/s with a sigma of city tends to give a more robust solution. In general,
0.02 deg/s was estimated for a static LN-100. These because vD is very weakly coupled with the
values are monitored continuously during a ZUPT horizontal channels and because of the inherent sta-
event to detect the vehicle vibration due to gusty bility of the vertical channel, only horizontal velocity
winds, passing cars, or movement inside the vehicle. components are normally used in the static calibra-
A look-ahead filter is used to detect the ZUPT tion. Under the assumption of no sensor errors, hor-
events, and if the threshold for acceptance of static izontal channels should sense zero specific force. In
conditions is not exceeded for 3 s, the processing fil- the case of no vertical aiding, the observability of the
ter switches automatically to the ZUPT mode. vertical orientation errors (heading) and the vertical
gyro biases depends on the system’s sensitivity to the
earth’s rotation, which in general requires much
STATIC CALIBRATION OF THE INS ERRORS:
more time than estimation of the errors determined
AN OVERVIEW
by the zero horizontal velocity condition. Since the
The INS navigation and sensor errors cannot be gyro biases are especially detrimental to inertial
estimated in the unaided inertial navigation mode. navigation, at least these biases should be included
To partially control the INS errors, a static calibra- in the static calibration, if observable [13].
tion, which relies on the fact that the vehicle is sta- The translatory error equations (2), augmented by
tionary on a rotating earth, must be performed. The the sensor dynamics, form the state dynamics
primary objective of a static calibration (ZUPT) is matrix for a static calibration (equation (3)) under
estimation of the systematic errors of the IMU sen- the assumption that the position errors, vertical
sors, as well as the attitude errors (if observable) velocity errors, and gravity anomaly errors are
that are crucial to accurate inertial navigation. removed from the state, and the sensor errors are
Consequently, in the case of an integrated GPS /INS constant biases. Also, in a static case, the vertical
system, a calibrated INS will aid the ambiguity res- specific force (accelerometer readout), fD, is replaced
olution after the GPS lock has been reestablished. A by g (equations (3a) and (3b)). Even a simple inspec-
short discussion of the INS error observability in tion of equations (3) allows partitioning the state
static calibration is presented below, primarily for vector into observable and unobservable parts. For
those readers not fully familiar with this problem. It example, g E and N in equation (3a) are both con-
should be pointed out, however, that there exists a stants whose influence on vN cannot be distin-
vast literature on this subject. Interested readers guished. Similarly, g E and E in equation (3b) are
are referred to, for example, [3 – 13]. inseparable. Examination of equation (3c) indicates
The GPS/INS Kalman filter [14, 15] used in the that ieD E cannot be directly distinguished from N.
analysis presented here implements the psi-angle Moreover, all three terms on the right side of equa-
approach to the INS error model [11] and uses the tion (3d), as well as both terms in equation (3e), can-
velocity error version of the translatory error equa- not be uniquely observed. Clearly, only linear combi-
tion according to equation (2). nations or a selected subset of the states are observ-
vœ (ie ) v f g able; see, for example, [11, 13].
rœ en r v vœN 2 ieD vE g E N (3a)
œ (2) vœE 2 ieD vN g N E (3b)
œN ieD E N (3c)
where v, r, and are the velocity, position, and ori-
entation error vectors, respectively; is the
œE ieD N ieN D E (3d)
accelerometer error vector; is the gyro drift error; œD ieN E D (3e)
Fig. 6 – Horizontal Velocity RMS Difference Between ZUPT and Fig. 8 – RMS for Heading, Pitch, and Roll of ZUPT Solution (pitch
GPS/INS Solutions and roll RMS coincide)
Vol. 48, No. 4 Grejner-Brzezinska et al.: Bridging GPS Gaps in Urban Canyons: The Benefits of ZUPTs 221
needed to separate sensor errors from attitude
errors. However, vertical acceleration remains the
weakest component as compared with its horizontal
counterparts.
Since the figures presented thus far are all based
on the static dataset, no transition between the kine-
matic GPS /INS and the ZUPT is shown. To illustrate
such a transition, Figures 11 and 12 present plots of
the RMS for heading, pitch, and roll and the RMS for
gyro north and east biases, respectively, based on the
dataset of January 31, 2001. In the figures, GPS / INS
navigation is followed by a free navigation period
starting at epoch 50, followed by a ZUPT, which
starts at epoch 80. Clearly, pitch and roll are of com-
parable quality in kinematic and static parts of the
trajectory, while heading is weaker for the static part
(see Figure 11). The north gyro bias is still observ-
Fig. 9 – Gyro Bias RMS of ZUPT Solution
able during ZUPT, while the observability of the east
gyro is weaker, as expected (see Figure 12).
100 30 6 – 20 13 92 0 8
100–200 30 16 – 20 6 83 0 0
200 30 5 – 20 4 100 0 0
100 60 5 – 20 9 78 11 11
100–200 60 10 – 20 7 100 0 0
200 60 5 – 20 3 34 33 33
100 90 8 – 20 8 50 0 50
100–200 90 16 – 20 6 83 0 17
200 90 5 – 20 3 67 0 33
100 30 13 62 15 8 15
100 – 200 30 6 83 0 0 17
200 30 4 75 0 0 25
100 60 9 44 0 44 22
100 – 200 60 7 72 14 28 0
200 60 3 0 34 66 0
100 90 8 0 38 50 12
100 – 200 90 6 17 17 66 0
200 90 3 34 33 33 0
Vol. 48, No. 4 Grejner-Brzezinska et al.: Bridging GPS Gaps in Urban Canyons: The Benefits of ZUPTs 223
Table 5 — Ambiguity Resolution Statistics for GPS-Only Solution
Ambiguities Ambiguities Ambiguities Ambiguities
Fixed Within Fixed Within Fixed Within First Fixed Within
First 2 Epochs First 3–20 Epochs 21 – 80 Epochs 80 Epochs Not Fixed
Checkpoint # (percent) (percent) (percent) (percent) (percent)
17 0 82 6 6 6
ambiguity resolution time was determined after the Table 5 presents independent results for ambi-
signal reacquisition as a function of GPS / INS cali- guity resolution based only on GPS data. The clas-
bration time, duration of free navigation mode, and sical four-measurement filter technique was used
length of the ZUPT events (relevant only to the case to obtain the float ambiguities, and the search
presented in Table 3). However, it appears that the in the ambiguity domain followed as described in
ZUPT duration does not make a significant differ- [14]. A sample of 17 checkpoints was analyzed in
ence, as our tests indicate that any uninterrupted which the ambiguity resolution was restarted. The
ZUPT of ~20 s provides enough calibration time. results indicate that while the majority of the
Naturally for more error decorrelation, very long ambiguities can be resolved within 3 – 20 epochs,
calibration periods may be required, but for the pur- no fixes can be accomplished within the first
pose of supporting fast ambiguity resolution, we 2 epochs, and it takes more than 80 epochs to
found that a ~20 s ZUPT generally suffices. resolve the remaining ambiguities. However, 6 per-
As can be observed in Table 3, the majority of the cent of the ambiguities remain unresolved.
ambiguities are fixed to their integer values (based Evidently, ambiguity resolution supported by INS
on the INS prediction and verified with the GPS positioning results reinforced by the ZUPT option
information) within the first two epochs of signal provides the best results.
reacquisition, and all of them are fixed within the
first 90 epochs. The length of the GPS-based INS
CONCLUSIONS
error calibration prior to the free navigation mode
makes a difference, as expected based on the analy- To demonstrate the impact of static calibration on
sis presented earlier of the navigation error growth the overall quality of georeferencing by GPS/INS,
during free navigation under different conditions. this paper has examined the error state observabil-
Generally, the longer is the calibration, the shorter ity, the rate of navigation error growth, and the
is the ambiguity resolution time, since the longer speed of ambiguity resolution after a GPS gap. The
calibration time normally ensures slower error mechanism of ZUPT detection and the operational
growth during the free navigation mode. However, aspects of dealing with the switches of GPS / INS
the dynamics of the trajectory is also important, as to/from free INS navigation and to/from ZUPT have
indicated earlier; thus a longer calibration over a also been presented. Analysis of the impact of the
straight portion of the trajectory may show weaker ZUPT points on the overall positioning quality of
results as opposed to a shorter calibration over more the GPS / INS system and the speed of on-the-fly
varying geometry (for example, a calibration of ambiguity resolution has been discussed and com-
100 – 200 s shows better results compared with a pared with the corresponding results obtained
200 s calibration for the gaps of 60 and 90 s in without applying the ZUPT conditions.
Table 3). A brief analysis of the error observability in a
Table 4 presents ambiguity resolution statistics static INS calibration has been presented as well.
similar to those in Table 3, but without ZUPT events It has been shown that the (partial) calibration
following the free navigation mode. A comparison of achievable during ZUPTs enables a significant
the results presented in Tables 3 and 4 indicates decrease in positioning errors at the end of the free
that if no ZUPT points are present, significantly navigation mode. In general, implementation of
fewer ambiguities can be fixed within the first ZUPTs was shown to significantly increase the
2 epochs after GPS signal reacquisition, as com- quality and reliability of GPS / INS positioning by
pared with the case of ZUPTs following the free lowering the rate of error growth during GPS loss of
navigation mode. Also, while the results in Table 3 lock. Consequently, after signal reacquisition, vir-
indicate that all the ambiguities were solved after tually all ambiguities were found to be fixed within
90 epochs, Table 4 still shows some percentage of at most 90 epochs for all the tests presented,
unresolved ambiguities. In general, the longer is the with the majority being fixed within the first
free navigation mode and the larger is the error 2 epochs, while some of the ambiguities remained
growth (especially in cases not supported by ZUPT), unresolved in the case of the no-ZUPT or GPS-only
the more time is needed to fix the ambiguities. search technique.
Vol. 48, No. 4 Grejner-Brzezinska et al.: Bridging GPS Gaps in Urban Canyons: The Benefits of ZUPTs 225