Low-Cost Inertial Aiding for Deep-Urban

Tightly-Coupled Multi-Antenna Precise GNSS

James E. Yoder, Todd E. Humphreys
Radionavigation Laboratory, The University of Texas at Austin

Abstract—A vehicular pose estimation technique is presented accurate positioning technique that differences a receiver’s
that tightly couples multi-antenna carrier-phase differential GNSS observables with those from a nearby fixed reference
GNSS (CDGNSS) with a low-cost MEMS inertial sensor and station to eliminate most sources of measurement error [3,
arXiv:2201.11776v1 [eess.SP] 27 Jan 2022

vehicle dynamics constraints. This work is the first to explore the

use of consumer-grade inertial sensors for tightly-coupled urban Sec. 26.3]. Previous work by this paper’s authors probed the
CDGNSS, and first to explore the tightly-coupled combination of limits of unaided CDGNSS in the deep urban environment,
multi-antenna CDGNSS and inertial sensing (of any quality) for finding that the combination of a GNSS measurement engine
urban navigation. An unscented linearization permits ambiguity optimized for urban positioning and robust estimation tech-
resolution using traditional integer least squares while both niques for outlier exclusion make CDGNSS feasible in the
implicitly enforcing known-baseline-length constraints and ex-
ploiting the multi-baseline problem’s inter-baseline correlations. deep urban environment [4]. But the the unaided CDGNSS
A novel false fix detection and recovery technique is developed system described in [4] suffers from availability gaps of up to
to mitigate the effect of conditioning the filter state on incorrect 90 seconds in duration, making it insufficient to serve as the
integers. When evaluated on the publicly-available TEX-CUP sole navigation sensor for an AGV.
urban positioning dataset, the proposed technique achieves, with
consumer- and industrial-grade inertial sensors, respectively, a A natural solution to bridging such availability gaps is to
96.6% and 97.5% integer fix availability, and 12.0 cm and 10.1 incorporate measurements from an inertial measurement unit
cm overall (fix and float) 95th percentile horizontal positioning
error. (IMU). These measurements are uniquely valuable due to
their invulnerability to environmental effects such as radio
Index Terms—urban vehicular positioning; CDGNSS; low-cost interference and weather. Combined GNSS and inertial naviga-
RTK positioning.
tion systems that incorporate only GNSS position solutions as
measurements for a downstream navigation filter are termed
I. I NTRODUCTION loosely coupled, whereas tightly coupled systems directly
The rise of connected and automated vehicles has created a incorporate raw GNSS observables (pseudorange, Doppler, or
need for robust globally-referenced positioning with lane-level carrier phase) [3, Sec. 28.8]. While both loosely- and tightly-
(e.g., sub-30-cm) accuracy [1]. Much automated ground vehicle coupled aiding can bridge availability gaps, tightly-coupled
(AGV) research focuses on use of LIDAR and cameras for aiding additionally reduces these gaps’ frequency and duration:
navigation, but these sensing modalities often perform poorly the probabilistic constraint between GNSS measurement epochs
in low illumination conditions or during adverse weather such provided by the inertial sensor increases the success rate
as heavy fog or snowy white-out. By contrast, positioning of carrier phase integer ambiguity resolution and makes the
techniques based on radio waves, such as automotive radar navigation solution observable with fewer GNSS measurements.
or GNSS, are robust to poor weather and lighting conditions AGV navigation filter performance can be further improved
[2]. Recent work has found that fusing measurements from by tightly coupling with so-called vehicle dynamics constraints
low-cost automotive radars with inertial sensing can provide (VDCs). One such technique exploits the natural motion
lane-level accuracy in urban environments [2]. But radar-based constraints of four-wheeled ground vehicles, commonly referred
positioning in a global coordinate frame requires the time- to as non-holonomic constraints (NHCs). A second VDC
consuming and costly production and maintenance of radar technique infers a lack of vehicle motion by monitoring, for
maps. example, wheel odometry ticks, or by detecting a lack of road
GNSS signals provide a source of high-accuracy all-weather vibration, and enforces this constraint as a strong zero-velocity
absolute positioning that does not require expensive investment pseudo-measurement, called a zero velocity update (ZUPT) in
in systems for map production, storage, maintenance, and the literature.
dissemination. If the so-called integer ambiguities associated
with the carrier phase measurements can be correctly resolved, This paper extends the navigation filter component of the
carrier-phase based GNSS positioning offers exquisite accuracy. CDGNSS system described in [4] by tightly coupling with an
However, GNSS signal blockage, diffraction, and multipath inertial sensor and with vehicle dynamics constraints, and by
effects make this family of techniques extremely challenging to incorporating measurements from multiple vehicle-mounted
use in urban areas. Carrier-phase differential GNSS (CDGNSS), GNSS antennas. It also develops a novel robust estimation
whose real-time variant for mobile platforms is commonly technique to mitigate the effects of multipath and allow graceful
known as real-time kinematic (RTK) GNSS, is a centimeter- recovery from incorrect integer fixes.

Copyright © 2022 by James E. Yoder January preprint of paper submitted for review
and Todd E. Humphreys
A. Related Work suburban environments. More recently, [17] demonstrated
This subsection reviews relevant existing literature on urban tightly-coupled PPP using both a geodetic-grade and a low-
GNSS positioning, inertial aiding, vehicle motion constraints, cost GNSS receiver and an industrial-grade MEMS sensor
and multi-antenna CDGNSS. along an urban route in downtown Toronto, Canada, but
1) Unaided urban CDGNSS: Performance of CDGNSS only achieved meter-level accuracy when using the low-cost
unaided by inertial sensing in urban environments has his- GNSS receiver. A drawback of PPP-based positioning is
torically been poor. Experiments in [5] suffered from poor that the aforementioned results all required a roughly 10-
availability (ă60%) and large positioning errors (ą9m RMS) minute convergence period before producing an accurate
in suburban and urban environments. A 2018 assessment of navigation solution. Short-baseline CDGNSS positioning with
commercial CDGNSS receivers found that no low-cost solution a modern multi-frequency, multi-constellation receiver, by
offered greater than 35% fixed-integer solution availability in contrast, typically yields instantaneous initialization.
urban environments [6]. Li et al. [7] achieved a 76.7% unaided 4) Vehicle dynamics constraints: Recent research has also
correct integer fixing rate in urban Wuhan, China using dual- explored the tight coupling of CDGNSS measurements with
frequency CDGNSS with a professional-grade receiver. In 2019, vehicle dynamics constraints. Nagai et al. [18] found in
Humphreys et al. [4] achieved an unaided correct integer fix a simulation study using a realistic 3D map of an urban
rate of 84.8% in the urban core of Austin, Texas. environment that a tightly-coupled CDGNSS system using
2) Inertial aiding: Tightly-coupled inertial aiding has long GPS only could feasibly provide high-integrity decimeter-level
been employed as a method to increase CDGNSS solution positioning when aided with vehicle-dynamics constraints,
availability and robustness. Early systems built around highly- a tactical-grade IMU, and odometry based on wheel-speed
accurate but expensive tactical-grade IMUs were capable of sensors. Yang et al. in [19] tightly coupled single-antenna
providing robust positioning in dense urban areas [8]–[11]. The CDGNSS with non-holonomic constraints and a tactical-grade
recent emergence of inexpensive consumer- and industrial-grade fiber-optic IMU, but only evaluated their system under open-sky
micro-electromechanical systems (MEMS) inertial sensors has GNSS conditions with simulated GNSS degradations.
led to a new chapter of research in low-cost inertial aiding for 5) Multi-antenna CDGNSS: Use of multiple GNSS an-
urban CDGNSS. tennas on the vehicle for CDGNSS offers four advantages.
In 2018, Li et al. [7] demonstrated that tight coupling of First, the full six-degree-of-freedom vehicle pose (position
single-antenna professional-grade GNSS measurements with and orientation) becomes instantaneously observable when
an industrial-grade MEMS IMU increased the integer fix CDGNSS measurements are combined with the gravity vector
availability of single-frequency CDGNSS from 44.7% to 86.1% as measured by an inertial sensor. With a single GNSS antenna,
on a test route in urban Wuhan, China. However, the authors the vehicle yaw is observable only over multiple epochs, and
did not provide the GNSS dataset, information on the incorrect only if the vehicle accelerates during the observations [20].
integer fix rate, or a full error distribution, making these results Second, the shared reference antenna creates redundancy in
difficult to assess. the measurement model that allows better ambiguity resolution
This paper, in contrast, is the first to demonstrate an performance than any CDGNSS baseline taken individually
increased CDGNSS integer fix rate in an urban environment [21]. Third, the additional set of GNSS measurements at the
via tightly coupling with a consumer-grade inertial sensor. second antenna provides reduced position estimation error.
Furthermore, it incorporates vehicle dynamics constraints and Fourth, a highly effective method for GNSS spoofing detection,
multiple vehicular GNSS baselines. The system’s performance the multi-antenna defense [22], can readily be implemented.
is evaluated on a publicly-available urban positioning dataset, Multi-antenna GNSS has long been used for attitude deter-
allowing for head-to-head comparison of techniques by the mination applications with snapshot estimation methods such
urban positioning research community. as C-LAMBDA [23] and MC-LAMBDA [24], which provide
3) Tightly-coupled urban PPP: One disadvantage of globally-optimal single-epoch maximum-likelihood solutions
CDGNSS is that it requires observations from a nearby base to the full nonlinear GNSS attitude determination problem,
station to eliminate modeling errors (e.g., for atmospheric and have been successfully extended to the pose estimation
delays or satellite clocks and orbits) common to both the base case [25]. Other work has incorporated special cases of a
station (the reference) and the vehicle (the rover). Short-baseline priori attitude information into the nonlinear solution process
CDGNSS, which offers the greatest robustness against urban [26]. These snapshot methods, however, are computationally
multipath [12], is limited to reference-rover baseline lengths demanding, and their extension to recursive estimation for tight
below approximately 10 km [13]. To avoid the requirement coupling with other sensors is not straightforward and remains
for a nearby base station, attention has recently focused on unexplored.
extending precise point positioning (PPP), which is based on Fan et al. [27] found that a hard constraint using an a priori
precise orbit, clock, and atmospheric corrections, to urban areas known vehicle attitude to combine CDGNSS observations from
by tightly coupling with inertial sensors. multiple vehicle antennas can increase ambiguity resolution and
Rabbou et al. in 2015 explored tight coupling of PPP with urban CDGNSS performance. However, this method requires
a tactical-grade inertial sensor in mostly open-sky conditions a highly-accurate independent source of attitude information,
with simulated GNSS outages, achieving centimeter accuracy such as from an expensive gyrocompass-capable tactical-grade
[14]. References [15] and [16] extended tightly-coupled PPP IMU following an initial static alignment period.
to industrial-grade MEMS inertial sensors in highway and Medina et al. [21] proposed pose estimation based on

2
multiple vehicle antennas for inland waterway navigation. 5) A detailed evaluation and breakdown of the positioning
This work sidestepped the complexity of C-LAMBDA or MC- and ambiguity resolution performance contribution of
LAMBDA by linearizing the attitude model in an extended various sensors and algorithmic components (Section V-E).
Kalman filter (EKF) update and propagating the state with a
simple motion model. This formulation was found to increase II. C OORDINATE AND N OTATION C ONVENTIONS
ambiguity resolution performance over either the positioning or
attitude determination problems taken independently. However, A. Vector notation, sensor platform, and coordinate frames
the authors made no attempt to incorporate an inertial sensor Superscripts indicate the coordinate frames associated with
or additional motion constraints. vectors and rotation matrices. For example, r w denotes a vector
Hirokawa et al. [28] developed a multi-antenna GNSS r expressed in the w frame, and Rwb denotes a rotation matrix
system for aircraft pose estimation that tightly coupled with that converts vectors from their representation in the b frame
a MEMS inertial sensor, but only used CDGNSS for attitude to their representation in the w frame, i.e., r w “ Rwb r b .
measurements, relying on standard pseudorange measurements The sensor platform described in this paper and used
for the estimator’s position component. in the evaluation in Section V is the University of Texas
Henkel et al. [29] tightly coupled triple-antenna CDGNSS Sensorium [32], a roof-mounted vehicular perception platform
with an industrial-grade inertial sensor for a micro air vehicle incorporating multiple grades of inertial sensor, two GNSS
navigation application, but only evaluated the system’s perfor- antennas (denoted primary and secondary), stereo cameras, and
mance over a single, short test flight in open-sky conditions, three automotive radars. Only the inertial sensors and GNSS
and did not compare against a “ground truth” reference. antennas are used in this work.
Previous work [30], [31] by this paper’s authors explored a Several coordinate frames are referenced in this paper:
suboptimal “federated filtering” approach to the tightly-coupled u: The IMU frame is centered at and aligned with the IMU
multi-antenna CDGNSS + inertial problem, additionally in- accelerometer triad.
corporating monocular vision measurements in [30]. But the b: The body frame has its origin at the phase center of the
approach did not properly model the multi-antenna CDGNSS Sensorium’s primary GNSS antenna. Its x axis points
measurement update, instead resolving the position and attitude towards the phase center of the secondary antenna, its y
baselines separately. axis is aligned with the boresight vector of the primary
antenna, and its z axis completes the right-handed triad.
v: The vehicle frame is a body-fixed frame, centered at
B. Contributions the vehicle’s center of rotation as determined by an
This paper makes five contributions: offline calibration using GNSS and IMU data. Its x axis
points in the direction of vehicle travel with no steering
1) An estimation technique that tightly couples multi-antenna
angle deflection, its z axis points upwards, and its y axis
CDGNSS with vehicle dynamics constraints and inertial
completes the right-handed triad.
measurements. To the best of the authors’ knowledge,
w: The world frame is a fixed geographic East-North-Up
this paper is the first in the open literature to explore the
(ENU) frame, with its origin at the phase center of the
tightly-coupled combination of multi-antenna CDGNSS
reference GNSS antenna, which is located at a fixed base
and inertial sensing for navigation in urban environments.
station with known coordinates.
Furthermore, it is the first to explore the use of consumer-
grade inertial sensors for tightly-coupled deep urban Fig. 1 shows the relationships between these frames.
CDGNSS (Sections III and IV). ẑ w
2) A novel application of the unscented transform for the GNSS reference
multi-baseline CDGNSS integer ambiguity resolution and w ẑ v
measurement update step, which widens the operating ŷ w w ẑ b
x̂ ẑ u
regime of the filter to allow significantly greater attitude v
uncertainty without suffering from the excessive integer b x̂b v ŷ v
b u x̂
GNSS primary ŷ u
least squares (ILS) failures seen by existing EKF ap- x̂
proaches (Section III). IMU ŷ u
GNSS secondary
3) A novel false fix detection and recovery technique that
limits the degree to which an incorrectly-resolved integer
ambiguity can corrupt the tightly-coupled CDGNSS
estimator’s state (Section IV-D).
4) Demonstration of state-of-the-art deep urban CDGNSS per-
formance, achieving, by tightly coupling with consumer-
grade and industrial-grade inertial sensors, respectively,
a 96.6% and 97.5% integer fix availability, and 12.0
cm and 10.1 cm overall (fix and float) 95th percentile Fig. 1: Diagram of relevant University of Texas Sensorium
horizontal positioning error on the publicly-available TEX- coordinate frames.
CUP urban positioning dataset [32] (Sections V-A to V-C).

3
B. State representation and error-state filtering are rotation matrices and 3-1-2 Euler angles for the state and
tangent space, respectively, leading to the definitions
The tightly-coupled navigation estimator described in this » fi » fi
φ cψcθ ´ sφsψsθ cθsψ ` cψsφsθ ´cφsθ
work is an unscented Kalman filter (UKF) that recursively fuses
Exp : – θ fl ÞÑ – ´cφsψ cφcψ sφ fl
inertial measurements, double-difference GNSS pseudorange
ψ cψsθ ` cθsφsψ sψsθ ´ cψcθsφ cφcθ
and carrier phase measurements, and vehicle dynamics pseudo- » fi » fi
measurements. The estimator’s state at epoch k is given by the R11 R12 R13 arcsin R23
R
ordered set Log : –R21 R22 R23 fl ÞÑ –arctan ´R3313 fl
R22
R31 R32 R33 arctan ´R21
` ˘
xk “ rkw , vkw , Rkwb , bu u
ak , bgk where arctan denotes the 4-quadrant arctangent (i.e., atan2),
cx denotes cospxq, and sx denotes sinpxq. These attitude
where r w P R3 is the position of the u frame origin in the w parameterizations could, of course, easily be substituted with
frame; v w P R3 is the velocity of the u frame origin relative alternate parameterizations such as unit quaternions and axis-
to the w frame, expressed in the w frame; Rwb P SOp3q is the angle rotation vectors, with appropriate redefinition of the Exp
attitude of the b frame relative to the w frame; and bu u 3 and Log maps following [35].
a , bg P R
are the IMU’s accelerometer and gyro biases, respectively. Probabilistic beliefs under this framework are taken as
Gaussian distributions over the tangent space. Let Z k be the
Because the set of 3D rotations, which can be represented
set of all measurements up to time k. Then with xk denoting
using the special orthogonal group SOp3q, is not a vector
the true system state at k, define the a priori state x̄k , its
space, certain adaptations are needed to the typical Kalman
error covariance P̄k , the a posteriori state x̂k , and its error
filter equations to properly account for its manifold structure. A
covariance P̂k , as follows:
popular and well-founded method is to use a full, nonsingular
attitude parameterization (e.g., quaternions or rotation matrices) ∆ “ ‰
x̄k “ E xk |Z k´1 P X
in the filter state, but to express uncertainties, velocities, and ∆ “ ‰
small increments of the state using a minimal vector space P̄k “ E pxk a x̄k qpxk a x̄k qT |Z k´1 P RNx ˆNx
∆ “ ‰
parameterization that is local to the nominal state. Examples x̂k “ E xk |Z k P X
of this approach appear in the robotics literature as as the error ∆ “ ‰
state Kalman filter [33], and in the aerospace literature as the P̂k “ E pxk a x̂k qpxk a x̂k qT |Z k P RNx ˆNx
multiplicative EKF [34]. The remarkable feature of this notational framework is that
This paper adopts the conventions and notation of Solá et the typical Kalman filtering equations can be adapted for on-
al. [35], which appeals to Lie theory to unify and generalize manifold estimation by simply replacing ` and ´ with the ‘
the various methods so that the filtering equations are agnostic and a operators as needed.
to the specific choice of attitude parameterization. The filter
∆
state xk is a point on the composite manifold X “ R3 ˆ III. A N U NSCENTED M ULTI -BASELINE CDGNSS
3 3 3
R ˆ SOp3q ˆ R ˆ R , which has Nx “ 15 independent M EASUREMENT U PDATE
degrees of freedom. A state increment δxk is defined on the
tangent space of X at xk , which can be parameterized with A. CDGNSS measurement model
the vector space RNx . These spaces are related using the At each GNSS measurement epoch, the estimator ingests Nk
operators ‘ : X ˆ RNx Ñ X and a : X ˆ X Ñ RNx , which pairs of double-difference (DD) GNSS observables, each pair
correspond to standard addition and subtraction for vector- composed of a pseudorange and a carrier phase measurement,
valued components of xk and to more complex operations for across all baselines. The baselines and relevant relative position
the attitude component: vectors are shown in Fig. 2. The measurement vector at epoch
k is
» fi
rkw ` δrkw ∆ “ T T
‰
zgk “ ρT T T
1k , φ1k , ρ2k , φ2k P R2Nk
— vkw ` δv w ffi
` k wb ˘ffi
∆— wb
xk ‘ δxk “ — R ˝ Exp δR ffi P X where ρmk and φmk are vectors of double-difference pseu-
— k u k ffi
– bak ` δbu ak
fl dorange and carrier phase measurements, both in meters, for
bu
gk ` δb u
gk
baseline m P t1, 2u at epoch k. The measurement zgk is a
function of the state xk , the integer ambiguity vector nk P ZNk ,
rjw ´ rkw
» fi
w w and zero-mean white Gaussian measurement noise gk [36]:
´ vj ´ vk
— ffi
— ¯ffi
∆— wb ´1 ffi
wb
xj a xk “ —Log Rj ˝ Rk ffi P RNx zgk “ hgk pb pxk q , nk q ` gk , gk „ N p0, Σgk q (1)
— ffi
bu u
aj ´ bak The function bpxk q relates the baseline vectors bw w
1k and b2k
– fl
u u
bgj ´ bgk shown in Fig. 2 to the position and attitude components of the
state:
where ˝ denotes rotation composition, Exp : R3 Ñ SOp3q, and
„ w „ w ` b ˘
∆ b1k rk ` Rkwb rp ´ ˘rub
Log : SOp3q Ñ R3 . The estimator’s attitude parameterizations bpxk q “ w “ ` (2)
b2k Rkwb rsb ´ rpb

4
 each DD carrier phase measurement. The geometry matrix
Gmk is defined as
i »
w w T
fi
pr̂ik ´ r̂1k q
∆ — .. ffi
Gmk “ —– .
ffi
fl
` w w
˘ T
w
j r̂ik ´ r̂Nmk k
rrik
for pivot satellite i and non-pivot satellites 1 to Nmk , where
w w w
rrjk rpjk r̂jk denotes a unit vector in the w frame directed from a GNSS
antenna to GNSS satellite j of baseline m at epoch k. Under
w
rpik the small-angle approximation, these unit vectors are assumed
bw
1k w
rsik w
rsjk to be approximately equal for all receiver antennas involved:
Reference w
rrik w
rpik rw
w
Primary r̂ik « « w « sik
bw
2k
w
krrik k krpik k w k
krsik
Secondary
The nonlinearity of (2) due to the manifold structure of
vehicle attitude Rkwb is nontrivial. Optimal snapshot estimators
for the nonlinear GNSS attitude problem, such as C-LAMBDA
Fig. 2: Baseline and antenna-to-satellite vectors of the multi- [23] and MC-LAMBDA [24], have been studied, but these
antenna CDGNSS measurement model, for pivot satellite i estimators are computationally expensive and their extension to
w w w
and non-pivot satellite j. rrnk , rpnk , and rsnk refer to vectors recursive filtering does not seem straightforward. Instead, extant
pointing from the reference, primary, and secondary GNSS Kalman filtering-based multi-baseline CDGNSS estimators
antennas, respectively, to the antenna phase center of GNSS [21], [29] typically linearize the baseline measurement model
satellite n. about the current state estimate with a simple first-order Taylor
expansion in order to perform integer ambiguity resolution
with a standard ILS solver such as the well-known LAMBDA
Here, rub , rpb , and rsb are the body-frame positions of the method [37]. This is essentially an extension of the LC-
IMU, primary GNSS antenna, and secondary GNSS antenna, LAMBDA method described by Teunissen et al. in [38] to
respectively. Under this formulation, the known length of bw 2k the multi-baseline recursive estimation case. As demonstrated
serves as an implicit constraint on integer ambiguity resolution in [38], this method performs poorly for ultra-short (length
due to the parameterization of bw 2k as a function solely of À 1m) baselines when the a priori attitude estimate and the
attitude. pseudorange measurements cannot together offer a sufficiently
Importantly, the off-diagonal blocks of the measurement accurate estimate about which to linearize.
noise covariance matrix Σgk are nonzero because baselines 1 1) Alternatives: Recursive Bayesian estimation of xk re-
and 2 share a GNSS antenna. By consuming the GNSS mea- quires finding the distribution of bpxk q given a Gaussian prior
surements for all baselines in a single update, this correlation for xk with mean x̄k and covariance P̄k . While the true
is exploited and typically yields a higher integer fix success distribution of bpxk q is nontrivial, it is desirable to approximate
rate than either baseline taken individually [21]. Additionally, it as Gaussian to enable ambiguity resolution with standard
the vehicle attitude Rkwb is often known a priori to sub- ILS techniques, which are computationally efficient and well
degree precision, providing a tight constraint on all 3 degrees understood. Dropping the k subscripts for notational clarity,
of freedom of bw 2k , which both strengthens the combined such an approximation produces a joint Gaussian distribution
integer model [27] and increases positioning accuracy, since over x and bpxq:
the off-diagonal blocks of Σgk encode the sensitivity of GNSS „  ˆ„  „ ˙
measurements on the secondary vehicle antenna to rkw . x x̄ P̄ Pxb
„N , T
bpxq b̄ Pxb Pbb

B. Linearization This can be parameterized in terms of mean b̄, Jacobian Hb ,

and additional baseline uncertainty Σ˚b that accounts for errors
The function hgk pb, nk q is accurately modeled as linear
due to linearization:
due to the extreme distance to GNSS satellites relative to the
∆
CDGNSS baseline lengths: Pbb “ Hb P̄ HbT ` Σ˚b
» fi » fi ∆
(3)
ρ1k G1k bw
1 Pxb “ P̄ HbT
—φ1k ffi —G1k bw 1 ` Λ1 n1k ffi
ffi
hgk pbpxk q, nk q “ —
– ρ2k fl “ –G2k bw
ffi —
fl The first-order Taylor expansion scheme used by extant multi-
2
φ2k w
G2k b2 ` Λ2 n2k baseline CDGNSS Kalman filters to obtain these parameters
can be described as in Algorithm 1.
Here, nmk P ZNmk is the vector“ of carrier-phase integer
T T Algorithm 1: linearizeEkf
‰
ambiguities for baseline m, nk “ nT
1k , n2k , and Λ m is a
diagonal matrix composed of the wavelengths in meters of Input : x̄, P̄

5
 ˚
„ w: b̄, H
Output  b , Σb 1.00
b px̄q EKF 3σ
1 b̄ “ 1w 0.75
b2 px̄q UKF 3σ
» w fi
Bb1 pxq True distribution
0.50
Bx fl
2 Hb “ –
Bbw
2 pxq 0.25
Bx x̄

Y (m)
3 Σ˚b “ 0 0.00

−0.25
An alternative approach is to approximate the distribution
of bpxq using a deterministic sampling technique such as the −0.50
unscented transform (UT), which is used in the UKF [39]. −0.75
The UT infers the probability distribution of a transformed
Gaussian by transforming a set of weighted “sigma points” −1.00
−0.5 0.0 0.5 1.0 1.5
through the nonlinearity and evaluating the statistics of the
X (m)
transformed points. An implementation of the UT is shown in
Algorithm 2.
Fig. 3: Simplified 2-dimensional example of Gaussian distri-
Algorithm 2: linearizeUkf butions produced by linearizeEkf and linearizeUkf
for a single constrained baseline of length 1 m with a Gaussian
Input : x̄, P̄ prior over the attitude angle. The lines represent the cost
Output : b̄, Hb , Σ˚b contours in the position domain associated with each Gaussian
4 α “ 0.001, κ “ 0 distribution at the 3σ likelihood; the ˆ symbols represent
5 λ “ α2 pNx ` κq ´ Nx the means. The EKF distribution is infinitely thin, as EKF
“ ‰ ` ˘T linearizations only consider baseline vectors on a plane tangent
6 S “ s1 , s2 , . . . , sNx “ chol P̄
to the true spherical distribution at the a priori estimate, causing
7 x̄p0q “ x̄ ILS failures when the a priori uncertainty is large enough that
„ w ` p0q ˘
b x̄ the sphere significantly diverges from the tangent plane.
8 b̄p0q “ 1w ` p0q ˘
b2 x̄
p0q λ
9 wm “ Nx `λ
p0q λ 2) UKF measurement update: A vector zgk of DD GNSS
10 wc “ ` 1 ´ α2 ` β
Nx `λ observables is ingested by the estimator at epoch k. Linearizing
11 for i P r1, 2N
# x s do ? about the a priori state estimate x̄k
x̄p0q ‘ Nx ` λ si i P r1, Nx s
12 x̄piq “ ? “ ‰ ` ˘
x̄p0q ‘ ´ N ´ λ si i P rNx ` 1, 2Nx s b̄k , Hbk , Σ˚bk “ linearizeUkf x̄k , P̄k
„ w ` piq ˘ x
b x̄
13 b̄piq “ 1w ` piq ˘ yields the following approximation of the measurement model
b2 x̄ ∆ ` ˘
piq piq for innovations vector νgk “ zgk ´ hgk b̄k , 0 :
14 wm “ wc “ 2pNx1`λq
» fi » fi
15 end G1k 0 0 0 „ 
2N —G1k 0 ffi
ÿx ffi Hbk δxk ` —Λ1 0 ffi n1k
— ffi
b̄ “ piq piq
wm b̄ νgk “ —– 0
16 G2k fl –0 0 fl lon 2kon
omo
i“0 0 G 0 Λ
2k 2
looooomooooon n
˜ looooooooooomooooooooooon k
„  2N
ÿx „ piq „ T ¸
P̄xx P̄xb piq x̄ a x̄p0q x̄piq a x̄p0q Hrk Hnk
17 T “ wc
P̄xb P̄bb
i“0
b̄piq ´ b̄ b̄piq ´ b̄ ` gk ` ˚gk
` ´1 ˘T “ Hrk δxk ` Hnk nk ` gk ` ˚gk
18 Hb “ P̄xx P̄xb ` T
˘
19
`
Σ˚b “ P̄bb ´ Hb P̄xx HbT
˘ ˚gk „ N 0, Hbk Σ˚bk Hbk

∆
Here the state estimate error vector δxk “ x̄k axk is expressed
A simplified 2-dimensional example of the two linearization
in the tangent space of X at x̄k , and ˚gk represents additional
schemes for a single constrained baseline is shown in Fig.
measurement error caused by approximation of b pxk q.
3. The UT yields an approximate Gaussian distribution over
the baseline vector that more closely matches the true mean
and covariance of bpxq. For this reason, ILS-based ambiguity
C. Square-root formulation
resolution has a higher success rate when linearization is based
on linearizeUkf rather than linearizeEkf under large The CDGNSS measurement update can be cast in square-root
a priori attitude uncertainty, as will be demonstrated in Section form for greater numerical robustness and algorithmic clarity
III-D. [40]. Given νgk , Hrk , Hnk , x̄k , and P̄k , the measurement

6
update can be defined as finding δxk and nk to minimize the is accepted (having passed validation via an integer aperture
cost function test), the a posteriori state and covariance are
2 2
Jk pδxk , nk q “ kνgk ´ Hrk δxk ´ Hnk nk kΣ´1 ` kδxk kP̄ ´1 x̂k “ x̄k ‘ δ x̌k
k k
` T ˘´1 (6)
where Σk “ Σgk ` Hb Σ˚bk HbT .
The vector cost compo- P̂k “ Rxxk Rxxk
nents can be normalized by left multiplying with square- If instead the float solution is accepted, the a posteriori state
root information
` matrices
˘ based on Cholesky
` ˘ factorization and covariance are found by marginalizing over the distribution
Rgk “ chol Σk ´1 , R̄xxk “ chol P̄k´1 : of nk :
x̂k “ x̄k ‘ δ x̃k
Jk pδxk , nk q (7)
„  „  „  2 P̂k “ pRkT Rk q´1r1:Nx ,1:Nx s
0 R̄xxk 0
“ ´ δxk ´ n
Rgk νgk Rgk Hrk Rgk Hrk k Here, r1 : n, 1 : ns denotes taking the first n rows and columns
„ 1 „  2 of the matrix.
Hrk δxk
“ νk1 ´ 1
Hnk nk
The cost Jk can be decomposed via QR factorization D. Evaluation of Unscented Multi-Antenna Update
ˆ„ 1 ˙
Hrk
rQk , Rk s “ qr 1 Correct fix rate
Hnk
0.75
where matrix Qk is orthogonal and Rk is upper triangular. UKF

Ps
0.50
Because Qk is orthogonal, the components of Jk inside the EKF
norm can be left-multiplied by QT k without changing the cost,
0.25
and Jk can be decomposed into 3 terms: False fix rate
„  2
δxk EK
F
Jk pδxk , nk q “ QT ν
k k
1
´ R k 0.1
Pf

nk
 2
» 2 fi » fi
ν1k Rxxk Rxnk „ UKF
δx 0.0
k
2
“ –ν2k fl ´ – 0 Rnnk fl (4) Float rate
2 nk
ν3k 0 0
2 2 EKF
2 2 0.50
Pu

ν1k ´ Rxxk δxk ´ Rxnk nk ` loooooooooomoooooooooon

“ looooooooooooooooooomooooooooooooooooooon ν2k ´ Rnnk nk
J1k pδxk ,nk q J2k pnk q 0.25 UKF
2 2
ν3k
` loomoon 0 20 40 60 80
Yaw uncertainty (deg)
J3k
Fig. 4: Integer aperture (IA) success (validated correct fix),
If both the measurement model and R̄xxk are not ill- failure (validated false fix), and float (failed IA validation) rates
conditioned, then Rxxk and Rnnk are invertible. J3k is found via Monte Carlo simulation with 106 samples for an
the irreducible cost, and, under a single-epoch ambiguity example multi-baseline CDGNSS measurement update with
resolution scheme, can be shown to be equal to the normalized varying a priori attitude uncertainty . UKF and EKF denote the
innovations squared (NIS) associated with the double-difference ILS success rates for models derived from linearizeUkf
pseudorange measurements. J2k is the extra cost incurred by and linearizeEkf, respectively. The vehicle pitch and roll
enforcing the integer constraint on nk , and can similarly be angles were assumed to be known to 2˝ (1σ), and the a
shown to be equal to the NIS of the double-difference carrier priori yaw angle uncertainty was varied from 0˝ to 90˝ (1σ).
phase measurements (again assuming single-epoch ambiguity The simulation assumed GPS L1 C/A signals visible from
resolution). If nk is allowed to take any real value (the “float the equator with a representative satellite constellation and a
solution”), J2k can be zeroed due to the invertibility of Rnnk . 15˝ elevation mask angle. Two vehicle GNSS antennas were
Similarly, J1k can be zeroed for any value of nk due to the simulated, with a baseline length of 1.0668 m (equivalent
invertibility of Rxxk . The “float solution,” tδ x̃k , ñk u, can to that of the Sensorium). Other GNSS measurement model
therefore be formed by choosing δ x̃k and ñk to zero J1k and parameters were selected as in Table II. The threshold function
J2k . Because Rk is upper triangular, these values can be found approximation to the fixed-failure rate distance test [41] was
by efficient backsubstitution. The “fixed solution,” tδ x̌k , ňk u, used with fixed failure rate P̄ “ 0.01. Linearization error
f
is via an ILS solver, yielding causes the true failure rate Pf of the EKF linearization to
ňk “ arg minn J2k pnk q greatly exceed P̄f as the yaw uncertainty increases beyond
nk PZ
` 2 ˘ (5) approximately 8˝ . In contrast, the integrity of the integer
´1
δ x̌k “ Rxxk ν1k ´ Rxnk aperture test is maintained for the UKF case due to the
˚
Rxxk is the a posteriori state vector square-root information approximate linearization error term Σb provided by the UKF
matrix conditioned on n “ ň . Therefore, if the fixed solution linearization. Note that Pu “ 1 ´ Pf ´ Ps [42].
k k

7
 The UKF linearization of bpxk q yields a Gaussian prior in 2) State dynamics: The estimator’s dynamics function
the position domain that better captures the true mean and xk`1 “ f pxk , uk , vk q is defined by discretizing the state
covariance of the constrained attitude baseline than the EKF dynamics, assuming a zero-order hold for the IMU measure-
linearization. Consequently, the UKF linearization achieves ment vector uk . The accelerometer and gyroscope biases are
a higher ILS success rate than the EKF linearization when modeled as Ornstein-Uhlenbeck random processes with time
attitude uncertainty is large. Fig. 4 demonstrates this effect constants τa and τg and steady-state uncertainties σba and
by evaluating integer aperture success, failure, and float rates σbg , respectively, as derived from IMU datasheet values (with
for a multi-baseline CDGNSS measurement update via Monte additional hand tuning, as datasheet values are often optimistic).
Carlo simulation. For a dual-antenna platform similar to that of The continuous-time state dynamics are
the University of Texas Sensorium with loose yaw knowledge,
r9 w “ v w
this effect becomes apparent as 1σ yaw uncertainty exceeds
approximately 8˝ , whereupon the ILS success rate with the EKF v9 w “ aw “ Rwb Rbu pfku ´ bu u
a ´ va q
“ ‰ “ ‰
linearization begins a rapid decline with increasing uncertainty, R9 wb “ Rwb ω b ˆ “ Rwb Rbu ω u ˆ
eventually falling below even that of the unconstrained multi-
b9 u 1 u
a “ ´ τa ba ` va2
u
antenna CDGNSS snapshot estimator.
The UKF linearization expands the operating regime of the b9 u 1 u
g “ ´ τg bg ` vg2
u

CDGNSS navigation estimator to greater levels of attitude

where vau2 and vgu2 are Gaussian white noise processes driving
uncertainty than with the EKF linearization. This effect is “ ‰
the evolution of the accelerometer and gyroscope biases, ¨ˆ
relevant for low-cost urban CDGNSS: while even a low-cost
denotes“ the‰ skew-symmetric “cross-product equivalent” matrix
accelerometer can provide pitch and roll angles with degree-
(i.e., aˆ b “ a ˆ b), and where ω u is obtained by solving
level accuracy, it is desirable to tolerate large yaw uncertainty,
for it in (8), which makes it a function of the measurement ω̃ku .
as may occur, for example, following a long GNSS outage
The full discrete-time process noise vector vk P RNv , with
in a parking garage. Upon emerging from such an outage, a
Nv “ 12, is
system using an EKF linearization may require re-initialization
“ u T u T u T u T ‰T
using a snapshot attitude estimator. This scheme also allows vk “ vak , vgk , va2 k , vg2 k „ N p0, Qk q
initialization of the estimator with loose attitude knowledge,
as may be provided, for example, by a magnetometer whose where
” ı
heading measurement may be uncertain in the presence of Qk “ diag σv2a I3ˆ3 , σv2g I3ˆ3 , σv2a2 I3ˆ3 , σv2g2 I3ˆ3
nearby buildings and vehicles.
The process noise parameters are determined by
IV. T IGHTLY-C OUPLED NAVIGATION E STIMATOR
Sa Sg
This section presents the remaining development of the full σva “ , σvg “
tightly-coupled multi-antenna CDGNSS recursive estimator ∆t ∆t
ˆ ˆ ˙˙ ˆ ˆ ˙˙
with vehicle dynamics constraints and false integer fix mitiga- ´∆t ´∆t
σva2 “ σb2a 1 ´ exp , σvg2 “ σb2g 1 ´ exp
tion. τa τg
where Sa and Sg are the accelerometer and gyroscope white
A. Propagation step
noise density, respectively, and ∆t is the inertial sensor sample
State propagation is based on a model replacement approach period.
in which IMU measurements supplant a vehicle dynamics and 3) Unscented Kalman filter propagation: When a new
kinematics model, permitting broader application. inertial measurement uk arrives at epoch k, the estimator
1) Inertial measurement model: The“ estimator‰ ingests a state is propagated from epoch k to k ` 1 using the standard
T
vector of inertial measurements uk “ fku T , ω̃ku T at each on-manifold UKF propagation step. Let n “ Nx ` Nv ; then
epoch k, triggering a propagation step (time update). The ˆ„ ˙T
accelerometer’s specific force measurement fku and the gyro- P̂k 0 “ ‰
Sk “ chol “ s1 , s2 , . . . , sn
scope’s angular velocity measurement ω̃ku are modeled as 0 Qk
fku “ Rub Rkbw paw w u u
k ´ g q ` bak ` vak A set of sigma points χp0q ¨ ¨ ¨ χpnq is formed by
(8)
ω̃ku
“ ωku ` Rkbw ωC w
` bu u
gk ` vgk “ T T
‰ “ T T
‰
χp0q “ x̂Tk , v̂k “ x̂T k, 0
where aw u
k and ωk are the true linear acceleration and angular # ?
w
rate, respectively, of the inertial sensor, g is the local gravity piq χp0q ‘ n ` λ si i P t1, nu
χ “ p0q
?
vector after compensation for centripetal force due to earth χ ‘ ´ n ´ λ si i P tn ` 1, 2nu
w
rotation, ωC is the earth’s angular rate, bu u
ak and bgk are time-
The sigma points are transformed through the dynamics
varying zero-mean accelerometer and gyroscope measurement
u u function xk`1 “ f pxk , uk , vk q as
biases, and vak and vgk are accelerometer and gyroscope
measurement noise, which are modeled as zero-mean white piq
x̄k`1 “ f pχpiq piq
i “ 0, . . . , 2n
x , uk , χv q,
Gaussian noise. Static biases and scale factor errors, although
piq piq
not shown in (8), are also modeled and calibrated, as detailed where χx and χv denote the selection of the state and
in [30]. process noise components, respectively, of χpiq . Finally, the

8
sigma points are recombined to produce the a priori state 2) Zero velocity update (ZUPT): Zero-velocity updates
estimate x̄k`1 and covariance P̄k`1 : also offer a valuable constraint on inertial sensor drift. Im-
« ff portantly for urban CDGNSS, they act as an anchor in the
n
ÿ
piq
´
piq
¯ presence of the worsened multipath errors that occur when the
x̄k`1 “ x̂k ‘ wm x̄k`1 a x̂k
vehicle is stopped. (Wavelength-scale vehicle motion decor-
i“0
n
ÿ ´ ¯´ ¯T relates multipath-induced measurement errors [43].) ZUPT
piq piq pseudo-measurements are modeled similarly to NHC pseudo-
P̄k`1 “ wcpiq x̄k`1 a x̄k`1 x̄k`1 a x̄k`1
i“0 measurements but act on the full 3-dimensional v-frame velocity
vector:
piq piq
The sigma point spread parameter λ and weights wm , wc
are formed as in Algorithm 2. zzupt,k “ hzupt pxq ` zupt P R3
“ v v ` zupt
“ ` ` ˘˘‰
“ Rvb Rbw v w ` ω b ˆ rvb ´ rub ` zupt
B. Vehicle dynamics constraints ` “ 2 2 2 ‰˘
zupt „ N 0, diag σzx , σzy , σzz
This paper adopts the VDC scheme of [2] with minor
modifications to the NHC sideslip model and ZUPT detection where rvb and rub are the b frame positions of the v and
mechanism. It is described in this subsection for completeness u frame origins, respectively, and ω b “ Rbu ω u . A ZUPT
and notational consistency with the rest of the paper. pseudo-measurement zzupt,k “ 0 is periodically applied when
the vehicle is detected to be stationary due to a lack of road
1) Non-holonomic constraint (NHC): The estimator exploits
vibration measured by the IMU. The ZUPT measurement
the natural constraints on the motion of four-wheeled ground
constraint is taken to be tighter in the y and z (lateral and
vehicles, known as non-holonomic constraints (NHCs), by
vertical) directions, as the vehicle may appear to be stationary
casting them as pseudo-measurements. The NHC model
via IMU vibrations when it is in reality traveling very slowly
assumes that the vehicle rotates about a fixed center of rotation
forwards.
(the origin of the v frame) when a steering input is applied,
Vehicle stationarity is detected via simple thresholding of
only moves in the v frame x-direction when no steering input is
time differences ∆fku and ∆ω̃ku of raw accelerometer and
applied, does not leave the surface of the road, and experiences
gyroscope samples, which are defined as
only a small, predictable amount of sideslip. The vehicle
∆ ∆
sideslip (y-component of the velocity of the v frame relative to ∆fku “ fku ´ fk´1
u
, ∆ω̃ku “ ω̃ku ´ ω̃k´1
u

the w frame, expressed in the v frame) is modeled as dependent

on the steering rate according to the second-order polynomial and are evaluated using threshold parameters γzupt,a and
model γzupt,g . If the conditions k∆fku k ă γzupt,a and k∆ω̃ku k ă
´ ¯2 γzupt,g have both been satisfied for all of the most recent Nzupt
v
v̂kpyq b
“ P0 ωkpzq b
` P1 ωkpzq inertial measurements, then a ZUPT pseudo-measurement is
applied at every GNSS measurement epoch. Finally, a simple
where r¨spyq and r¨spzq denote taking the y and z components, χ2 innovations test is applied with threshold parameter Pf,zupt
respectively, of a vector. The vehicle-specific polynomial to mitigate the effect of falsely-detected ZUPTs.
coefficients P0 and P1 are found via offline calibration. The
estimator adopts the following measurement model for the C. Outlier Rejection using Pseudorange Innovations
NHC pseudo-measurements: T
Let νρk “ rνρk1 , . . . , νρkn s contain the elements of the
GNSS innovation vector νgk that correspond to pseudorange
znhc,k “ hnhc pxk q ` nhc P R2 measurements. Its covariance matrix νρk is
v
“ vpy,zq ` nhc „  „ T
“ vb ` bw w ` ˘˘‰ G1k 0 G1k 0
“ R R v ` ω b ˆ rvb ´ rub py,zq ` nhc P̄ρρk “
0 G2k
Pbbk
0 G2k
` Σρk (9)
ˆ „ 2 ˙
σ 0 where Pbbk is formed as in (3), and Σρk is formed by
nhc „ N 0, nhc,y 2
0 σnhc,z selecting the rows and columns of Σgk that correspond to
pseudorange measurements. If the estimator is consistent and
The sideslip error standard deviation σnhc,y is typically taken
the pseudorange measurements `are` not corrupted by large
to be looser than the vertical motion error standard deviation ˘ ˘
multipath errors, then νρkn „ N 0, P̄ρρk nn . A test can be
σnhc,z . The pseudo-measurement
used to detect outlier pseudorange measurements using the
„ v 
v̂ detection statistic
znhc,k “ kpyq
0 2
` ˘
qkn “ νρk,n { P̄ρρk nn
is periodically applied as an unscented measurement update to with a threshold γ 2 selected to yield a sufficiently low false-
the navigation estimator. The v frame origin rvb and orientation positive rate. If a pseudorange outlier is detected for some index
Rvb are calibrated offline in batch mode using the method n (i.e., qkn ą γ 2 ), then the corresponding DD pseudorange
developed in [2]. measurement is assumed to be corrupted by multipath. Because

9
the effect of multipath on GNSS signals is primarily a function state on fixed and validated integers at each epoch, as is done
of the line-of-sight vector to the transmitter, all pseudorange under a single-epoch ambiguity resolution scheme, is perilous
and carrier phase measurements associated with the offending because false integer fixes eventually corrupt the filter state
non-pivot satellite on all frequencies and baselines are assumed with incorrect but highly confident priors. This causes future
to be corrupt. They are removed from the measurement vector GNSS measurement updates to accept a similarly incorrect
zgk before proceeding with the measurement update. fix with high probability, repeatedly conditioning the filter
state on incorrect ambiguities. While the simple rectilinear
motion model of the unaided estimator in [4] contains sufficient
D. False fix detection and recovery
process noise that the filter eventually re-fixes to the correct
1) Single-epoch ambiguity resolution: An optimal CDGNSS ambiguities, the tight epoch-to-epoch constraints of the present
filter (in the maximum a posteriori sense) must append a new paper’s tightly-coupled estimator can cause these cycles of
carrier phase integer ambiguity to its state each time a cycle slip false fixes to persist indefinitely.
is detected in a carrier tracking loop [44]. This causes the update To mitigate the effect of conditioning on incorrect integer
and ILS solve operations to quickly become computationally ambiguities, this paper’s estimator employs a fault detection and
intractable when applied in an urban environment where cycle exclusion technique based on solution separation. A float-only
slips are common and detection of discrete cycle slips is often filter, configured to never attempt fixing integer ambiguities,
impossible. is operated in parallel to the primary navigation filter. Under
A scheme could be imagined whereby cycle slips are the single-epoch ambiguity resolution scheme, the float-only
modeled to occur with some probability PCS at each epoch, filter’s behavior is equivalent to accepting only pseudorange
and a suboptimal dynamic multiple-model estimator such as measurements, discarding carrier phase measurements entirely.
the interacting multiple-model or generalized pseudo-Bayesian 3) Carrier phase innovations testing: The primary filter’s
estimator [45, Sec. 11.6] is used to handle the resulting multiple carrier phase measurement innovations sequence is monitored
hypotheses over past cycle slips. However, for large PCS , to detect filter inconsistency, which is assumed to be caused
as holds for urban CDGNSS, these methods would again by false integer fixes. The carrier phase measurement NIS is
require measurement updates and ILS solve operations for defined as
a quickly-growing number of integer ambiguity candidates, ∆ T ´1
making such a method computationally prohibitive without φk “ ν̌φk P̄φφk ν̌φk
aggressive hypothesis pruning.
This paper’s estimator adopts a posture of maximum pes- where ν̌φk is the vector of integer-resolved DD carrier phase
simism regarding cycle slips: each carrier tracking loop in measurement innovations at epoch k, and P̄φφk is the innova-
the GNSS receiver is assumed to slip cycles between each tion covariance matrix, formed similarly to P̄ρρk of (9). The
pair of measurement epochs (i.e., PCS “ 1). Ambiguity state NIS φk can be calculated during the square-root measurement
growth is curtailed by discarding all ambiguity states at every update by φk “ J2k pňk q from (4).
GNSS measurement epoch, either by conditioning the state The test statistic used to detect false fixes is the windowed
vector on the candidate fixed solution (if the candidate fix carrier phase NIS Ψk over a moving window of fixed length
is validated), as in (5) and (6), or by accepting the float l of past GNSS measurement epochs. It has NΨk degrees of
solution and marginalizing over the ambiguities, as in (7). freedom and is calculated by
This single-epoch ambiguity resolution scheme is suboptimal k k
because it discards what continuity may be present in the ∆
ÿ ∆
ÿ
Ψk “ φn , NΨk “ Nn
integer ambiguities from epoch to epoch, weakening the integer n“k´l`1 n“k´l`1
model strength. But it renders the navigation filter entirely
insensitive to cycle slips, which may be a practical necessity where Nk is the number of DD carrier phase measurements at
for urban CDGNSS, and is computationally efficient. Moreover, epoch k. If the filter is consistent and the integer ambiguities
the conditioning operation, when applied, greatly increases the are correctly resolved, then the innovations sequence should be
integer model strength for subsequent epochs, allowing the approximately white and Gaussian, and Ψk should be approxi-
filter to “hold on” to fixes by virtue of the tight epoch-to-epoch mately χ2 -distributed with with NΨk degrees of freedom. (This
position domain constraint provided by the inertial sensor and distribution is approximate due to the “tail clipping” effect of
the vehicle dynamics pseudo-measurements. integer fixing: large phase residuals are not possible because
2) False integer fixes: Even in the absence of modeling of integer-cycle phase wrapping.) A statistical consistency test
errors and measurement outliers, the integer fixing procedure can be performed by choosing a desired false-alarm rate P̄f,Ψ
of Section III-C occasionally yields an incorrect integer vector and declaring a false fix if Ψk ą γΨk , where the threshold γΨk
ňk that passes ambiguity validation tests, a situation referred to is calculated by evaluating the inverse cumulative distribution
as a false integer fix or an ambiguity resolution failure [42]. This function (CDF) of χ2 pNΨk q at P̄f,Ψ .
occurs because validation tests can only provide probabilistic Use of a window of NIS values over multiple epochs
guarantees of integer correctness [46]. In an urban environment, increases the statistical power of the consistency test and helps
measurement outliers due to multipath and diffraction cause avoid premature declaration of a false fix due to sporadic
the true false integer fix rate Pf to greatly exceed the specified measurement outliers. However, increasing the window length
rate P̄f for the validation test [4]. Thus, conditioning the filter l also increases the latency to detect a false fix event.

10
 4) False fix recovery: If a false fix is detected (Ψk ą γΨk ), equivariant estimation to the class of elliptically contoured
the estimator performs a soft reset, discarding the primary distributions in [48] may offer a means of providing a better
navigation filter’s state estimate and covariance and replacing re-seed estimate for the float-only filter in urban environments,
them with a copy of the float-only filter’s state and covariance, but it has not been tested with empirical urban data. Meanwhile,
as shown in Fig. 5. application of this paper’s re-seeding technique with the criteria
in Table I will be shown in the next section to significantly
!
increase integer fix availability while respecting a low false fix
Primary x̂1 x̂2 x̂3 x̂4 x̂5
rate.
et
res
ft
so
Float-Only x̂1 x̂2 x̂3 x̂4 x̂5
V. P ERFORMANCE E VALUATION IN A D EEP U RBAN
Fig. 5: A false-fix detection and recovery event. “!” denotes E NVIRONMENT
a window NIS test failure at k “ 4. The primary navigation A. Experimental setup
filter’s state estimate and covariance matrix are assumed to
The tightly-coupled CDGNSS estimator described in the
be contaminated by false integer fixes at past epochs and are
foregoing sections was implemented in C++ as a new version of
discarded. They are replaced after epoch 4 with the less-certain
the PpEngine sensor fusion engine [4], and was experimentally
but uncontaminated state and covariance of the float-only filter.
evaluated against the publicly-available TEX-CUP urban posi-
tioning dataset. TEX-CUP comprises raw GNSS intermediate-
X frequency (IF) samples and inertial data collected on 9 and 12
Primary x̂1 x̂2 x̂3 x̂4 x̂5 May, 2019, using the University of Texas Sensorium vehicular
re-
see
d
perception research platform [32]. The dataset consists of a
Float-Only x̂1 x̂2 x̂3 x̂4 x̂5
total of over 2 hours of driving in Austin, Texas in conditions
ranging from light to dense urban; routes are shown in Fig. 7.
Fig. 6: A “re-seed” event.“X” denotes re-seed criteria being met
at epoch 4, indicating very high confidence in the correctness
of the primary filter’s integer fix and consistency of its
state estimate. The estimator replaces the state estimate and
covariance of the float-only filter with a copy of that of the
primary navigation filter.

5) Float-only estimator re-seeding: To increase the proba-

bility of a correct fix after a soft reset, the “float-only” filter
is occasionally re-seeded with the state of the primary filter
during epochs over which a set of heuristic criteria indicate
that a correct fix is extremely likely. This operation carries the
risk that the float-only filter could also be contaminated with
information from a false integer fix in the primary filter. But
in practice, heuristic criteria can be set to strictly limit this
event’s probability. The four criteria used in the next section’s
evaluation are given in Table I.

Carrier phase measurement NIS φk{Nk ď 1.0

Windowed carrier phase NIS Ψk{NΨk ď 0.5
Last fix number of DD measurements Nk ě 10
Time since last soft reset tsr,k ě 2.0 s

TABLE I: Re-seed criteria used in the evaluation of Section

V. A re-seed operation is performed if all four conditions are
met. Fig. 7: Overview of the CDGNSS reference station position
and routes driven through the urban core of Austin, Texas in
One might argue that these criteria for re-seeding the float- the TEX-CUP urban positioning datasets. Routes differ slightly
only filter are redundant because an integer aperture test for from May 9 to May 12 due to road closures.
validating the primary filter’s integer estimate can be made
arbitrarily strict, obviating additional validation. But integer Two-bit-quantized IF samples were captured at the Senso-
aperture theory is founded on modeling measurement error rium and at the reference station through RadioLynx, a low-
distributions as Gaussian [47], which is a poor approximation in cost L1+L2 GNSS front end with a 5 MHz bandwidth at
the urban environment, leading to low fixed solution availability each frequency, and were processed with the PpRx software-
[4]. Teunissen’s recent extension of so-called best integer defined GNSS receiver [4]. The Sensorium RadioLynx was

11
connected to two Antcom G8 GNSS antennas separated by CDGNSS parameters
Carrier-to-noise ratio threshold C{N0 ě 40 dB-Hz
1.0668 meters in the vehicle Y direction, and the reference Phase lock statistic threshold sθ ě 0.8
RadioLynx was connected to a Trimble Zephyr II geodetic- Elevation mask θel ě 10˝
grade GNSS antenna. Integer aperture (IA) validation test FF-difference test [41]
IA fixed failure rate P̄f “ 0.001
The system’s performance was separately evaluated using Undifferenced zenith pseudorange std σρ “ 1.5 m
inertial data from each of the Sensorium’s two MEMS inertial Undifferenced zenith phase std σφ “ 0.006 m
sensors. The first, a LORD MicroStrain 3DM-GX5-25, is an Pseudorange outlier threshold std γ “ 1.5σ
False fix detection window length l “ 10
industrial-grade sensor. The second, a Bosch BMX055, is a False fix detection threshold P̄f,Ψ “ 10´15
surface-mount consumer-grade sensor. Their relevant datasheet
IMU parameters
specifications are compared in reference [49]. ? ? ˚
Accelerometer noise density Sa “ 100, 300 µg{ Hz
The system’s positioning performance was evaluated by Accelerometer bias steady-state std σba “ 0.5, 10 mg ˚
comparing against TEX-CUP’s forward-backward smoothed Accelerometer bias time constant τ
a a
“ 100 s
? ˚
ground-truth reference trajectory. This ground truth was gen- Gyroscope noise density Sg “ 0.01, 0.05 p˝ {sq{ Hz
˝ ˚
Gyroscope bias steady-state std σbg “ 8, 30 {hr
erated by post-processing the data from an iXblue ATLANS-
Gyroscope bias time constant τg “ 100 s
C mobile mapping system comprising a Septentrio AsteRx4
NHC parameters
RTK receiver and a high-end tactical-grade IMU. The reported Lateral std σnhc,y “ 0.1 m/s
accuracy of the ground truth trajectory varied between 2 and 15 Vertical std σnhc,z “ 0.2 m/s
cm (1-σ) along the route. Because it is impossible to directly ZUPT parameters
evaluate integer ambiguity resolution performance, the integer Longitudinal std σzx “ 0.05 m/s
fixing performance in the following sections was evaluated by Lateral/vertical std σzy , σzz “ 0.01 m/s
Accelerometer noise threshold γzupt,a “ 0.8 m/s2
considering integer fixes to be correct if the 3D distance to Gyroscope noise threshold γzupt,g “ 0.006, 0.018 rad/s*
the ground truth was below 30 cm, following [4]. ZUPT detection window Nzupt “ 10, 30*
Because TEX-CUP contains several minutes of no motion Innovations test threshold P̄f,zupt “ 10´30 , 10´6 *
in an open-sky environment at the beginning and end of each
TABLE II: Baseline PpEngine configuration parameters. Pairs
capture, the estimator was run on a subset of each capture
marked with “*” indicate separate parameter values used with
beginning approximately 10 seconds before first motion and
the industrial-grade and consumer-grade IMU, respectively.
ending 10 seconds after last motion. On the May 9 dataset,
IMU noise parameters were increased from datasheet values
the estimator was run and evaluated from GPS time of week
for consistency with empirical observations.
(TOW) 411003 s to 415029 s, and on the May 12 dataset, from
TOW 63770 s to 67972 s.
Consumer-grade IMU
1.00
B. Baseline configuration
CDF(error)

The performance of the tightly-coupled estimator with all 0.95

proposed features and signals enabled (the “baseline configura-
tion”) was evaluated using each IMU on the May 9 and May 12
0.90
TEX-CUP datasets. PpRx was configured as described in [4],
tracking the GPS L1 C/A, GPS L2C (combined CL+CM codes), Industrial-grade IMU
Galileo E1 (combined B+C codes), and L1 SBAS (WAAS) 1.00
CDF(error)

signals on the reference and both rover antennas. Data bit

prediction and wipeoff were performed on the GPS L1 C/A 0.95
and SBAS signals. Reference and rover GNSS observables were Fixed epochs
All epochs
produced at a rate of 5 Hz. PpEngine’s baseline configuration 0.90
parameters are given in Table II.
0.0 0.2 0.4 0.6 0.8 1.0
The phase center variation of the Sensorium antennas with 3D Positioning Error (m)
respect to signal elevation angle was calibrated as in [4]. The
orientation, body-frame position, axis scale factors, and steady- Fig. 8: Cumulative distribution functions of 3D positioning
state biases of both inertial sensors were calibrated offline error of baseline estimator configuration across both days of
using a short period of dynamic open-sky GNSS data at the the TEX-CUP urban positioning dataset.
beginning of the May 12 dataset.
The position and attitude states of the tightly-coupled
navigation estimator were initialized with the first available the primary and secondary vehicle antennas, as determined
batch of GNSS observables and inertial measurements. The with a brute-force attitude-domain search.
position state was initialized with the standard (single-ended)
pseudorange position solution for the primary antenna. The C. Baseline performance
attitude state was initialized by combining the gravity vector 1) Ambiguity resolution and positioning: The achieved
as determined by the IMU’s accelerometers with a constrained- integer-fix availability was 97.52% and 96.62% when tightly
baseline snapshot CDGNSS solution for the baseline connecting coupled with the industrial-grade and consumer-grade IMUs,

12
 Overall positioning performance (fix & float epochs)
Ambiguity Resolution 3D Horizontal Vertical
Dataset PV (%) Pf (%) d95 (cm) RMSE (cm) d95h (cm) RMSE (cm) d95v (cm) RMSE (cm)

May 9 Unaided 77.31% 0.33% 742.1 1799.6 471.4 1713.3 209.3 550.5
Consumer-grade IMU 94.56% 0.19% 20.2 34.4 16.4 28.6 7.3 19.2
Industrial-grade IMU 97.28% 0.21% 13.8 20.6 10.6 18.9 4.0 8.1

May 12 Unaided 76.94% 0.49% 860.5 571.9 525.5 406.1 467.9 402.7
Consumer-grade IMU 98.63% 0.56% 13.7 7.7 10.6 6.0 5.7 4.8
Industrial-grade IMU 97.74% 0.43% 12.3 13.8 9.7 12.5 4.9 5.8

TABLE III: Baseline estimator ambiguity resolution and positioning performance on each day of the TEX-CUP dataset. “Unaided”
indicates the use of the motion model of [4] in lieu of inertial tight coupling, as described in Sec. V-D. Quoted 95th percentile
and RMS error quantities are over the entire dataset (i.e., for both float and fixed epochs). PV denotes the availability of an
aperture-test-validated fixed solution for conditioning in the primary filter. Pf denotes the false fix rate, as determined by an
excursion of the primary filter beyond 30 cm from the ground truth when conditioned on fixed ambiguities.

Unaided
5
0.2 E
Error (m)

N
0 0.0
U
−0.2
−5
Consumer-grade IMU
5
0.2
Error (m)

0 0.0

−0.2
−5
Industrial-grade IMU
5
0.2
Error (m)

0 0.0

−0.2
−5
0 1000 2000 3000 4000 0 1000 2000 3000 4000
Elapsed time (s) Elapsed time (s)

Fig. 9: Baseline estimator positioning error over time in the East, North, and Up directions for the May 9 TEX-CUP dataset.
This day had worse GNSS satellite geometry and therefore lower positioning performance than on May 12. Gray shading
indicates float epochs (periods when the estimator accepted the float CDGNSS solution). “Elapsed time” indicates time since
dataset start at GPS TOW 411003 s.

respectively, across both days of the dataset. When tightly 2) Attitude: The attitude performance of the baseline estima-
coupled with the industrial-grade IMU, the 95th-percentile tor is excellent when tightly coupled with either the industrial-
horizontal positioning error was 8.4 cm when fixed and 10.1 grade or consumer-grade inertial sensor, achieving single-
cm overall (fixed and float). Using the consumer-grade IMU, degree-level precision in all three axes with the consumer-grade
the 95th-percentile horizontal error was 9.2 cm when fixed sensor, and sub-degree precision with the inertial sensor. Better
and 12.0 cm overall. The empirical CDF of 3D positioning performance with the industrial-grade sensor is as expected
errors using both grades of inertial sensor is shown in Fig. 8. due to its significantly better gyroscope noise properties.
Detailed statistics of the estimator’s positioning performance in
the baseline configuration are given in Table III, and statistics D. Effect of inertial tight coupling
of its attitude performance in Table IV. Figs. 9 and 10 show
To evaluate the benefit of inertial tight coupling, PpEngine
the position and attitude error, respectively, over time for the
was run in an “unaided” mode, using the nearly-constant-
May 9 portion of TEX-CUP.
velocity motion model described in [4] for propagation in
place of an inertial sensor. Attitude dynamics were modeled as

13
 Consumer-grade IMU antenna. CDGNSS observables were formed for only a single
2 baseline, between the reference antenna and the primary Senso-
rium GNSS antenna. Despite having fewer integer ambiguities
Error (deg)

1
(only one baseline’s worth) to fix at each measurement epoch,
0
the integer fix availability was lower than for the multi-antenna
−1 case in this configuration because the estimator was no longer
−2 able to exploit the measurement noise cross-covariance between
Industrial-grade IMU baselines due to the shared antenna. Positioning performance
2
is slightly worse in the single-antenna case by all metrics,
Error (deg)

1 as would be expected due to the loss of half of all GNSS

Pitch measurements.
0
Roll 2) Performance without vehicle motion constraints: The
−1
Yaw non-holonomic constraints (Sec. IV-B1) were disabled in
−2 Configuration 3, and zero-velocity updates (Sec. IV-B2) were
0 1000 2000 3000 4000
disabled in Configuration 4. It can be seen that the non-
Elapsed time (s)
holonomic constraints have only a small effect on performance
Fig. 10: Baseline estimator attitude estimation error over time when tightly coupling with the industrial-grade IMU. With
for the May 9 TEX-CUP dataset. Heading-dependent pitch and the consumer-grade IMU, however, the effect is much greater.
roll errors are evident when tightly coupled with the industrial- Clearly, non-holonomic constraints (Sec. IV-B1) provide a
grade IMU; these are likely due to residual calibration errors. major performance boost in the consumer-grade IMU case,
cutting the number of float GNSS measurement epochs in half
Roll (°) Pitch (°) Yaw (°)
(fix availability increased from 94.86% to 96.62%).
The incorporation of zero-velocity updates improves all of
Dataset p95 RMS p95 RMS p95 RMS
the presented statistics when tightly coupling with either the
May 9 industrial-grade or consumer-grade IMU. But for the consumer-
Consumer-grade IMU 0.99 0.59 0.99 0.60 0.99 0.26 grade IMU, the benefit of ZUPTs is minimal. This is likely
Industrial-grade IMU 0.28 0.21 0.28 0.18 0.28 0.22 because the poor noise properties of the consumer-grade IMU
May 12 required such strict thresholds to limit false detections that
Consumer-grade IMU 0.84 0.43 0.84 0.41 0.84 0.27 many opportunities to apply a ZUPT were missed. However,
Industrial-grade IMU 0.29 0.21 0.29 0.17 0.29 0.18 ZUPTs have a clearly positive benefit on overall RMS position
error, as they help to constrain against float-solution position
TABLE IV: Baseline estimator attitude performance on each error when the vehicle is stopped, which is when urban code
day of the TEX-CUP dataset. Quoted 95th percentile and RMS multipath errors are largest.
error quantities are over the entire dataset (i.e., for both float 3) Performance without pseudorange outlier exclusion: The
and fixed epochs). pseudorange innovations-based outlier exclusion mechanism
(Sec. IV-C) was disabled in Configuration 5. Without this
mechanism, the availability of validated integer-fixed solutions
a simple integrated white noise process, with noise intensity of
? ? decreased drastically due to the presence of outlier measure-
0.3 ˝{ s in the vehicle pitch and roll axes, and 5.7 ˝{ s in the
ments caused by multipath. Interestingly, the false fix rate Pf
vehicle yaw axis. Because vehicle pitch (rotation about the axis
was elevated, but not to extreme levels. This was likely because
connecting the primary and secondary Sensorium antennas) is
false fixes due to multipath-induced outliers were reverted by
not strongly observable, a weak pseudo-measurement of zero
the false-fix recovery mechanism.
pitch was added with standard deviation 10˝ at a rate of 5 Hz,
4) Performance without false-fix detection and recovery:
which was found to provide good estimation performance.
In Configuration 6, the “re-seed” mechanism described in Sec.
The positioning and ambiguity resolution performance of the
IV-D5 was disabled, and in Configuration 7 the entire false-fix
unaided estimator is shown in Table III. Tight coupling with
detection and recovery mechanism (Sec. IV-D) was disabled.
even a consumer-grade IMU has a clearly beneficial effect on
Integer fix performance without the re-seed mechanism is
ambiguity resolution, greatly increasing the fraction of fixed-
appreciably reduced, as the estimation performance of the
integer epochs for comparable Pf , and reducing 95th-percentile
float-only estimator suffers without the ability to re-seed from
and RMS positioning errors from meters to centimeters.
especially trustworthy integer fixes. Disabling the false-fix
detection and recovery mechanism has a catastrophic effect on
E. Performance in alternate configurations false fixing rate Pf , for the reasons given in Sec. IV-D.
Next, the estimator was run in a collection of alternate 5) Performance on subsets of GNSS signals: In Configura-
configurations (with various features disabled) in order to tion 8, the estimator was run using only L1 GNSS signals (i.e.,
study its performance’s sensitivity to the presence of various GPS L2C was disabled), and in Configuration 9, SBAS L1
algorithmic components. Results are given in Table V. signals were disabled. Ambiguity resolution and positioning
1) Performance in single-antenna mode: In Configuration performance on these GNSS signal subsets was fairly close
2, the estimator was run using only a single vehicle-mounted to the baseline case when using the industrial-grade IMU, but

14
 Industrial-grade IMU Consumer-grade IMU
Configuration PV (%) Pf (%) d95 (cm) RMSE (cm) PV (%) Pf (%) d95 (cm) RMSE (cm)

1 Baseline (all features enabled) 97.52 0.33 13.0 17.5 96.62 0.38 16.2 24.8
2 Single vehicle antenna 96.89 0.55 16.2 25.3 95.73 0.53 19.3 30.6
3 Sans non-holonomic constraints (§IV-B1) 97.48 0.33 13.2 18.3 94.86 0.60 26.1 116.0
4 Sans zero-velocity updates (§IV-B2) 97.39 0.34 13.8 38.2 96.39 0.40 16.9 27.1
5 Sans pseudorange outlier exclusion (§IV-C) 93.09 0.48 19.1 45.5 87.73 1.25 241.8 133.6
6 Sans re-seed (§IV-D5) 97.00 0.91 15.0 36.8 96.58 0.46 16.5 34.2
7 Sans false-fix detection & recovery (§IV-D) 99.00 24.28 590.8 194.7 97.27 10.03 121.6 46.2
8 Single frequency (L1 only) 97.65 0.96 14.7 24.2 90.35 0.38 138.9 87.1
9 Sans SBAS 95.73 0.55 17.1 25.6 88.86 0.86 191.2 80.3
10 EKF CDGNSS Update (linearizeEkf) 97.52 0.33 13.0 17.5 96.62 0.38 16.2 24.8

TABLE V: Estimator ambiguity resolution and positioning performance on the combined TEX-CUP May 9 and May 12 datasets
with various estimator features disabled.

with the consumer-grade IMU a substantial loss of integer- GNSS antenna significantly increased integer-fix availability,
fix availability occurred (down from 96.62% to 90.35% and decreased false-fix rate, and improved both root-mean-square
88.86% in each of these configurations, respectively). The and 95th-percentile positioning performance as compared to a
weaker motion constraints provided by the consumer-grade single-baseline CDGNSS configuration.
IMU cause the estimator to require more GNSS signals for
acceptable integer fix availability. ACKNOWLEDGMENTS
6) Performance with EKF-based linearization: The es- This work was supported by the U.S. Department of
timator was run using linearizeEkf in place of Transportation under Grant 69A3552047138 for the CAR-
linearizeUkf in Configuration 10. Due to the excellent MEN University Transportation Center, and by the Army
attitude performance of the estimator with either inertial sensor, Research Office under Cooperative Agreement W911NF-19-2-
no significant difference in performance arises on the TEX-CUP 0333. The views and conclusions contained in this document
dataset by using the UKF update. This can be explained by are those of the authors and should not be interpreted as
referring to the results of Fig. 4, which show that the benefit representing the official policies, either expressed or implied,
of the UKF linearization is significant for the Sensorium’s of the Army Research Office or the U.S. Government. The
inter-antenna distance only when attitude uncertainty exceeds U.S. Government is authorized to reproduce and distribute
approximately 8˝ on a single axis, whereas the attitude error reprints for Government purposes notwithstanding any copy-
on the TEX-CUP dataset never exceeded 2˝ . right notation herein. Map data © OpenStreetMap contributors
(https://www.openstreetmap.org/copyright).
VI. C ONCLUDING R EMARKS
R EFERENCES
A vehicular pose estimation technique has been presented
and evaluated that tightly-couples multi-antenna CDGNSS, [1] T. G. Reid, S. E. Houts, R. Cammarata, G. Mills, S. Agarwal, A. Vora,
and G. Pandey, “Localization requirements for autonomous vehicles,”
a low-cost MEMS IMU, and vehicle dynamics constraints arXiv preprint arXiv:1906.01061, 2019.
(non-holonomic constraints and zero-velocity updates). The [2] L. Narula, P. A. Iannucci, and T. E. Humphreys, “Towards all-weather sub-
unscented transform was used to linearize the multi-antenna 50-cm radar-inertial positioning,” Field Robotics, 2021. To be published.
[3] P. Teunissen and O. Montenbruck, eds., Springer handbook of global
CDGNSS update, allowing the use of a linear integer least navigation satellite systems. Springer, 2017.
squares solver for ambiguity resolution while exploiting [4] T. E. Humphreys, M. J. Murrian, and L. Narula, “Deep-urban unaided
between-baseline correlations and respecting the constraints precise global navigation satellite system vehicle positioning,” IEEE
Intelligent Transportation Systems Magazine, vol. 12, no. 3, pp. 109–122,
provided by known vehicle antenna geometry, even under 2020.
large attitude uncertainties. Robust estimation techniques were [5] R. B. Ong, M. G. Petovello, and G. Lachapelle, “Assessment of
developed to mitigate the effects of urban multipath and GPS/GLONASS RTK under various operational conditions,” in Pro-
ceedings of the ION GNSS Meeting, pp. 3297–3308, 2009.
signal blockage, and to recover from false integer fixes. The [6] J. Jackson, B. Davis, and D. Gebre-Egziabher, “An assessment of low-
estimator was evaluated using the publicly-available TEX- cost RTK GNSS receivers,” in Proceedings of the IEEE/ION PLANSx
CUP urban positioning dataset, yielding a 96.6% and 97.5% Meeting, (Monterey, CA), 2018.
[7] T. Li, H. Zhang, Z. Gao, Q. Chen, and X. Niu, “High-accuracy positioning
integer fix availability, and 12.0 cm and 10.1 cm overall in urban environments using single-frequency multi-GNSS RTK/MEMS-
(fix and float) 95-th percentile horizontal positioning error IMU integration,” Remote Sensing, vol. 10, no. 2, p. 205, 2018.
with a consumer-grade and industrial-grade inertial sensor, [8] M. Petovello, M. Cannon, and G. Lachapelle, “Benefits of using a tactical-
grade IMU for high-accuracy positioning,” Navigation, Journal of the
respectively, over more than two hours of driving in the urban Institute of Navigation, vol. 51, no. 1, pp. 1–12, 2004.
core of Austin, Texas. A performance sensitivity analysis [9] B. M. Scherzinger, “Precise robust positioning with inertially aided RTK,”
showed that the false-fix detection and recovery scheme is Navigation, vol. 53, no. 2, pp. 73–83, 2006.
[10] H. T. Zhang, “Performance comparison on kinematic GPS integrated
key to achieving an acceptably low false integer fixing rate of with different tactical-grade IMUs,” Master’s thesis, The University of
0.3%and 0.4%, respectively. Having a second vehicle-mounted Calgary, Jan. 2006.

15
[11] S. Kennedy, J. Hamilton, and H. Martell, “Architecture and system [34] F. L. Markley, “Multiplicative vs. additive filtering for spacecraft attitude
performance of SPAN—NovAtel’s GPS/INS solution,” in Position, determination,” Dynamics and Control of Systems and Structures in
Location, And Navigation Symposium, 2006 IEEE/ION, p. 266, IEEE, Space, no. 467-474, p. 48, 2004.
2006. [35] J. Sola, J. Deray, and D. Atchuthan, “A micro Lie theory for state
[12] M. Murrian, C. Gonzalez, T. E. Humphreys, and T. D. Novlan, “A dense estimation in robotics,” arXiv preprint arXiv:1812.01537, 2018.
reference network for mass-market centimeter-accurate positioning,” in [36] M. Psiaki and S. Mohiuddin, “Modeling, analysis, and simulation of GPS
Proceedings of the IEEE/ION PLANS Meeting, (Savannah, GA), 2016. carrier phase for spacecraft relative navigation,” Journal of Guidance,
[13] D. Odijk, Fast Precise GPS Positioning in the Presence of Ionospheric Control, and Dynamics, vol. 30, no. 6, p. 1628, 2007.
Delays. No. no. 52 in Fast precise GPS positioning in the presence of [37] P. J. Teunissen, “The least-squares ambiguity decorrelation adjustment: a
ionospheric delays, NCG, Nederlandse Commissie voor Geodesie, 2002. method for fast GPS integer ambiguity estimation,” Journal of Geodesy,
[14] M. Abd Rabbou and A. El-Rabbany, “Tightly coupled integration of vol. 70, no. 1-2, pp. 65–82, 1995.
GPS precise point positioning and MEMS-based inertial systems,” GPS [38] P. Teunissen and G. Giorgi, “To what extent can standard GNSS ambiguity
Solutions, vol. 19, no. 4, pp. 601–609, 2015. resolution methods be used for single-frequency epoch-by-epoch attitude
[15] Z. Gao, H. Zhang, M. Ge, X. Niu, W. Shen, J. Wickert, and H. Schuh, determination?,” in Proceedings of the ION GNSS Meeting, pp. 235–242,
“Tightly coupled integration of multi-GNSS PPP and MEMS inertial 2009.
measurement unit data,” GPS Solutions, vol. 21, no. 2, pp. 377–391, [39] S. J. Julier and J. K. Uhlmann, “Unscented filtering and nonlinear
2017. estimation,” Proceedings of the IEEE, vol. 93, pp. 401–422, Mar. 2004.
[40] M. L. Psiaki and S. Mohiuddin, “Relative navigation of high-altitude
[16] S. Vana, “Low-cost, triple-frequency multi-GNSS PPP and MEMS
spacecraft using dual-frequency civilian CDGPS,” in Proceedings of the
IMU integration for continuous navigation in urban environments,” in
ION GNSS Meeting, pp. 1191–1207, 2005.
Proceedings of the ION GNSS+ Meeting, pp. 3234–3249, 2021.
[41] L. Wang and S. Verhagen, “A new ambiguity acceptance test threshold
[17] A. Elmezayen and A. El-Rabbany, “Ultra-low-cost tightly coupled triple-
determination method with controllable failure rate,” Journal of Geodesy,
constellation GNSS PPP/MEMS-based INS integration for land vehicular
vol. 89, no. 4, pp. 361–375, 2015.
applications,” Geomatics, vol. 1, no. 2, pp. 258–286, 2021.
[42] G. N. Green and T. E. Humphreys, “Data-driven generalized integer
[18] K. Nagai, M. Spenko, R. Henderson, and B. Pervan, “Evaluating aperture bootstrapping for high-integrity positioning,” IEEE Transactions
INS/GNSS availability for self-driving cars in urban environments,” in on Aerospace and Electronic Systems, vol. 55, no. 2, pp. 757–768, 2018.
Proceedings of the ION International Technical Meeting, pp. 243–253, [43] K. M. Pesyna, Jr., T. Novlan, C. Zhang, R. W. Heath, Jr., and T. E.
2021. Humphreys, “Exploiting antenna motion for faster initialization of
[19] Z. Yang, Z. Li, Z. Liu, C. Wang, Y. Sun, and K. Shao, “Improved robust centimeter-accurate GNSS positioning with low-cost antennas,” IEEE
and adaptive filter based on non-holonomic constraints for RTK/INS Transactions on Aerospace and Electronic Systems, vol. 3, Aug. 2017.
integrated navigation,” Measurement Science and Technology, 2021. [44] M. L. Psiaki, “Kalman filtering and smoothing to estimate real-valued
[20] S. Hong, M. H. Lee, H.-H. Chun, S.-H. Kwon, and J. L. Speyer, states and integer constants,” Journal of Guidance, Control, and Dynam-
“Observability of error states in GPS/INS integration,” IEEE Transactions ics, vol. 33, no. 5, pp. 1404–1417, 2010.
on Vehicular Technology, vol. 54, no. 2, pp. 731–743, 2005. [45] Y. Bar-Shalom, X. R. Li, and T. Kirubarajan, Estimation with Applications
[21] D. Medina, J. Vilà-Valls, A. Hesselbarth, R. Ziebold, and J. Garcı́a, “On to Tracking and Navigation. New York: John Wiley and Sons, 2001.
the recursive joint position and attitude determination in multi-antenna [46] P. Teunissen and S. Verhagen, “The GNSS ambiguity ratio-test revisited:
GNSS platforms,” Remote Sensing, vol. 12, no. 12, p. 1955, 2020. a better way of using it,” Survey Review, vol. 41, no. 312, pp. 138–151,
[22] M. L. Psiaki, B. W. O’Hanlon, S. P. Powell, J. A. Bhatti, K. D. Wesson, 2009.
T. E. Humphreys, and A. Schofield, “GNSS spoofing detection using [47] P. Teunissen, “Integer aperture GNSS ambiguity resolution,” Artificial
two-antenna differential carrier phase,” in Proceedings of the ION GNSS+ Satellites, vol. 38, no. 3, pp. 79–88, 2003.
Meeting, (Tampa, FL), Institute of Navigation, 2014. [48] P. Teunissen, “Best integer equivariant estimation for elliptically con-
[23] P. Teunissen, “The LAMBDA method for the GNSS compass,” Artificial toured distributions,” Journal of Geodesy, vol. 94, no. 9, pp. 1–10, 2020.
Satellites, vol. 41, no. 3, pp. 89–103, 2006. [49] J. E. Yoder, “Low-cost inertial aiding for deep-urban tightly-coupled
[24] G. Giorgi and P. J. Teunissen, “Carrier phase GNSS attitude determination multi-antenna precise GNSS,” Master’s thesis, The University of Texas
with the multivariate constrained LAMBDA method,” in 2010 IEEE at Austin, 2021.
Aerospace Conference, pp. 1–12, IEEE, 2010.
[25] S. Wu, X. Zhao, C. Pang, L. Zhang, Z. Xu, and K. Zou, “Improving
ambiguity resolution success rate in the joint solution of GNSS-
based attitude determination and relative positioning with multivariate
constraints,” GPS Solutions, vol. 24, no. 1, pp. 1–14, 2020.
[26] P. Henkel and C. Günther, “Reliable integer ambiguity resolution:
multi-frequency code carrier linear combinations and statistical a priori
knowledge of attitude,” Navigation, Journal of the Institute of Navigation,
vol. 59, no. 1, pp. 61–75, 2012.
[27] P. Fan, W. Li, X. Cui, and M. Lu, “Precise and robust RTK-GNSS
positioning in urban environments with dual-antenna configuration,”
Sensors, vol. 19, no. 16, p. 3586, 2019.
[28] R. Hirokawa and T. Ebinuma, “A low-cost tightly coupled GPS/INS for
small uavs augmented with multiple GPS antennas,” Navigation, Journal
of the Institute of Navigation, vol. 56, no. 1, pp. 35–44, 2009.
[29] P. Henkel, A. Sperl, U. Mittmann, T. Fritzel, R. Strauss, and H. Steiner,
“Precise 6D RTK positioning system for UAV-based near-field antenna
measurements,” in 2020 14th European Conference on Antennas and
Propagation (EuCAP), pp. 1–5, IEEE, 2020.
[30] J. E. Yoder, P. A. Iannucci, L. Narula, and T. E. Humphreys, “Multi-
antenna vision-and-inertial-aided CDGNSS for micro aerial vehicle pose
estimation,” in Proceedings of the ION GNSS+ Meeting, (Online), 2020.
[31] T. E. Humphreys, R. X. T. Kor, and P. A. Iannucci, “Open-world virtual
reality headset tracking,” in Proceedings of the ION GNSS+ Meeting,
(Online), 2020.
[32] L. Narula, D. M. LaChapelle, M. J. Murrian, J. M. Wooten, T. E.
Humphreys, J.-B. Lacambre, E. de Toldi, and G. Morvant, “TEX-CUP:
The University of Texas Challenge for Urban Positioning,” in Proceedings
of the IEEE/ION PLANSx Meeting, 2020.
[33] J. Sola, “Quaternion kinematics for the error-state Kalman filter,” arXiv
preprint arXiv:1711.02508, 2017.

16