JOURNAL OF LATEX CLASS FILES, VOL. 18, NO.

9, SEPTEMBER 2020 1

Online IMU-odometer Calibration using GNSS

Measurements for Autonomous Ground Vehicle
Localization
Baoshan Song, Xiao Xia, Penggao Yan, Yihan Zhong,
Weisong Wen, Member, IEEE, and Li-Ta Hsu, Senior Member, IEEE

Abstract—Accurate calibration of intrinsic (odometer scaling need of the dead-reckoning, accurate calibration of ego-sensor
arXiv:2510.08880v1 [[Link]] 10 Oct 2025

factors) and extrinsic parameters (IMU–odometer translation intrinsic parameters and inter-sensor extrinsic parameters are
and rotation) is essential for autonomous ground vehicle localiza- prerequisites to reduce the effect of systematic errors. For
tion. Existing GNSS-aided approaches often rely on positioning
results or raw measurements without ambiguity resolution, and examples, inaccurate IMU bias, odometer scaling factors and
their observability properties remain underexplored. This paper extrinsic spatial transformation between IMU-odometer may
proposes a tightly coupled online calibration method that fuses cause rapid drift [7], [8]. Thus, their calibrations are required.
IMU, odometer, and raw GNSS measurements (pseudo-range, Camera and LiDAR aided IMU-odometer calibration is
carrier-phase, and Doppler) within an extendable factor graph feasible but could drift in long term: In recent years,
optimization (FGO) framework, incorporating outlier mitigation
and ambiguity resolution. Observability analysis reveals that two visual sensors and LiDAR have became popular in navigation
horizontal translation and three rotation parameters are observ- systems and they are also used to calibrate the IMU-odometer
able under general motion, while vertical translation remains un- parameters [9]. Based on the baseline like VINS [10], MSCKF
observable. Simulation and real-world experiments demonstrate [11] and LOAM [12], the camera and the LiDAR have been
superior calibration and localization performance over state-of- employed to aid IMU-odometer calibration. These methods
the-art loosely coupled methods. Specifically, the IMU-odometer
positioning using our calibrated parameters achieves the absolute leverage sliding-window estimation and tightly coupled mea-
maximum error of 17.75 m while the one of LC method is surements to mitigate nonlinearity and non-Gaussian noise.
61.51 m, achieving up to 71.14% improvement. To foster further Variations between these works exist in the number of en-
research, we also release the first open-source dataset that coders and motion models used, ranging from simple 1D
combines IMU, 2D odometer, and raw GNSS measurements from single-wheel models [2], [4], [9], [13], [14], [15], 2D multi-
both rover and base stations. To support future work, we release
the first open-source dataset combining IMU, 2D odometer, and encoder systems [3], [16], [17], [18] and even barometer-
raw GNSS measurements from rover and base stations. augmented 3D setups [19]. Specifically, a 1D single-wheel
type is to employ a single wheel encoder to provide forward
Index Terms—GNSS, Sensor fusion, extrinsic calibration, fac-
tor graph optimization, autonomous vehicles. linear velocity [14]. After that, some authors employ various
models, such as skid-steering, differential and instantaneous
centers of rotation (ICR) [3], to fuse measurements from more
I. I NTRODUCTION encoders to generate 2D measurements. Recently, barometer is
Localization for autonomous ground vehicles: Localization integrated in the system to provide measurement in altitude to
is a fundamental requirement for AGV, supporting intelligent provide full 3D measurements with multiple wheel encoders
transportation applications such as delivery, patrolling, search, [19]. However, a barometer is usually used on aerial vehicles
and rescue [1]. An IMU and an odometer are two common but not common on ground vehicles. Observability analysis has
sensors to provide acceleration, velocity and angular velocity also been conducted in some works to understand calibration
for navigation [2]. Generally, they are less susceptible to mechanisms [3], [14], [16]. With the help of a camera, the
environmental changes and can be used as dead-reckoning extrinsic and intrinsic of related to 2D odometer can be
sensors which can incorporate other external sensors (e.g., calibrated. Despite these advances, camera and LiDAR aided
camera [3], light detection and ranging (LiDAR) [4] and GNSS methods are limited by their properties: (1) visual features are
[5]) to achieve driftless positioning. The problem is, these not always reliable in all environments due to light changes
external sensors are sensitive to environmental conditions. For and LiDAR measurements could be limited by poor geometry;
examples, the camera is affected by light changes; LiDAR (2) camera and LiDAR measuring relies on precise intrinsic
needs signal reflection from blocks; and GNSS is degraded parameters and is sensitive to camera hardware configuration;
in urban canyons due to multipath effects and none-line- and (3) camera and LiDAR aided calibration based on local
of-sight (NLOS) receptions. Consequently, it is critical to odometry could drift after long-term operation in large-scale
enhance the robustness of the IMU and the odometer fused environments. These limitations lead to the first research gap.
dead-reckoning method (e.g., IMU-odometer) [6]. To meet the GNSS provides driftless measurements but affected by
outliers in degraded environments: Global Navigation Satel-
The authors are with The Department of Aeronautical and Aviation
Engineering, Hong Kong Polytechnic University, Hong Kong. (e-mail: lite System (GNSS) provides globally consistent, drift-free
[Link]@[Link]). position measurements, making it a natural complement for
JOURNAL OF LATEX CLASS FILES, VOL. 18, NO. 9, SEPTEMBER 2020 2

TABLE I
R ELATED WORKS ON SENSOR - AIDED IMU- ODOMETER CALIBRATION

Representative Supported Odometer Estimator Couple Open- Amb. Obs.

Work Sensors Meas. Mode source Res. Analysis
Liu [2], Zhao camera-IMU- 1D Graph TC ✗ – ✗
[13], Zhang [15] odometer
Zhao [14] camera-IMU- 1D Graph TC ✗ – ✓
odometer
Zuo [3], [16] camera-IMU- 2D Graph TC ✗ – ✓
odometer
Lee [17] camera-LiDAR- 2D Graph TC ✓ – ✗
IMU-odometer
Zhang [4] LiDAR-IMU- 1D Filter TC ✗ – ✗
odometer
Sun [20] UWB-IMU- 2D Graph TC ✗ – ✗
odometer
Wu [21] GNSS-IMU- 1D Filter LC ✗ ✗ ✓
odometer
Bai [22] GNSS-IMU- 2D Graph LC ✗ ✗ ✗
odometer
Bai [23], [24] GNSS-IMU- 1D Graph LC ✗ ✗ ✗
odometer
Che [19] GNSS-camera- 2D Graph TC ✗ ✗ ✗
IMU-odometer-
barometer
Ours GNSS-IMU- 2D Graph TC Dataset ✓ ✓
odometer

IMU-odometer calibration in outdoor environments [5]. Con- These gaps motivate the development of a robust GNSS-
sidering the number of wheel encoders, the GNSS-aided IMU- aided IMU-odometer calibration framework that can reliably
odometer calibration works also include 1D and 2D cases. operate in degraded environments, handle raw GNSS mea-
To the best of our knowledge, [21] is among the earliest surements with ambiguity resolution, and provide a benchmark
works that utilize GPS measurements to aid IMU-odometer dataset for the community. Therefore, our major objective is to
misalignment calibration. We call the GNSS positioning aided fill the mentioned two groups of gaps. To achieve this goal, this
calibration as loose coupling (LC) and there are still some paper proposes a GNSS-aided INS-odometer online calibration
limitations: (1) reliance on sufficient satellites; and (2) in- method employing raw GNSS carrier phase measurements
accurate uncertainty of GNSS positioning. A potential to and perform analysis on both observability and AR. The
solve the problems is the use of raw GNSS measurements. comparison of our proposed method and existing methods is
To overcome the limitations of LC, fuse the raw GNSS illustrated in TABLE I. Our main contributions are as follows:
measurements in a tightly coupled manner. Recently, raw • Utilize raw GNSS measurements to aid IMU-
GNSS measurements are applied in multi-sensor fusion local- odometer online calibration with outlier mitigation
ization. Compared with the former LC systems, the raw GNSS and ambiguity resolution: The integration method is
measurements aided methods are termed as tight coupling based on an extendable factor graph optimization (FGO)
(TC). TC approaches integrate raw GNSS such as pseudo- framework [28]. In this FGO, the navigation state and
range and Doppler measurements have been used to couple IMU-odometer extrinsic parameters are estimated jointly,
with a LiDAR [25], a camera [26] and an ultra wide band in which GNSS raw measurements including pseudo-
(UWB) [27]. In these papers, pseudo-range measurements range, carrier-phase and Doppler are used. To the best
provide the globally consistent range at meter level accuracy of our knowledge, this paper is the first to employ these
and Doppler measurements provide the range rate at decimeter GNSS raw measurements together with outlier mitigation
per second level accuracy. To eliminate the effect of raw and ambiguity resolution to calibrate the IMU-odometer
GNSS pseudo-range outliers, carrier-phase measurements with extrinsic and odometer’s intrinsic parameters.
ambiguity variables are also used as range constraints at • Derive the observability of the online calibration:
centimeter level accuracy. In recent years, the carrier-phase Observability analysis is conducted to show that with gen-
measurement is fused with the odometer for positioning [5] eral motion two horizontal translation and three rotation
and calibration [19]. The results show a superior performance parameters between IMU-odometer are observable, and
over the methods only using pseudo-range. However, there are only the vertical translation parameter at the body frame
also some limitations in current works: 1) the impact of carrier- is unobservable.
phase ambiguity resolution in a GNSS-aided IMU-odometer • Experimental verification and open-sourced datasets:
calibration system is not discussed; 2) observability analysis Experiments are conducted with both simulated and real-
is lacked; 3) there are no open-sourced dataset including both world datasets. Ablation tests are also investigated to
base/rover station GNSS measurements and IMU-odometer evaluate the effect of the GNSS-IMU lever-arm in the
measurements. These limitations lead to the second gap. system. The proposed calibration method outperforms
JOURNAL OF LATEX CLASS FILES, VOL. 18, NO. 9, SEPTEMBER 2020 3

the loosely coupled methods using RTK results and forward linear velocity and bearing angular velocity [31]. This
shows superior accuracy and robustness in various GNSS- paper adopts the same IMU model as the one in [30], hence
degraded environments. Also, this is the first paper to only GNSS and odometer models are introduced as below
open source dataset including IMU, 2D-odometer and raw while the details of the IMU model can be referred to [30]
GNSS measurements from both rover and base stations. and [32].
The rest of the paper is organized as follows. Sec II to
Sec. IV introduce the tightly-coupled factor graph framework, A. GNSS Modeling
including the framework overview in Sec. II, key sensor GNSS raw measurement models including pseudo-range P ,
modeling in Sec. III and factor graph structure in Sec. IV. carrier-phase L and Doppler shift D between the satellite s
Afterwards, Sec. V derives the observability analysis. Then and receiver r can be formulated as [33]:
Sec. VI and Sec. VII evaluate the performance by simulation 
and real-world experiments. Finally, conclusions are drawn, 

 P = ρ + I + T + c(tr − ts ) + c(br − bs ) + ϵP ,

together with an outlook in Sec. VIII.
L = ρ − I + T + c(tr − ts ) + cλ(Br − B s ) + λN + ϵL ,

 λD = ∆pT ∆v + c(t˙ − t˙s ) + ϵ .


II. GNSS- CONSTRAINED IMU- ODOMETER PARAMETER ρ r D
CALIBRATION OVERVIEW (2)
where ρ is the geometric distance between the phase center of
The proposed IMU-odometer parameters calibration method
the satellite and receiver antennas; ∆p and v are the relative
is based on a GNSS-INS-odometer tightly coupled integrated
position and velocity between the satellite and receiver antenna
estimator via a nonlinear factor-graph optimization pipeline,
phase center (ρ = ∥∆p∥); I and T denote the ionosphere
which is designed for a differential-drive based robot car using
and troposphere delays; c is the speed of light; tr and ts are
GNSS receivers (both rover and reference stations), an IMU
the clock offsets of the receiver and satellite; t˙r and t˙s are
and an 2D odometer. The framework of the pipeline consists
the clock shift rate of the receiver and satellite; br and bs
of two parts: (1) sensor modeling of GNSS-IMU-odometer
are the receiver and satellite code hardware bias; Br and B s
using sensor inputs to eliminate systematic noises of raw mea-
are the receiver and satellite phase bias; λ and N are the
surements; (2) factor graph optimization based on the sensor
wavelength and integer ambiguity of carrier signal; ϵP , ϵL
modeling to estimate the optimal calibration parameters. The
and ϵD represent the sum noise and multipath error of the
symbols frequently used in this paper are also introduced to
corresponding measurements.
define the mathematic models clearly. We present the multi-
In view of the systematic errors in GNSS raw measure-
sensor coordinate frame denotation as follows:
ments, the double-differenced (DD) model is usually adopted
• w: the local world frame fixed to the origin and direction
to eliminate most of systematic errors in pseudo-range P
of the first IMU pose (i.e., position and attitude) [26]; and carrier-phase L [28]. Specifically, the DD model includes
• n: the local navigation frame with the same origin as w
single-differenced (SD) model between satellites and receivers:
frame and east-north-up (ENU) direction [29]; SD model between satellites can eliminate the receiver-related
• e: the global Earth-Centered, Earth-Fixed (ECEF) frame
bias; the SD model between receivers can eliminate the
of GNSS fixed to the Earth [7]; satellite-related bias; the DD model can greatly reduce the
• b: the moving body frame with the origin fixed to IMU
atmospheric delays. After applying DD model, the GNSS DD
center and sensor defined direction [30]; measurements can be expressed as:
• m: the moving vehicle-mounted frame with the origin 
fixed to the vehicle center and right-forward-up (RFU)  P
DD = ρDD + ϵPDD ,
direction . (3)
 L =ρ +λ N +ϵ .
DD DD DD DD LDD
Additionally, the transformation from a to b with a as the
reference frame is denoted as translation pab and rotation Rba where PDD and LDD are the DD pseudo-range and DD
separately. In addition, the skew-symmetric operator is defined carrier-phase; ρDD is the DD geometric distance between
as:   satellite and receiver antenna phase center; λDD and NDD
0 −a3 a2 are the wavelength and ambiguity of DD carrier-phase; ϵPDD
⌊a×⌋ =  a3 0 −a1  (1) and ϵLDD represent the noise and residual bias of the corre-
−a2 a1 0 sponding measurements.
 T Notably, although the DD operator eliminates most system-
where a = a1 a2 a3 ∈ R3 .
atic errors within the GNSS measurements, the unmodeled
noises and outliers including multi-path and NLOS effects
III. S ENSOR M ODELING are amplified. Therefore, in this work, we also employ our
In this section, we describe in detail the sensor modeling previously proposed two-stage GNSS outlier detection method
based on raw sensor measurement inputs. In terms of inputs for [34], which first leverages Doppler measurements to detect
sensor modeling, GNSS rover and reference stations provide pseudo-range outliers and then uses pre-integrated IMU and
broadcast ephemeris and raw measurements such as pseudo- odometer constraints to refine the rejection of remaining errors,
range, carrier-phase and Doppler measurements; IMU outputs resulting in significantly improved positioning accuracy and
linear accelerations and angular velocities; odometer offers robustness.
JOURNAL OF LATEX CLASS FILES, VOL. 18, NO. 9, SEPTEMBER 2020 4

B. Odometer Modeling
Different from existing work [22] for autonomous driving,
we focus on a general odometer model for ground robots.
One of the general models is the differential drive model
which can provide planar kinematic constraints provided that
measurements between time k and k + 1 are collected on a
supporting planar manifold [35]. According to the differential
drive model, we model the true odometer output fixed to the
m frame as: Fig. 1. The proposed factor graph with a sliding window overview, where the
 red box denotes the marginalization factor, the orange, blue, and green boxes
 v m = (1 + s )(v̂ m − ϵ m ), denote GNSS, IMU, odometer measuring and relative state constraints, and
v v
(4) the orange, blue, and green circles denote GNSS, IMU, and odometer states.
 ω m = (1 + s )(ω̂ m − ϵ m )
ω ω

where v m and ω m are the true linear velocity in the forward where X denotes the robot state set with n state in the sliding
direction and angular velocity in the up direction separately; window, and xk signifies the state at epoch k consisting of the
sv and sω are the scaling factors of linear velocity and angular IMU state xI , GNSS state xG and odometer state xO :
velocity from the odometer; v̂ m and ω̂ m present the forward
linear velocity and bearing angular velocity measurements of xk = [xTI,k , xTG,k , xTO,k ]T (7)
the odometer; ϵvm and ϵωm are the corresponding measuring
Although xG and xO will not be added to the same state
noises. The odometer scaling factors are modeled as random
due to asynchronous measuring of GNSS and odometer. For
walk [36]: s˙v = ϵsv and s˙ω = ϵsω .
presentation simplicity, we describe them explicitly in the
Under the planar motion assumption, the odometer measure- same state vector. Specifically, the INS states can be written
ments can be projected to vm ∈ R3 and ω m ∈ R3 in three as:
dimensions. The projected odometer measuring model can be T w T w T
xI,k = [pw T T T
b,k , vb,k , qb,k , ba,k , bg,k ] (8)
expressed as:
 mT T T where pw w
b is the robot position in the world frame; vb is the
ωm T = P vm ωm .

v (5) robot velocity; qw is the robot attitude quaternion; b a and bg
b
T are the bias of the accelerometer and gyroscope. For the GNSS
where P = 0 eT2 0 0 0 eT3

is the projection states, considering the convenience of selecting new reference
matrix with ei being the 3×1 unit vectors with the ith element satellite, we keep the single-differenced ambiguity set NSD,k
of 1. between two receiver stations together with the receiver clock
shift rate ṫr,k in the state vector (because of D is not single
IV. FACTOR G RAPH O PTIMIZATION differenced) as below:
In this section, we introduce the factor graph optimization xG,k = [ṫr,k , NSD,k ] (9)
framework based on the sensor modeling section.
The proposed factor graph framework is shown in Fig. For the odometer state, both of odometer scaling factors
1, containing two kinds of nodes-variable nodes and factor and IMU-odometer extrinsic transformation (translation pbm
nodes. Variable nodes denote the states needed to be estimated and rotation qbm ) are included. Note that new scaling factors
and factor nodes represent the sensor observation constraints are inserted at every odometer update moment and there is
related to the states [24]. All the nodes are formulated in a only one IMU-odometer extrinsic transformation in the sliding
sliding window for numerical programming by a nonlinear window.
T T
estimator after which the estimator will output the navigation xO,k = [svk , sωk , pbm , qbm ]T (10)
state, and use the latest estimated IMU-odometer parameters The cost function r to be minimized includes the combina-
for sensor calibration at the next factor graph estimating tion of sensor measurement constraints is formulated as:
moment. The following section firstly describes the state
equations and then derives multiple sensor constraints, includ- r = r M + r G + r I + rO (11)
ing GNSS, IMU and odometer constraints. To simplify the
where rM denotes the prior marginalization constraint; rG
expression, we use G, I and O to denote GNSS , IMU and
denotes the GNSS constraint; rI and rG represent the inertial
odometer, respectively.
constraint and the odometer constraint, both of which are
constructed by measurement integration. Due to page limit,
A. State Formulation we omit the details of rM and readers are encouraged to refer
to [37]. The rest constraints are introduced as below.
To begin with, we define the complete states vector X
including multiple variable nodes xk within a sliding window
as: B. GNSS Constraints
Since we define the position fixed to IMU b frame in the
X = [xT1 , ..., xTk , ..., xTn ]T (6) state, a lever arm pgb between IMU and GNSS is utilized to
JOURNAL OF LATEX CLASS FILES, VOL. 18, NO. 9, SEPTEMBER 2020 5

associate the GNSS antenna position peg and velocity vge with D. Odometer constraints
the IMU-center position pw w
b and velocity vb . Specifically, Based on the projected odometer measuring model (5) and
considering the lever arm between IMU and GNSS antenna state model (10), we can transform the odometer measure-
is constant, the GNSS antenna position used in (3) is: ments from m-frame to b-frame with IMU-odometer extrinsic
transformation. The odometer velocity in b-frame can be

 pe = Re Rn (pw + Rw · pb ),
g n w b b g expressed by:
(12)
 ve = Re Rn (vw + ω b × · pb ). 
g n w b ib g  vb = Rb vm − ω b × pb ,
m ib m
Based on the GNSS model (2) (3) and state transforming (18)
 ω b = Rb ω m .
model (12), the cost function of GNSS could be generically m
written in the following form: Then similar to the IMU constraints, the integral measurements
2 2 2 in b-frame between two epochs k and k + 1 can be written
rR = PDD − P̂DD ) + LDD − L̂DD ) + D − D̂) as the equations (19), where δpkO,k+1 and δqkO,k+1 denote
ΣP ΣL ΣD
(13) integral odometer position and attitude measurements; Ω′
where ΣP , ΣL and ΣD present the covariance matrix of denote the quaternion right-hand multiplier in [30].
corresponding measurements computed by the elevation an- To derive the uncertainty of the integral measurements,
gle based model [32]. Moreover, we utilize the LAMBDA continuous error propagation model of the integral odometer
method to perform ambiguity resolution in this paper [38]. If measurements is given as:
the DD ambiguity resolution succeeds, the integer ambiguity  k   k   
constraints will also be employed as a factor in the later sliding δ ṗO,t δpO,t ϵv m
δ q̇k   k 
window:  O,t  = Ft · δqO,t  + Gt · ϵωm 
 
 δ ṡv,t   δsv,t   ϵsv  (21)
2
i j
rAR = N̂DD,ij − g(NSD − NSD ) (14) δ ṡω,t δsω,t ϵsω
ΣAR

where N̂DD,ij denotes the integer double differenced ambi- The details of the Jacobian matrix Ft and Gt in the
guity estimated by the LAMBDA; NSD i
denotes the single uncertainty propagation equation are given as in (20) and (22).
differenced integer ambiguity state between rover and base −Rbbkt Rbm e2 (1 + sv )
 
0 0 0
station for satellite i; function g denotes the single differenced  0 −Rbm e3 (1 + sω ) 0 0
Gt =   (22)
model between satellite i and j; ΣAR denotes the covariance  0 0 1 0
matrix of the integer ambiguity and it is defined as a diagonal 0 0 0 1
matrix with the element 1 × 10−6 in this paper. Based on the integral odometer measurements, we derive
Hence, the residual model of GNSS can be presented as the the cost function of odometer kinematic in the form:
sum of them: 2
rG = rR + rAR (15) rO = ∥xk ⊟ f (xk−1 , Ov , Oω )∥ΣO (23)
where Ov and Oω denote the odometer measurements between
C. Inertial Constraints epoch k − 1 and k; ΣO represents the covariance matrix of
To deal with high-rate IMU measurements, a pre-integration the integral measurements, which is given by (21).
model is employed to provide relative motion constraints
between two consecutive epochs. Plenty of existing works have V. O BSERVABILITY A NALYSIS
introduced the pre-integration model of IMU in details and In addition to the FGO framework, observability analysis for
comprehensive definition and derivation can be found in [30]. specific parameters is derived in this section. This is because
In short, the IMU pre-integration constraint factor provides the the calibration method utilizes projected 3D kinematic mea-
relative connection between two states xk−1 and xk , which surements from differential-drive based odometer, so it will be
can be written as: interesting to analyze the local observability of the odometer
2 scaling factor and transformation parameter between IMU and
rIM U = ∥xk ⊟ f (xk−1 , Ia , Ig )∥ΣI (16)
the vehicle with the raw planar odometer constraints. Since
where ⊟ denotes the generalized minus on manifold defined in the local observability of scaling factors is derived in [39], we
[30]; Ia and Ig denote the IMU accelerometer and gyroscope focus on observability of extrinsic transformation parameters
measurements between epoch k − 1 and k; ΣI represents following the method proposed in [3]. To simplify the process,
the covariance matrix of the integral measurements, which we assume that the IMU bias is estimated accurately with
is calculated by the same model in [32]. In addition, we GNSS constraints and the system is able to provide virtual
also use the IMU measurements to detect the motion of the linear and angular velocity measurements in the IMU’s body
AGV and apply motion constraints in [32], such as zero- frame:
velocity update (ZUPT), non-holonomic constraint (NHC) and v̆b = vb + ϵvb , ω̆ b = ω b + ϵωb (24)
heading measurement constraint (HMC), to enhance the state
where v̆b and ω̆ b are the virtual measurements from the
estimation. Therefore, the inertial constraint can be written as:
GNSS/IMU integration. Here we introduce a lemma about the
rI = rIM U + rmotion (17) observability.
JOURNAL OF LATEX CLASS FILES, VOL. 18, NO. 9, SEPTEMBER 2020 6

 Z k+1
bk b m
bk w bk w
 b  b



 R w · p bk+1
= Rw pb k
+ R bt
[R m e2 (1 + s v )v̂ − ωib × pm ] · dt,

 k

 | {z }
δpk

O,k+1
Z k+1 (19)
1 ′ b
bk w
qw ⊗ qbk+1 = Ω [Rm e3 (1 + sω )ω̂ m ]qbbkt · dt .



 2
|k



 {z }

δqk
O,k+1

 j k 
− Rbbkt (Rbm e2 (1 + sv )v̂tm − ωib Rbbkt Rbm e2 v̂tm
 b  b
0 × pm )× 0
   
0
Ft =  − Rbm e3 (1 + sω )ω̂tm × 0 Rbm e3 ω̂tm 
 (20)
0 0 0 0 
0 0 0 0

Lemma 1. With the assumption of planar motion, the Algorithm 1 FGO-based Raw GNSS Aided IMU-Odometer
translation parameter along the vehicle rotation axis is not Online Calibration
locally identifiable, while translation and rotation extrinsic Require: Raw GNSS pseudo-range/carrier-phase/Doppler
parameters along other directions are locally identifiable measurements from rover and base station GNSS
under general motion. receivers, IMU and odometer measurements; initial state
Proof. Substituting (24) into (18) leads to X0 including IMU-odometer parameter xO,0 and its
 b  b initial covariance P0
[(Rm b m
   
r̆v
= b (v̆ + ω̆ × pm ) − v ) (25) Ensure: refined states X̂ including online calibration param-
r̆ω Rm b
b ω̆ − ω
m
eters x̂O ,
Note that the equation only contains virtual measurements 1: % Preprocess raw GNSS data, detect and reject outliers
and the IMU-odometer transformation parameters. A nec- via two-stage outlier mitigation
essary and sufficient condition of the IMU-odometer trans- 2: % Initialize IMU-odometer parameters (xO ) included in
T T
formation [pbm , qbm ]T to be locally identifiable is that the state variables (X)
following observability matrix M has full column rank: 3: for each sliding window k = 1, . . . , K do
T 4: Formulate factor graph with inertial, odometer, GNSS
M = M(t0 )T M(t1 )T ... M(ts )T

(26) factors using residual r and state X
where 5: Perform nonlinear optimization to jointly estimate X̂,
 m b 
Rb ω̆ × Rm b
 b  b  resolve DD ambiguities N̂DD using LAMBDA
M(t) = b (v + ω̆ × pm )× (27) 6: If ambiguity ratio test passes, fix integer ambiguities
0 Rm b
b ω̆ ×
7: Update calibration parameters x̂O online
A necessary and sufficient condition of the extrinsic transfor- 8: end for
mation to be locally identifiable is that M has full column 9: return X̂
rank. However, if the robot b car is operating on a plane,
the third column of Rm b ω̆ × is zero, which demonstrates
that the translation parameter along the Z-axis of m-frame measurements of both rover and base stations are simulated
is not identifiable. For other parameters, if the robot can based on the GREAT software [41] with precise ephemeris
provide sufficient acceleration and rotation, they will be locally and clock products from NASA CDDIS [42] and the above
identifiable. This completes the proof. ground-truth trajectory from Gazebo. The IMU and odometer
measurements are generated with given sensor configuration.
VI. S IMULATION A NALYSIS The brief introduction of simulated sensor configuration is
A. Introduction of the Simulation presented in Table II.
1) simulation setup: To analyze the performance of the To evaluate the observability conditions for general ground
proposed calibration method, a simulation test is performed robots, a typical trajectory with only horizontal motion is sim-
using numerically synthetic raw GNSS, IMU, and odome- ulated for the AGV. The ground-truth trajectory and velocity
ter measurements. With the simulation test, it is possible in the body frame are illustrated in Fig. 2, consisting of several
to set true extrinsic parameters and appropriate measuring types of motion to test the calibration process (e.g., standing
noise to evaluate the accuracy and observability of calibration stationary, accelerating in straight lines and going around a
while mostly eliminating uncontrolled uncertainties like GNSS circle).
multi-path effect, IMU bias instability and wheel slip. For 2) Evaluated Methods: Due to the bias instability of
the simulation, a Gazebo 11.0 simulator [40] is employed to MEMS-IMU, it is difficult for IMU-odometer fusion to cali-
provide the ground-truth trajectory, which has been widely ap- brate the IMU-odometer parameters. Besides, current sensors
plied to support autonomous robot systems. Then, raw GNSS used in IMU-odometer calibration methods includes camera
JOURNAL OF LATEX CLASS FILES, VOL. 18, NO. 9, SEPTEMBER 2020 7

B. Calibration results in simulation tests

We first compare the calibration results from the methods
above in Fig. 3, including errors of IMU-odometer extrinsic
transformation and odometer scaling factors. At 20 s, the AGV
quickly raise its forward velocity to 0.25 m/s and keep constant
velocity in about 20 s. Since the velocity in most of this
period is constant. Most of the translation and rotation extrinsic
parameters hardly change except yaw angle and scaling for v
(sv ). At 45 s, the AGV starts to circle around and most of
the parameters converge in seconds. In the following time,
all the extrinsic parameters keep stable until the end of the
Fig. 2. Simulated trajectory and odometer measurements. simulation test. After the calibration, the absolute calibration
errors at final epochs are concluded in TABLE III. Notably,
TABLE II translation error for Z-axis of m-frame does not change during
S IMULATION CONFIGURATION
the 2D motion. This verifies that the vertical translation at m-
Simulation settings values frame is unidentifiable, so we do not analyze its error below.
GNSS constellations GPS, Galileo, BDS
For scheme TC, the translation errors are 0.06 m for X-axis
and 0.05 m for Y-axis. The rotation errors are 0.24°, 0.05°
Satellite number 19∼21
and 3.29° for roll, pitch and yaw angle. The estimation errors
Position dilution of precision 1.25∼1.45
(PDOP)
for linear/angular velocity scaling factors are 6.0 × 10−4 and
3.1 × 10−3 separately. For scheme TC-LE-C, the calibration
GNSS frequency L1, L2
p accuracy is worse. The translation errors are 0.11 m for X-axis
Pseudo-range noise 0.3 ∗ (1 + 1/sin(El))m
p and 0.09 m for Y-axis. The rotation errors are 0.24°, 0.16°
Carrier-phase noise 0.003 ∗ (1 + 1/sin(El))m and 4.26° for roll, pitch and yaw angle. The estimation errors
GNSS sampling rate 1 Hz for linear/angular velocity scaling factors are 2.0 × 10−3 and
Gyroscope bias (not employed in 900 °/h 1.1 × 10−4 separately. Compared to the TC-LE-C scheme, the
estimation) calibration accuracy of TC-LE-OL is improved, especially at
Accelerometer bias (not employed 5 mGal the translation part. The translation errors are 0.06 m for X-
in estimation)
p axis and 0.05 m for Y-axis. Other estimated parameters are
Angle random walk noise 20 °/ (h) almost the same. The rotation errors are 0.26°, 0.11° and
p
Velocity random walk noise 0.1 m/s/ (h) 4.60° for roll, pitch and yaw angle. The estimation errors
IMU sampling rate 100 Hz for linear/angular velocity scaling factors are 2.0 × 10−3 and
odometer linear velocity noise 0.01 m/s 3.1 × 10−3 separately.
odometer angular velocity noise 1 °/s Apart from GNSS-IMU lever-arm, a further insight is given
odometer scaling factor 0, 0 into the initial values and the uncertainty of the calibration. As
p p is shown in Fig. 4, we disturb the initial value with Gaussian
odometer random walk noise 0 (Hz), 0 (Hz)
noise and perform 100 Monte-Carlo simulations for statistics.
odometer sampling rate 25 Hz
In the random simulation results, it can be observed that the
choice of initial values does have a noticeable influence on the
convergence process. Nevertheless, the translation extrinsic pa-
[43], GNSS [24] and other sensors [17] while each kind of
rameters are consistently estimated within 0.2 m accuracy, with
sensor is designed for specific scenes. In this simulation, we
the exception of the z-axis which remains unobservable under
focus on the accuracy and observability performance with
the given configuration. Moreover, the standard deviation (std)
ideal measuring configurations and the comparison with other
of the estimates provides a certain level of indication regarding
methods is performed in the field test. In the following, we
the convergence behavior, but it does not fully cover the actual
compare the proposed method with the ground truth to validate
error distribution. This suggests that in practical applications
the calibration accuracy. At the same time, we also consider
the reported uncertainty should be treated as a reference rather
the effect of GNSS-IMU lever-arm error. The introduction of
than a strict bound.
compared methods is shown as below:
(a) TC-LE-C: the tightly coupled (TC)
GNSS/IMU/odometer integration method with constant
C. Discussion: Performance and Limitation of the Online
GNSS-IMU lever-arm (LE-C). The LE error is 0.1 m in each
Calibration
axis of b-frame.
(b) TC-LE-OL: the TC method with online GNSS-IMU The results show the observability of the proposed method.
lever-arm estimation (LE-OL). The LE error is 0.1 m in each Most of the parameters are locally observable and converge
axis of b-frame. in a several seconds under general planar motion and the
(c) TC: the TC method without GNSS-IMU lever-arm precision of the calibration result is corresponding to the
errors. navigation state level. Notably, there is an error within 5° for
JOURNAL OF LATEX CLASS FILES, VOL. 18, NO. 9, SEPTEMBER 2020 8

0.2
X (m) 0.2
Y (m) 0.2
Z (m) 0.010
V Scale
TC-LE-C
0.1 TC-LE-OL 0.1 0.1 0.005
TC
0.0 0.0 0.0 0.000
0.1 0.1 0.1 0.005
0.2 0.2 0.2 0.010
0 50 100 0 50 100 0 50 100 0 50 100
Time (s) Time (s) Time (s) Time (s)
Roll (degrees) Pitch (degrees) 10
Yaw (degrees) 0.010
Omega Scale
0.4 0.4
0.2 0.2 5 0.005
0.0 0.0 0 0.000
0.2 0.2 5 0.005
0.4 0.4
10 0.010
0 50 100 0 50 100 0 50 100

50

0
Time (s) Time (s) Time (s)

10
Time (s)

Fig. 3. Calibration errors in the Simulation, where the pink line denotes the TC method with constant GNSS-IMU lever-arm error (TC-LE-C), the blue line
denotes the TC method with online estimated GNSS-IMU lever-arm (TC-LE-OL), the green one denotes the TC method without GNSS lever-arm disturbance
(TC).

X (m) Y (m) Z (m) V Scale

1.00
0.4 0.4 0.4 0.75

0.50
0.2 0.2 0.2
0.25

0.0 0.0 0.0 0.00

0.25
0.2 0.2 0.2
0.50

0.4 0.4 0.4 0.75

1.00
s

0s

0s

0s

0s

0s

0s

0s

0s
0

60

60

20

60

60

20

60

60

20

60

60

20
12

12

12

12
Err

Std

Err

Std

Err

Std

Err

Std
1

1
Err

Std

Err

Std

Err

Std

Err

Std
Err

Std

Err

Std

Err

Std

Err

Std

Roll (deg) Pitch (deg) Yaw (deg) Omega Scale

10.0 10.0 10.0 1.00

7.5 7.5 7.5 0.75

5.0 5.0 5.0 0.50

2.5 2.5 2.5 0.25

0.0 0.0 0.0 0.00

2.5 2.5 2.5 0.25

5.0 5.0 5.0 0.50

7.5 7.5 7.5 0.75

10.0 10.0 10.0 1.00

0s

0s

0s

0s

0s

0s

0s

0s

0s

0s

0s

0s

0s

0s

0s

0s
60

60

60

60

60

60

60

60
12

12

12

12

12

12

12

12
Err

Std

Err

Std

Err

Std

Err

Std
Err

Std

Err

Std

Err

Std

Err

Std
Err

Std

Err

Std

Err

Std

Err

Std

Fig. 4. Statistics for calibration results and uncertainty in 100 Monte-Carlo simulation with random initial value. ”Err 0s”, ”Err 60s” and ”Err 120s” mean
the calibration error at corresponding epochs. ”Std 0s”, ”Std 60s” and ”Std 120s” mean the standard deviation of these calibrated results.
JOURNAL OF LATEX CLASS FILES, VOL. 18, NO. 9, SEPTEMBER 2020 9

TABLE III
M EAN CALIBRATION ERRORS OVER 40 M ONTE -C ARLO SIMULATIONS . T RANSLATION ERRORS ARE IN METERS , ROTATION ERRORS IN DEGREES ,
SCALING FACTOR ERRORS UNITLESS .

Scheme X (m) Y (m) Z (m) Roll (deg) Pitch (deg) Yaw (deg) V Scale Omega Scale

TC 0.06 0.05 – 0.24 0.05 3.29 6.0 × 10−4 3.1 × 10−3

TC-LE-C 0.11 0.09 – 0.24 0.16 4.26 2.0 × 10−3 1.1 × 10−4
TC-LE-OL 0.06 0.05 – 0.26 0.11 4.60 2.0 × 10−3 3.1 × 10−3

at Section V. In fact, the lack of observability for height

translation has little effect on the navigation performance
because the odometer could sample in high-frequency and
the motion in such a short time could be approximated as
2D motion. Moreover, to further lower down the impact
of unobservable extrinsic parameters and raise the accuracy
IMU-odometer integrated navigation, 2D motion detection is
helpful and promising. We guess that if utilizing multiple
IMU or perception sensors, the detection of 2D motion can
be more accurate and the influence of IMU-odometer vertical
translation parameter error can be further eliminated.

D. Discussion: Impact of the GNSS-IMU lever-arm parameter

disturbance
As mentioned in the methodology, the proposed tightly-
coupled factor graph can be divided into GNSS-IMU and
Fig. 5. AGV platform for field data collection. Blue boxes denote the
evaluating sensor and the Red box denotes the ground-truth system. IMU-odometer fusion parts. In the open-sky scenes, GNSS-
IMU tightly coupled integration is able to perform positioning
at centimeter level with predefined precise lever-arm param-
eters [7]. Here, we analyze the impact of the GNSS-IMU
lever-arm in the calibration process. In the proposed online
calibration framework, the IMU-odometer extrinsic parameters
4 are correlated with the navigation state. Thus the extrinsic
transformation, especially the translation part will absorb the
systematic error of the GNSS-IMU lever-arm. For the scheme
TC-LE-C, it can be noticed that there are obvious offsets in
2 X/Y translation, which is caused by GNSS-IMU lever-arm
3 perturbation. Luckily, thanks to accurate GNSS measurements,
GNSS-IMU lever-arm can also be estimated online in the
same factor graph. For the scheme TC-LE-OL, the accuracy of
1
estimated lever-arm comes to (0.01, 0.012, 0.12) m. Results
in Section VI-B show that the final extrinsic transformation
accuracy in TC-LE-OL is almost the same as the one in
TC. Therefore, the GNSS-IMU lever-arm can be measured
Fig. 6. Google earth trajectory. The blue dots denote measurements over all
by tools at within centimeter level or estimated together in the
the trajectory are used for calibration. The orange dots denote the trajectory estimator. Therefore, the impact of GNSS-IMU lever-arm with
where the measurements are used to test the performance of calibration results. disturbance in centimeter level can be neglected after online
calibration.

extrinsic yaw angle because of the low velocity and short IMU- VII. F IELD T EST
odometer extrinsic translation, which might be decreased if
we improve the precision of heading initial alignment. On the A. Introduction of the field test
other hand, the extrinsic yaw angle is related to the estimated 1) Data collection setups: To verify the performance of
attitude state so high-level IMU can enhance the attitude the proposed calibration method, field experimental tests are
estimation precision [6]. For the translation along vehicle conducted with a real-world ground robot dataset. To the best
rotating direction, according to the result, the translation at the of our knowledge, Existing public ground robot datasets lack
Z-axis of v-frame is not observable, proving that the lemma critical measurements from RTK base station [44] or odometer
JOURNAL OF LATEX CLASS FILES, VOL. 18, NO. 9, SEPTEMBER 2020 10

by a low-cost u-blox-F9P dual-frequency receiver with a

1 3 NovAtel GPS-702-GG antenna at 1 Hz; IMU measurements
are collected by a MEMS-IMU Xsens MTi-30 at 400 Hz;
odometer measurements are generated by an Agilex scout-mini
mobile base with its open-source driver at 50 Hz. Besides, the
ground-truth of navigation states is solved by NovAtel Inertial
Explorer 8.9 software [46] in the RTK/INS tightly coupled and
2 4
smoothing filter mode using measurements from a NovAtel
GNSS receiver and a high-precision tactical-level SPAN-CPT
IMU. The key parameters of the MEMS-IMU and tactical IMU
are presented in [45], which is also shown in TABLE IV. The
baseline between the rover and the base GNSS station is about
4 kilometers, so the atmospheric delay could be eliminated
Fig. 7. Four selected scenes for calibration evaluation. clearly.
In this work, we adopt the NHC when the mean angular
velocity from the gyroscope is less than 5 °/s and the mean
forward velocity from the odometer is greater than 1 m/s
in one second. We also adopt the ZUPT when the mean
angular velocity from gyroscope is less than 0.05 °/s in one
second. The initial uncertainty of the IMU-odometer extrinsic
parameters are 0.5 m and 5 °. The GNSS measurement
uncertainty is 0.3 m, 0.003 m and 0.3 m/s for raw pseudo-
range, carrier-phase andp Doppler. The IMU biaspuncertainty
1 2 3 4
is set as 0.01 rad ∗ (Hz) and 0.02 m/s ∗ (Hz). For
the input of LC, we adopt the RTK result from RTKLIB
with default configuration, including dual-frequency measure-
ments, elevation weighting model and continuous ambiguity
Fig. 8. GPS Satellite number over time in the field test.
resolution mode. In this work, we adopt the NHC when
the mean angular velocity from the gyroscope is less than
[45], making them unsuitable for this paper. Therefore, a 5 °/s and the mean forward velocity from the odometer
self-assembled AGV robot is employed for real-world data is greater than 1 m/s in one second. We also adopt the
collection on Nov. 4, 2024 in the Hong Kong West Kowloon ZUPT when the mean angular velocity from gyroscope is
Cultural park. As shown in Fig. 6 and 7, the change of less than 0.05 °/s in one second. The initial uncertainty of
height is about 1.5 m. There are four typical scenes in the the IMU-odometer extrinsic parameters are 0.5 m and 5 °.
real-world tests, including open-sky, pathway covered by few The GNSS measurement uncertainty is 0.3 m, 0.003 m and
trees, boulevard covered by more trees and half-sky blocking 0.3 m/s for raw pseudo-range, carrier-phase and p Doppler.
scenes. The satellite number of GPS is illustrated in Fig. The IMU biaspuncertainty is set as 0.01 rad ∗ (Hz) and
8, where the red lines denotes the epoch at four typical 0.02 m/s ∗ (Hz). For the input of LC, we adopt the
scenes. The employed ground robot is illustrated in Fig. 5. RTK result from RTKLIB with default configuration, including
During the experiment, GNSS measurements are collected dual-frequency measurements, elevation weighting model and
continuous ambiguity resolution mode. All the data is collected
and time-synchronized using a robot operation system (ROS)
[47], and the local computer time is aligned with the GPS
3 time through Pulse Per Second (PPS) signal and the GPRMC
2 message [26]. Note that the effect of time inconsistency is
related to the motion of the vehicle. For example, for the
1
Velocity (m/s)

ground robot used in this paper with a average forward velocity

0 of about 1.5 m/s, a time misalignment of 10 ms only leads
to 0.015 m, while for a car with a velocity of 20 m/s,
1 the same time error could lead to 0.2 m. Therefore, it is
2 East Velocity (m/s) necessary to consider the time misalignment instability. An
North Velocity (m/s)
Up Velocity (m/s) similar work with a ROS software synchronization method
3
has limited the time misalignment error below 30 ms [48],
equal to 0.045 m positioning error for our AGV. Therefore, the
0

0
00

50

00

50

00

50
15

15

16

16

17

17

time misalignment is insignificant compared to our translation

GPS Second of Day (s) calibration error at decimeter-level. In our experiment, we
adopt a sliding window size as 3 and the mean computation
Fig. 9. AGV velocity ground-truth at the local n-frame. time is less than 0.5 s per epoch in an intel NUC 11 computer
JOURNAL OF LATEX CLASS FILES, VOL. 18, NO. 9, SEPTEMBER 2020 11

Fig. 10. Calibration results of the field test, where orange dashed line denotes the initial guess; the red line denotes the loosely coupled (LC) method; the blue
line denotes the tightly coupled method without ambiguity resolution (TC-WAR); the green line denotes the tightly coupled method with ambiguity resolution
(TC-AR). Note that LC cannot estimate roll angle and omega scale since it only employs 1D odometer measurement (i.e., the linear velocity)

TABLE IV
IMU HARDWARE CONFIGURATIONS IN REAL TESTS

Bias stability Random walk

IMU Grade
p p
Gyro. (°/h Acc. (mGal) Ang. (°/ (h) Vel. (m/s/ (h))

NovAtel SPAN-CPT Tactical 1 0.75 0.0667 -

Xsens-MTi30 MEMS 18 0.015 1.8 5.88

with i5-1135G7 CPU and 16 GB RAM . Therefore, it is RTKLIB [49].

suitable for real-time applications. (c) TC-WAR: the proposed TC method without ambiguity
2) Evaluation Scenes and Methods: The velocity in n- resolution (WAR).
frame is shown in Fig. 9. There are sufficient acceleration (d) TC-AR: the proposed TC method with ambiguity reso-
and turning to stimulate the extrinsic calibration process. . To lution (AR).
evaluate the robustness of our method under limited satellite
observations, we only employ GPS to aid the IMU-odometer
calibration. The average satellite number is 5.39. To test the B. Estimation of IMU-odometer parameters at the calibration
accuracy and robustness the proposed method, we employ a stage
two-stage test to evaluate the performance of the proposed Fig. 10 shows the results of IMU-odometer extrinsic calibra-
calibration method: (1) the calibration stage: parameters are tion at the calibration stage. Since the LC method only utilize
estimated with measurements; (2) the testing stage: calibrated 1D odometer measurements, the roll angle and bearing rotating
parameters at four typical scenes of the calibration stage rate scale sω are not estimated. Notably, the fluctuation of the
are extracted and IMU-odometer positioning with simulated translation estimation is more obvious for the LC method.
GNSS outage in about 1000 s is used to evaluate the calibration We consider the following reasons. On the one hand, the
results. As is shown in Fig. 11, the test trajectory is a loop RTK positioning for the input of the LC method utilizing
which is convenient for dead-reckoning based positioning GPS requires at least three satellites and could degrade or
evaluation. Based on these experimental settings, the compared fail with poor satellite geometry. On the other hand, the LC
methods are listed below: method can only deal with GNSS outliers at result level.
(a) initial guess: coarse initial IMU-odometer parameters Instead, the proposed TC method can process raw GNSS
manually provided. measurements with less information loss and therefore is more
(b) LC: the loosely coupled GNSS positioning constrained robust to GNSS outliers. Despite the absence of the ground
INS-odometer calibration [24] using RTK positioning from truth of the IMU-odometer parameters, we can still evaluate
JOURNAL OF LATEX CLASS FILES, VOL. 18, NO. 9, SEPTEMBER 2020 12

TABLE V
IMU- ODOMETER DEAD - RECKONING POSITIONING ERRORS USING THE CALIBRATED RESULTS FROM THE CALIBRATION STAGE , INCLUDING MAXIMUM
ABSOLUTE ERROR (MAX) AND ROOT MEAN SQUARES ERROR (RMSE). T HE ’ INITIAL GUESS ’ COLUMN IS THE SAME ACROSS ALL FOUR SCENARIOS .

scene initial guess LC TC-WAR TC-AR

MAX/RMSE
(m)
1 (open) 63.32/30.81 38.67/18.44 19.05/7.74
2 (few trees) 61.51/30.99 32.92/15.22 17.75/7.02
168.65/84.54
3 (more trees) 51.40/31.53 34.27/15.80 25.30/10.81
4 (half-sky) 30.48/14.17 36.66/17.12 27.08/11.77

the calibration results indirectly using IMU-odometer dead-

reckoning test. Specially, the pitch angle of both TC-WAR and
TC-AR are drifting severely. We speculate that the pitch angle
is highly related to the Z-translation during forward moving.
Since the Z-translation error is not limited, the pitch angle of
the IMU-odometer also absorbs that error. Moreover, since the
GNSS-based sensor integrated navigation could be affected by
environments, the calibration results at the four typical scenes

North (m)
mentioned above are evaluated.

C. Performance of IMU-odometer localization with calibrated

results at the test stage
In this section, the calibrated results are evaluated in a
IMU-odometer dead-reckoning localization system. For fair
comparison, we simulate GNSS outage after 100 seconds
for a consistent initialization using GNSS measurements.
Since the vertical translation at v-frame of the IMU-odometer
extrinsic parameters is unidentifiable and may include large
errors. We only compare the horizontal positioning errors in East (m)
the following. It is worth noting that due to the inherent
unobservability of the vertical extrinsic parameter and the
Fig. 11. Horizontal trajectory comparison of IMU-odometer dead-reckoning
coupling with extrinsic pitch and lever arm parameters, the TC test in scene 3 (more trees).
method accumulates larger errors in elevation compared to LC.
Nevertheless in the horizontal plane which is the practically
relevant dimension the TC method consistently outperforms 2 with longer stimulation. Nevertheless, when facing more
LC. For direct illustration, we plot the horizontal trajectory challenging environments at scene 3 and 4, the MAX increases
in n-frame in Fig. 11 using the calibrated at scene 3 (more to 34.27 m and 36.66 m. Finally, the TC-AR provides the most
trees). The trajectories show that our proposed TC-AR method accurate calibration results for this localization test. The trend
achieves the most accurate dead-reckoning performance in of TC-AR follows a similar pattern as TC-WAR. TC-AR also
horizontal east-north directions. achieves the best calibration at scene 2 with MAX of 17.75
Their horizontal positioning errors are presented in Table m and the worst one at scene 4 with MAX of 27.08 m.
V. Notably, MAX presents maximum absolute positioning Therefore, both TC-WAR and TC-AR are more accurate
error and RMSE denotes root mean square positioning error. and robust than the initial guess and LC method, with TC-
Since the LC method does not output roll angle and the AR performing the best under these test conditions. TC-AR
odometer scaling factor of angular velocity, we replace them achieves up to 89.09% and 71.14% improvement in maximum
with the same parameters from the initial guess during testing horizontal positioning errors, respectively, compared to the
stage. The results show that MAX is 168.65 m and RMSE is initial guess and LC method.
84.54 m for the initial guess scheme, which are the maximum
ones among all compared schemes. Compared to that, the LC
scheme gets smaller errors. From scene 1 to scene 4, the MAX D. Discussion: Impact of the GNSS ambiguity resolution on
shrinks from 63.32 m to 30.48 m and the RMSE reduces the calibration performance
from 30.81 m to 14.17 m. For TC-WAR, the calibration is In open-sky environments, only a few outliers in pseudo-
more accurate and robust. The MAX and RMSE of TC-WAR range measurements are caused by multipath and NLOS
are 38.67 m and 18.44 m separately at scene 1. After that, effects. Then the GNSS can provide accurate relative ranging
the errors are reduced to 32.92 m and 15.22 m at scene reference for calibration. In [43], ionosphere-free (IF) GNSS
JOURNAL OF LATEX CLASS FILES, VOL. 18, NO. 9, SEPTEMBER 2020 13

measurements are employed to aid visual-inertial extrinsic [2] J. Liu, W. Gao, and Z. Hu, “Visual-Inertial Odometry Tightly Coupled
calibration. Ambiguity resolution is not used in that work, with Wheel Encoder Adopting Robust Initialization and Online Extrinsic
Calibration,” in 2019 IEEE/RSJ International Conference on Intelligent
likely because the authors utilized geodetic-grade receivers to Robots and Systems (IROS), pp. 5391–5397, Nov. 2019. ISSN: 2153-
refine the GNSS measurements. With such clean pseudo-range 0866.
and carrier-phase data, the ambiguity variable is stable and [3] X. Zuo, M. Zhang, Y. Chen, Y. Liu, G. Huang, and M. Li, “Visual-
Inertial Localization for Skid-Steering Robots with Kinematic Con-
the time-differenced carrier-phase is equal to a high-precision straints,” Nov. 2019. arXiv:1911.05787 [cs].
relative range measurement. Hence the extrinsic parameters [4] S. Zhang, Y. Guo, Q. Zhu, and Z. Liu, “Lidar-IMU and Wheel Odometer
can be estimated accurately, similar in principle to hand- Based Autonomous Vehicle Localization System,” in 2019 Chinese
Control And Decision Conference (CCDC), pp. 4950–4955, June 2019.
eye calibration using accurate relative transform measurements ISSN: 1948-9447.
[50]. [5] Z. Gao, M. Ge, Y. Li, Q. Chen, Q. Zhang, X. Niu, H. Zhang, W. Shen,
In contrast, this paper focuses on low-cost sensors, which and H. Schuh, “Odometer, low-cost inertial sensors, and four-GNSS data
to enhance PPP and attitude determination,” GPS Solutions, vol. 22,
are more susceptible to measurement outliers. Although the p. 57, Apr. 2018.
double-differenced (DD) model eliminates most systematic [6] Q. Chen, Q. Zhang, and X. Niu, “Estimate the Pitch and Heading
errors, gross errors caused by the surrounding environment be- Mounting Angles of the IMU for Land Vehicular GNSS/INS Integrated
come more prominent in relative terms. Furthermore, low-cost System,” IEEE Transactions on Intelligent Transportation Systems,
vol. 22, pp. 6503–6515, Oct. 2021.
GNSS receivers often adopt loose signal quality thresholds [7] X. Li, Z. Qin, Z. Shen, X. Li, Y. Zhou, and B. Song, “A High-
to maintain continuous tracking, allowing low-quality mea- Precision Vehicle Navigation System Based on Tightly Coupled PPP-
surements to pass. To mitigate these environmental outliers, RTK/INS/Odometer Integration,” IEEE Transactions on Intelligent
Transportation Systems, vol. 24, pp. 1855–1866, Feb. 2023. Conference
the OM method applied prior to AR removes a portion of Name: IEEE Transactions on Intelligent Transportation Systems.
the pseudo-range outliers, leading to an improvement up to [8] J. Georgy, A. Noureldin, M. J. Korenberg, and M. M. Bayoumi,
50.74% in the maximum horizontal error when comparing TC- “Modeling the Stochastic Drift of a MEMS-Based Gyroscope in
Gyro/Odometer/GPS Integrated Navigation,” IEEE Transactions on In-
AR to TC-WAR. Nevertheless, cycle-slips or remaining gross telligent Transportation Systems, vol. 11, pp. 856–872, Dec. 2010.
errors may still lead to incorrect ambiguity resolution results, Conference Name: IEEE Transactions on Intelligent Transportation
which could be researched in the future. Systems.
[9] G. Huang, “Visual-Inertial Navigation: A Concise Review,” in 2019 In-
ternational Conference on Robotics and Automation (ICRA), pp. 9572–
VIII. C ONCLUSION 9582, May 2019. ISSN: 2577-087X.
[10] T. Qin and S. Shen, “Robust initialization of monocular visual-inertial
In this paper, we presented an online IMU–odometer param- estimation on aerial robots,” in 2017 IEEE/RSJ International Conference
eter calibration method for AGV localization using raw GNSS on Intelligent Robots and Systems (IROS), pp. 4225–4232, Sept. 2017.
measurements. The proposed method is built on a tightly ISSN: 2153-0866.
[11] A. I. Mourikis and S. I. Roumeliotis, “A Multi-State Constraint Kalman
coupled factor graph optimization framework, which incorpo- Filter for Vision-aided Inertial Navigation,” in Proceedings 2007 IEEE
rates pseudo-range, carrier-phase, and Doppler measurements International Conference on Robotics and Automation, (Rome, Italy),
together with outlier mitigation and ambiguity resolution. We pp. 3565–3572, IEEE, Apr. 2007. ISSN: 1050-4729.
[12] J. Zhang and S. Singh, “LOAM: Lidar Odometry and Mapping in
further conducted observability analysis, showing that two Real-time,” in Robotics: Science and Systems X, Robotics: Science and
horizontal translation and three rotation parameters are iden- Systems Foundation, July 2014.
tifiable, while the vertical translation remains unobservable. [13] H. Zhao, X. Ji, D. Wei, and J. Zhang, “Online IMU-Odometer Extrinsic
Both simulation and field experiments verified that our method Calibration Based on Visual-Inertial-Odometer Fusion for Ground Vehi-
cles,” in 2022 IEEE 12th International Conference on Indoor Positioning
significantly improves calibration accuracy and robustness and Indoor Navigation (IPIN), pp. 1–8, Sept. 2022. ISSN: 2471-917X.
compared with state-of-the-art loosely coupled approaches. [14] H. Zhao, X. Ji, and D. Wei, “Vehicle-Motion-Constraint-Based Visual-
Moreover, results indicate that the method can remain effective Inertial-Odometer Fusion With Online Extrinsic Calibration,” IEEE
Sensors Journal, vol. 23, pp. 27895–27908, Nov. 2023.
even without ambiguity resolution, enhancing its availability [15] H. Zhang, D. Wang, and J. Huo, “Mounting Misalignment and Time
in practice. Offset Self- Calibration Online Optimization Method for Vehicular
Finally, some limitations still exist. The calibration problem Visual-Inertial-Wheel Odometer System,” IEEE Transactions on Instru-
mentation and Measurement, vol. 73, pp. 1–13, 2024.
is inherently non-convex, requiring reliable initialization for [16] X. Zuo, M. Zhang, M. Wang, Y. Chen, G. Huang, Y. Liu, and
convergence. Future work will explore convex relaxation or M. Li, “Visual-Based Kinematics and Pose Estimation for Skid-Steering
certifiable optimization techniques to achieve global guaran- Robots,” IEEE Transactions on Automation Science and Engineering,
vol. 21, pp. 91–105, Jan. 2024.
tees. [17] W. Lee, P. Geneva, C. Chen, and G. Huang, “MINS: Efficient and
Robust Multisensor-aided Inertial Navigation System,” Sept. 2023.
ACKNOWLEDGEMENT arXiv:2309.15390 [cs].
[18] L. Xiong, R. Kang, J. Zhao, P. Zhang, M. Xu, R. Ju, C. Ye, and
The data processing of LC method [24] is supported by T. Feng, “G-VIDO: A Vehicle Dynamics and Intermittent GNSS-
Dr. Shiyu Bai. Besides, the authors would like to thank the Aided Visual-Inertial State Estimator for Autonomous Driving,” IEEE
Transactions on Intelligent Transportation Systems, vol. 23, pp. 11845–
open-source software GICI-LIB [32] from Shanghai Jiao Tong 11861, Aug. 2022. Conference Name: IEEE Transactions on Intelligent
University. We develop our method based on this library. Transportation Systems.
[19] Y. Che and J. Dong, “Visual–Inertial–Wheel Tightly Couple GNSS and
R EFERENCES Barometer With Online Calibration Adopting Differential Model,” IEEE
Sensors Journal, vol. 25, pp. 33437–33446, Sept. 2025.
[1] T. G. R. Reid, S. E. Houts, R. Cammarata, G. Mills, S. Agarwal, A. Vora, [20] J. Sun, W. Sun, J. Zheng, X. Fang, J. Liu, and A. Mian,
and G. Pandey, “Localization Requirements for Autonomous Vehicles,” “UWB–IMU–Odometer Fusion for Simultaneous Calibration and Lo-
SAE International Journal of Connected and Automated Vehicles, vol. 2, calization,” IEEE Internet of Things Journal, vol. 12, pp. 950–963, Jan.
pp. 12–02–03–0012, Sept. 2019. arXiv:1906.01061 [cs, eess]. 2025.
JOURNAL OF LATEX CLASS FILES, VOL. 18, NO. 9, SEPTEMBER 2020 14

[21] Y. Wu, C. Goodall, and N. El-Sheimy, “Self-calibration for Applications,” IEEE Sensors Journal, vol. 22, pp. 4545–4557, Mar.
IMU/Odometer Land Navigation: Simulation and Test Results,” pp. 839– 2022. Conference Name: IEEE Sensors Journal.
849, Jan. 2010. [44] J. Yin, A. Li, W. Xi, W. Yu, and D. Zou, “Ground-Fusion: A Low-
[22] S. Bai, J. Lai, P. Lyu, Y. Cen, and B. Ji, “Improved Preintegration cost Ground SLAM System Robust to Corner Cases,” Feb. 2024.
Method for GNSS/IMU/In-Vehicle Sensors Navigation Using Graph arXiv:2402.14308 [cs].
Optimization,” IEEE Transactions on Vehicular Technology, vol. 70, [45] L.-T. Hsu, F. Huang, H.-F. Ng, G. Zhang, Y. Zhong, X. Bai, and W. Wen,
pp. 11446–11457, Nov. 2021. Conference Name: IEEE Transactions “Hong Kong UrbanNav: An Open-Source Multisensory Dataset for
on Vehicular Technology. Benchmarking Urban Navigation Algorithms,” NAVIGATION: Journal
[23] S. Bai, J. Lai, P. Lyu, Y. Cen, B. Wang, and X. Sun, “Graph-optimisation- of the Institute of Navigation, vol. 70, no. 4, p. navi.602, 2023.
based self-calibration method for IMU/odometer using preintegration [46] Novatel, “Waypoint 8.9 User Manunal,” 2020.
theory,” Journal of Navigation, vol. 75, pp. 594–613, May 2022. [47] Stanford Artificial Intelligence Laboratory et al., “Robotic Operating
[24] S. Bai, J. Lai, P. Lyu, B. Wang, X. Sun, and W. Yu, “An Enhanced System,” May 2018.
Adaptable Factor Graph for Simultaneous Localization and Calibration [48] h. Hu, J. Wu, and Z. Xiong, “A Soft Time Synchronization Framework
in GNSS/IMU/Odometer Integration,” IEEE Transactions on Vehicular for Multi-Sensors in Autonomous Localization and Navigation,” in
Technology, vol. 72, pp. 11346–11357, Sept. 2023. 2018 IEEE/ASME International Conference on Advanced Intelligent
[25] X. Liu, W. Wen, and L.-T. Hsu, “GLIO: Tightly-Coupled Mechatronics (AIM), pp. 694–699, July 2018. ISSN: 2159-6255.
GNSS/LiDAR/IMU Integration for Continuous and Drift-Free State [49] T. Takasu and A. Yasuda, “Development of the low-cost RTK-GPS
Estimation of Intelligent Vehicles in Urban Areas,” IEEE Transactions receiver with an open source program package RTKLIB,” in Interna-
on Intelligent Vehicles, vol. 9, pp. 1412–1422, Jan. 2024. tional Symposium on GPS/GNSS, (International Convention Center Jeju,
[26] S. Cao, X. Lu, and S. Shen, “GVINS: Tightly Coupled GNSS-Visual- Korea), Nov. 2009.
Inertial Fusion for Smooth and Consistent State Estimation,” Aug. 2021. [50] M. Giamou, Z. Ma, V. Peretroukhin, and J. Kelly, “Certifiably Globally
arXiv:2103.07899 [cs]. Optimal Extrinsic Calibration from Per-Sensor Egomotion,” Oct. 2020.
[27] J. Wang, Y. Gao, Z. Li, X. Meng, and C. Hancock, “A Tightly-Coupled arXiv:1809.03554.
GPS/INS/UWB Cooperative Positioning Sensors System Supported by
V2I Communication,” Sensors, vol. 16, p. 944, June 2016.
[28] W. Wen and L.-T. Hsu, “Towards Robust GNSS Positioning and Real-
time Kinematic Using Factor Graph Optimization,” in 2021 IEEE
International Conference on Robotics and Automation (ICRA), (Xi’an,
China), pp. 5884–5890, IEEE, May 2021.
[29] Z. Shen, X. Li, and X. Li, “Advancing High-Precision Navigation:
Leveraging Homogeneous Sensors in Tightly Coupled PPP-RTK/IMU
Integration,” IEEE Transactions on Industrial Electronics, vol. 71,
pp. 15100–15110, Nov. 2024. Conference Name: IEEE Transactions
on Industrial Electronics.
[30] C. Forster, L. Carlone, F. Dellaert, and D. Scaramuzza, “On-Manifold
Preintegration for Real-Time Visual–Inertial Odometry,” IEEE Transac-
tions on Robotics, vol. 33, pp. 1–21, Feb. 2017.
[31] C. De Giorgi, D. De Palma, and G. Parlangeli, “Online Odometry
Calibration for Differential Drive Mobile Robots in Low Traction
Conditions with Slippage,” Robotics, vol. 13, p. 7, Dec. 2023.
[32] C. Chi, X. Zhang, J. Liu, Y. Sun, Z. Zhang, and X. Zhan, “GICI-LIB: A
GNSS/INS/Camera Integrated Navigation Library,” IEEE Robotics and
Automation Letters, vol. 8, pp. 7970–7977, Dec. 2023. arXiv:2306.13268
[cs].
[33] X. Li, J. Huang, X. Li, Z. Shen, J. Han, L. Li, and B. Wang, “Review
of PPP–RTK: achievements, challenges, and opportunities,” Satellite
Navigation, vol. 3, p. 28, Dec. 2022.
[34] B. Song, P. Yan, X. Xia, Y. Zhong, W. Wen, and L.-T. Hsu, “Two
stage GNSS outlier detection for factor graph optimization based GNSS-
RTK/INS/odometer fusion,” Oct. 2025. arXiv:2510.00524 [cs].
[35] Y. He, Y. Guo, A. Ye, and K. Yuan, “Camera-odometer calibration and
fusion using graph based optimization,” in 2017 IEEE International
Conference on Robotics and Biomimetics (ROBIO), pp. 1624–1629, Dec.
2017.
[36] S. Thrun, W. Burgard, and D. Fox, Probabilistic robotics. Intelligent
robotics and autonomous agents, Cambridge, Massachusetts London:
MIT Press, 2006.
[37] T. Ando, “Generalized Schur complements,” Linear Algebra and its
Applications, vol. 27, pp. 173–186, 1979.
[38] P. Teunissen, “The Lambda Method for the GNSS Compass,” Artificial
Satellites, vol. 41, Jan. 2006.
[39] N. Vasilyuk and D. Tokarev, “Identification of geometric displacements
of odometers in a GNSS/inertial navigation system installed on a
land vehicle,” in 2020 IEEE/ION Position, Location and Navigation
Symposium (PLANS), pp. 197–207, Apr. 2020. ISSN: 2153-3598.
[40] N. Koenig and A. Howard, “Design and use paradigms for Gazebo,
an open-source multi-robot simulator,” in 2004 IEEE/RSJ International
Conference on Intelligent Robots and Systems (IROS) (IEEE Cat.
No.04CH37566), vol. 3, pp. 2149–2154 vol.3, Sept. 2004.
[41] X. Li, L. Yuan, X. Li, J. Huang, Z. Shen, and Y. Tan, “GREAT-PVT:
an open-source software for multi-frequency and multi-GNSS PPP-AR
and RTK,” GPS Solutions, vol. 29, p. 157, July 2025.
[42] C. E. Noll, “The crustal dynamics data information system: A resource
to support scientific analysis using space geodesy,” Advances in Space
Research, vol. 45, pp. 1421–1440, June 2010.
[43] Y. Zhou, S. Li, C. Xia, Z. Shen, X. Wang, and X. Li, “Online Visual-
Inertial Extrinsic Calibration Utilizing GNSS Measurements for Vehicle