GPS-IMU Sensor Fusion For Reliable Autonomous Vehicle Position Estimation
GPS-IMU Sensor Fusion For Reliable Autonomous Vehicle Position Estimation
Abstract—Global Positioning System (GPS) navigation pro- fusion aims to leverage the global positioning capabilities
arXiv:2405.08119v1 [eess.SY] 13 May 2024
vides accurate positioning with global coverage, making it a of GPS with the relative motion insights from IMUs, thus
reliable option in open areas with unobstructed sky views. How- enhancing the robustness and accuracy of navigation systems
ever, signal degradation may occur in indoor spaces and urban
canyons. In contrast, Inertial Measurement Units (IMUs) consist in autonomous vehicles. The application of advanced Bayesian
of gyroscopes and accelerometers that offer relative motion filtering techniques, particularly the Extended Kalman Filter
information such as acceleration and rotational changes. Unlike (EKF) [3] and Unscented Kalman Filter (UKF) [4], facili-
GPS, IMUs do not rely on external signals, making them useful in tates the effective integration of these sensors. This approach
GPS-denied environments. Nonetheless, IMUs suffer from drift ensures seamless and reliable navigation, which is vital for
over time due to the accumulation of errors while integrating
acceleration to determine velocity and position. Therefore, fusing the reliable operation of autonomous vehicles, especially in
the GPS and IMU is crucial for enhancing the reliability and environments where GPS signals are compromised.
precision of navigation systems in autonomous vehicles, especially This project, which utilizes the KITTI GNSS and IMU
in environments where GPS signals are compromised. To ensure datasets for validation, demonstrates its potential through
smooth navigation and overcome the limitations of each sensor,
realistic experimental setups. These initiatives contribute to the
the proposed method fuses GPS and IMU data. This sensor
fusion uses the Unscented Kalman Filter (UKF) Bayesian filtering development of the technological infrastructure for self-driving
technique. The proposed navigation system is designed to be vehicles and tackle the crucial safety and efficacy issues facing
robust, delivering continuous and accurate positioning critical for the industry at present.
the safe operation of autonomous vehicles, particularly in GPS-
denied environments. This project uses KITTI GNSS and IMU
datasets for experimental validation, showing that the GNSS- II. L ITERATURE R EVIEW
IMU fusion technique reduces GNSS-only data’s RMSE. The
RMSE decreased from 13.214, 13.284, and 13.363 to 4.271, The GPS and IMU fusion is essential for autonomous
5.275, and 0.224 for the x-axis, y-axis, and z-axis, respectively. vehicle navigation. It addresses limitations when these sensors
The experimental result using UKF shows promising direction operate independently, particularly in environments with weak
in improving autonomous vehicle navigation using GPS and or obstructed GPS signals, such as urban areas or indoor
IMU sensor fusion using the best of two sensors in GPS-denied
environments.
settings. Given the rising demand for robust autonomous nav-
Index Terms—Autonomous Vehicle Position, Global Positioning igation, developing sensor fusion methodologies that ensure
System, Inertial Measurement Unit, Sensor Fusion, Unscented reliable vehicle navigation is essential.
Kalman Filter Various filtering techniques are used to integrate GNSS/GPS
and IMU data effectively, with Kalman Filters [5] and their
I. I NTRODUCTION variants, such as the Extended Kalman Filter (EKF), the Un-
In autonomous vehicle navigation, integrating Global Posi- scented Kalman Filter (UKF), etc. Caron et al. [6] introduced
tioning System (GPS) and Inertial Measurement Units (IMU) a multisensor Kalman filter technique incorporating contextual
has become a cornerstone for achieving reliable and precise variables to improve GPS/IMU fusion reliability, especially in
location tracking, particularly in challenging environments. signal-distorted environments. Lee et al. [7] put forth a sensor
While GPS offers extensive coverage and high positional fusion method that combines camera, GPS, and IMU data,
accuracy in open spaces [1], its performance degrades in utilizing an EKF to improve state estimation in GPS-denied
indoor or urban canyons where signals are obstructed [2]. scenarios. Different innovative sensor fusion methods push
Conversely, IMUs provide valuable motion data independently the boundaries of autonomous vehicle navigation. Suwandi
of external signals, making them indispensable in GPS-denied et al. [8] demonstrated a cost-effective approach to vehicle
areas. However, the utility of IMUs is hindered by their navigation by focusing on low-cost IMU and GPS sensor
susceptibility to drift over time, which accumulates errors in fusion to improve navigation. Atia et al. [9] combined MEMS,
velocity and position estimations derived from acceleration IMU, GPS, and road network maps with an EKF and Hidden
data. Markov model-based map-matching to provide accurate lane
To mitigate the limitations of each sensor type, the fusion determination without high-precision GNSS technologies. Li
of GPS and IMU data emerges as a crucial strategy. This and Xu [10] introduced a method for sensor fusion navigation
in GPS-denied areas. This approach fuses cheaper sensors with
a sliding mode observer and a federated Kalman filter.
Liu et al. [11] developed an enhanced Adaptive Kalman
Filter (IAE-AKF) with an attenuation factor to handle noise
effectively, resulting in a 20% improvement in navigation
accuracy. Meng et al. [12] employed the Global Navigation
Satellite System (GNSS), IMU, DMI, and LiDAR to coun-
teract inaccuracies caused by GNSS signal jumps and multi-
path interference in urban settings. Tao et al. [13] introduced
a multisensor fusion strategy that integrates GNSS, IMU,
and visual data with global pose graph optimization, yielding
superior results on a ROS simulation platform using the Fig. 1: The proposed architecture.
KITTI dataset. Yusef et al. [14] employed a Deep Visual
Inertial Odometry framework to improve accuracy in low-
GNSS areas. At the same time, Park [15] used an adaptive IMU and enhances its output, providing an accurate estimate
Kalman filter for vehicle position estimation to address GPS of the vehicle’s ’Pose’—its precise location and orientation in
outages, validated in real-world tests. Gruyer and Pollard space. This pose estimate is not merely a set of coordinates; it’s
[16] enhanced navigation in GPS-denied environments using a dynamic vehicle’s position representation. It is continuously
proprioceptive sensors with an Interacting Multiple Model updated and adjusted, ensuring the vehicle ’knows’ where it
(IMM) filter. Godoy et al. [17] proposed a Particle Swarm is, where it needs to go, and how to get there safely. This
Optimization-based filter integrating multiple sensors with is vital for the vehicle’s decision-making processes, such as
digital maps, showing promise in reducing reliance on GPS navigating complex routes, avoiding obstacles, and ensuring
in accurate vehicle testing. passenger safety. The proposed work illustrates an equilibrium
Although these studies propose novel methods to improve between robustness and accuracy, aiming to achieve seamless
autonomous vehicle navigation, most rely on locally collected navigation functionality, particularly in demanding settings
data, making their results difficult to reproduce. This work where GNSS signals may not be dependable or accessible.
addresses this issue using a publicly available dataset, allow-
The mathematical representation of the system comprises
ing others to improve autonomous driving navigation with
process and measurement models. The dynamic connection
real, representative, and accessible data. This proposed fusion
between the states at two consecutive time steps is managed
technique leverages the strengths of both GNSS and IMU
by the process truth model, which can be defined as follows:
to maintain continuous operation, even if one sensor fails. It
uses the publicly accessible KITTI dataset for testing, allowing
others to replicate and validate the results. By experimenting xt = f (xt−1 , ut−1 ) + wt−1 , (1)
with GPS-only and fused data, this study demonstrates the
importance of sensor fusion for safe navigation in real-world
conditions. This approach bridges the gap between theoreti- where xt indicates the estimated state after an interval σ,
cal models and practical applications, experimenting with a derived from the preceding state vector xt−1 . The variable
dataset representative of the real driving scenario. ut−1 functions as the input to the state space equations, and
wt−1 reflects the noise affecting the process.
III. M ETHOD Data from the GNSS receiver introduce errors in the ve-
Fig. 1 shows the proposed sensor fusion model for au- hicle’s position, velocity, and orientation estimates. When a
tonomous vehicle navigation. It fuses data from two primary measurement from the GNSS receiver is obtained, the GNSS
sources: an IMU and a GNSS. With its accelerometers and measurement model is defined as:
gyroscopes, the IMU provides real-time information about
the vehicle’s acceleration and rotational movements, offering yp = p + np , (2)
continuous data crucial for navigating without relying on
external signals.
where p ∈ (x, y, z), np represents the measurement noise
On the other hand, GNSS offers absolute positioning and G e e e
and yp = CE [x , y , z ]. Here, (xe , y e , z e ) are Earth-centered
velocity data that are essential for high-precision location
Earth-fixed (ECEF) rectangular coordinates [12], [19].
tracking when the vehicle has a clear path to the sky. The
UKF fuses the data from the two sensors. Unlike simpler e
models, the UKF is uniquely suited for handling the non-linear x (RN + h) cos ϕ cos λ
nature of the data integration challenge posed by autonomous y e = (RN + h) cos ϕ sin λ (3)
ze
vehicle navigation. By processing the high-rate, raw motion RN (1 − e2 ) + h sin ϕ
data from the IMU and the periodic positional corrections from
the GNSS, the UKF corrects for any potential drift from the where
Then, the update process continues:
x̄t|t−1 = f (xt−1 , ut−1 )
RN is the normal radius 2n
X
h is the ellipsoidal height x−
t = Wim x̄i,t|t−1
λ is the longitude i=0 (6)
2n
ϕ is the latitude Pt− =
X
Wic x̄i,t|t−1 − x−
x̄i,t|t−1 − x−
⊤
+ Q,
t t
e is the eccentricity. i=0
measurement covariance Pyt and the state-measurement cross- the inherent higher precision of GPS in the horizontal plane.
covariance matrix Pxt yt are also derived from this process. Kt Overall, these errors are essential in assessing the effectiveness
denotes the Kalman gain, vt is the innovation term, and xt and of the fusion system.
Pt reflect the updated state and covariance at time t. Fig. 4 demonstrating the position estimation in the xy-plane
Overall, the UKF is a more robust approach in situations illustrates that the UKF estimated track closely aligns with the
where the Kalman Filter’s linear assumptions are insufficient, GPS measurements. The degree of tightness of the UKF track
especially in autonomous navigation and robotics, where ac- around the GPS measurements indicates the accuracy of the
curacy in the face of non-linear behaviors is crucial. filter’s performance. Similarly, the RMSE across the X, Y, and
Z axes are calculated to show the fusion model’s robustness.
IV. E XPERIMENTS
The RMSE using GNSS-only shows 13.214, 13.284, and
The experiment uses the KITTI GNSS and IMU dataset 13.363 for the X, Y, and Z axes, respectively. The fusion of
[18]. The GNSS and IMU data are essential for tasks that the GNSS and IMU reduces the RMSE to 4.271, 5.275, and
involve vehicle pose estimation and tracking in autonomous 0.224 for the x, y, and z axes, respectively, as shown in Table
driving. The GNSS provides geographic location information, I. For real-time implementation, certain challenges, such as
while the IMU provides insights into the vehicle’s movements sensor calibration and synchronization, need to be considered
through space, such as acceleration and orientation. Together, when incorporating sensor fusion technology into autonomous
these data sets enable high-precision tracking and navigation, vehicles. Developing strong software algorithms, ensuring
which is crucial in developing and testing autonomous sys- adequate hardware infrastructure, and conducting extensive
tems. The dataset provides a dynamic and challenging real- testing in real-world scenarios are essential to addressing
world environment, with traffic scenarios recorded to reflect these challenges. This is crucial to guaranteeing autonomous
various driving conditions. It has served as a benchmark vehicles’ accuracy, reliability, and safety.
for many studies focusing on improving the accuracy and
robustness of autonomous driving technologies. TABLE I: Comparison of RMSE for GNSS-only and GNSS-
The experiment is conducted using a GNSS observation IMU fusion.
frequency of 1 Hz. Similarly, the IMU uses a 0.01 rad/s gyro- RMSE
scope, a 0.05 m/s2 accelerometer, a 0.000001 rad/s2 gyroscope Method
X-axis Y-axis Z-axis
bias, and a 0.0001 m/s3 accelerometer bias. The position and
GNSS 13.214 13.284 13.363
velocity estimation using UKF results are demonstrated in the GNSS-IMU 4.271 5.275 0.224
experiments. The UKF is a variant of the Kalman Filter that is
more suitable for dealing with non-linear systems. It captures
the mean and covariance of the probability distribution of the V. C ONCLUSION
system state by passing carefully chosen sample points through The fusion model leveraging the UKF employs the KITTI
the non-linearities. The errors observed in the x, y, and z dataset for validation, combining real-time IMU data with
coordinates, as depicted in Figures 2 and 3, serve as indicators GNSS absolute positioning to refine autonomous vehicle nav-
of tracking accuracy over time. The GNSS-only results show igation. The IMU’s real-time motion data and the GNSS’s
the errors are larger than the fusion of GNSS and IMU data absolute positioning capabilities are integrated to estimate
for navigation, as shown in Fig. 2. The fusion figure, Fig. 3, precise and reliable poses. This is crucial to helping individuals
for the x and y directions reveals the GNSS frequency that remain aware of their surroundings and perform important
provides the most precise location estimates, characterized by navigation-related tasks, especially when faced with obstacles
a lower mean error and minimal fluctuations around zero. On that might interfere with GNSS signals. This work highlights
the other hand, the z-coordinate aims to identify the frequency the UKF’s superiority in processing nonlinear sensor outputs
that yields the smallest and most uniform error, highlighting a and reinforces the fusion model’s potential for enhancing
better tracking performance in three-dimensional space despite the reliability and efficiency of autonomous vehicle systems.
Fig. 3: Position errors in x, y, and z-coordinates using fusion of GNSS and IMU.
Sensor fusion typically combines data from multiple sensors improve performance.
to improve navigation. It is also crucial during practical
applications for maintaining position and orientation estimates R EFERENCES
if one of the sensors fails, allowing the navigation process [1] D. M. Bevly, ”Global positioning system (GPS): A low-cost velocity
to continue without interruption. Adding LiDAR or radar sensor for correcting inertial sensor errors on ground vehicles,” J. Dyn.
sensors to autonomous vehicle navigation systems can improve Sys., Meas., Control, vol. 126, pp. 255–264, 2004.
[2] X. Li, W. Chen, and C. Chan, ”A reliable multisensor fusion strategy
the precision and dependability of position and orientation for land vehicle positioning using low-cost sensors,” Proceedings of
estimates by enhancing GNSS and IMU fusion. LiDAR’s high- the Institution of Mechanical Engineers, Part D: Journal of Automobile
resolution 3D mapping capabilities enable accurate obstacle Engineering, vol. 228, pp. 1375–1397.
[3] H. W. Sorenson, ”Kalman filtering techniques,” Advances in control
detection and localization, while radar’s long-range detection systems, vol. 3, pp. 219–292, 1966.
and weather-resilience qualities enhance overall robustness. [4] E. A. Wan and R. Van Der Merwe, ”The unscented Kalman filter for
Fusing these sensors with GNSS and IMU can enhance the nonlinear estimation,” Proceedings of the IEEE 2000 adaptive systems
for signal processing, communications, and control symposium (Cat. No.
navigation system’s reliability, ensuring continuous operation 00EX373), pp. 153–158, 2000.
even during sensor failure. This leads to improved safety [5] R. E. Kalman and others, ”Contributions to the theory of optimal
and navigation accuracy across various environments and control,” Bol. soc. mat. mexicana, vol. 5, pp. 102–119, 1960.
[6] F. Caron, E. Duflos, D. Pomorski, and P. Vanheeghe, ”GPS/IMU data
conditions. Additionally, carefully developing existing filtering fusion using multisensor Kalman filtering: introduction of contextual
techniques, such as EKF, or customizing existing filters may aspects,” Information fusion, vol. 7, pp. 221–230, 2006.
[7] Y. Lee, J. Yoon, H. Yang, C. Kim, and D. Lee, ”Camera-GPS-IMU sen-
sor fusion for autonomous flying,” 2016 Eighth International Conference
on Ubiquitous and Future Networks (ICUFN), pp. 85–88, 2016.
[8] B. Suwandi,T. Kitasuka, and M. Aritsugi, ”Low-cost IMU and GPS
fusion strategy for apron vehicle positioning,” TENCON 2017-2017
IEEE Region 10 Conference, pp. 449–454, 2017.
[9] M. M. Atia, A. R. Hilal, C. Stellings, E. Hartwell, J. Toonstra, W. B.
Miners, and O. A. Basir,”A low-cost lane-determination system using
GNSS/IMU fusion and HMM-based multistage map matching,” IEEE
Transactions on Intelligent Transportation Systems, vol. 18, pp. 3027–
3037, 2017.
[10] X. Li and Q. Xu, ”A reliable fusion positioning strategy for land
vehicles in GPS-denied environments based on low-cost sensors,” IEEE
Transactions on Industrial Electronics, vol. 64, pp. 3205–3215, 2016.
[11] Y. Liu, X. Fan, C. Lv, J. Wu, L. Li, and D. Ding, ”An innovative
information fusion method with adaptive Kalman filter for integrated
INS/GPS navigation of autonomous vehicles,” Mechanical systems and
signal processing, vol. 100, pp. 605–616, 2018.
[12] X. Meng, H. Wang, and B. Liu, ”A robust vehicle localization approach
based on gnss/imu/dmi/lidar sensor fusion for autonomous vehicles,”
Sensors, vol. 17, pp. 2140, 2017.
[13] X. Tao, B. Zhu, S. Xuan, J. Zhao, H. Jiang, J. Du, and W. Deng,
”A multi-sensor fusion positioning strategy for intelligent vehicles
using global pose graph optimization,” IEEE transactions on vehicular
technology, vol. 71, pp. 2614–2627, 2021.
[14] A. Yusefi, A. Durdu, F. Bozkaya, S. Tığlıoğlu, A. Yılmaz, and C. Sungur,
”A Generalizable D-VIO and Its Fusion with GNSS/IMU for Improved
Autonomous Vehicle Localization,” IEEE Transactions on Intelligent
Vehicles, 2023.
[15] G. Park, ”Optimal vehicle position estimation using adaptive unscented
Kalman filter based on sensor fusion,”Mechatronics, vol. 99, pp. 103144,
2024.
[16] D. Gruyer and E. Pollard, ”Credibilistic IMM likelihood updating
applied to outdoor vehicle robust ego-localization,” 14th International
Conference on Information Fusion, pp. 1–8, 2011.
[17] J. Godoy, D. Gruyer, A. Lambert, and J. Villagra, ”Development of an
particle swarm algorithm for vehicle localization,” 2012 IEEE Intelligent
Vehicles Symposium, pp. 1114–1119, 2012.
[18] A. Geiger, P. Lenz, C. Stiller, and R. Urtasun, ”Vision meets robotics:
The kitti dataset,” The International Journal of Robotics Research,
vol. 32, pp. 1231–1237, 2013.
[19] A. Noureldin, T. B. Karamat, and J. Georgy, ”Fundamentals of inertial
navigation, satellite-based positioning and their integration,” Springer
Science & Business Media 2012.