Professional Documents
Culture Documents
Project
Project
Peter Blumenstein
Department of Mechanical Engineering
Carnegie Mellon University
pblumens@andrew.cmu.edu
Abstract—VINS-Fusion is a proven state estimation algorithm it can estimate its relative state to objects in its environment.
used in Simultaneous Localization and Mapping (SLAM). In this The state of the vehicle can change based on the use case of
paper, we attempt to improve the state estimation accuracy of the UV, but in the general sense usually includes the position,
VINS-Fusion by coupling it with a Kalman Filter. Specifically,
we compare the results of VINS-Fusion by itself to the outputs of orientation, and speed of the UV.
VINS-Fusion plus an Extended Kalman Filter and an Unscented Two proven algorithms for state estimation are VINS-
Kalman Filter. The respective algorithms will be tested with the Fusion and the Kalman Filter (KF). VINS-Fusion, a form of a
EuROC MAV dataset. Our results will show that the accuracy Visual Inerital System (VINS), has already proven to be very
does not significantly change when VINS-Fusion is combined with effective [1]. However, we wanted to see if we can improve
either variant of the Kalman Filter. Our assumption of constant
covariance matrices for both Kalman Filters is hypothesized to its effectiveness by coupling it with another proven algorithm.
have strongly impacted the performances of our algorithms. As a result, in this project, we attempt to improve the
Index Terms—VINS-Fusion, SLAM, state estimation, Kalman state estimation of VINS-Fusion by modifying it with the KF.
Filter Since, there are multiple forms of the KF, the two specific KF
variations we have chosen are the Extended Kalman Filter
I. I NTRODUCTION (EKF) and the Unscented Kalman Filter (UKF). We have
Over the last few years, we have seen a rise in the demand chosen these two variations for the following reasons:
for autonomous vehicles (AVs). This rise in demand is largely 1 EKF and UKF provide different approaches to nonlinear-
due to the fact that the social, economical, and environmental ity. We wanted to compare these two approaches.
benefits of AVs are becoming more realized. AVs limit the 2 Simple implementation
need for a human driver, alleviating humans from a taxing 3 Low computational cost
obligation. AVs can also be turned into Robotaxis when not
To summarize, our project will be a comparison of the
in use, providing economical gain for the owner. This can also
following:
help limit the number of vehicles on the road, decreasing the
carbon emissions caused by vehicle use. 1 VINS-Fusion
The benefits of AVs don’t stop there. An AV is essentially 2 VINS-Fusion + EKF
an unmanned vehicle (UV), so its application extend beyond 3 VINS-Fusion + UKF
the road. A good example would be autonomous robots that
II. R ELATED W ORKS
operate in closed settings such as robot vacuums used to
clean homes or factory robots used to transport goods around There has already been a lot of proven work on VINS,
manufacturing facilities. Another good example would be showing its advantages. VINS uses both visual data and
drones. Drones can be utilized in construction environments inertial measurement units (IMUs) as its inputs [1]. With the
for safe inspection of hard-to-reach areas. They can also be use of these two kinds of inputs, it allows VINS to estimate
used during search-and-rescue operations. roll and pitch angles, in addition to the common positions and
However, the effectiveness of UVs is evidently based on the velocities. With the incorporation of IMUs, VINS can mitigate
methods behind it that help the UV navigate its environment situations in which the visual sensors don’t perform as well.
safely and effectively. However, despite its many advantages, VINS has drawbacks
One of said methods is Simultaneous Localization and too. VINS is a nonlinear algorithm. It also doesn’t have direct
Mapping (SLAM), which has proven to be an integral part distance measurements. Both of these characteristics make
in the effectiveness of UVs. The essential goal of a SLAM initializing VINS extremely difficult. Furthermore, VINS is
algorithm is to create a map of the vehicle’s environment, so prone to the drift issue. A lot of additional sub-algorithms
need to be run to handle the drift, increasing the algorithm’s OpenVINS was tested in simulation and in the real world
complexity. against VINS-Fusion, an extension of VINS-Mono, and other
There have been extensions of VINS to improve it. One such VINS algorithms. Results showcased OpenVINS not only able
extension is called VINS-Mono [1]. VINS-Mono has a novel to reduce drift but also outperformed the other algorithms in
initialization procedure to deal with the initialization problem. state estimation. This is seen in OpenVINS’ relatively lower
It also incorporates a online relocalization methodology to absolute pose error (see Fig. 3).
handle the drift issue. Another key advantage of VINS-Mono
is that it can merge previous pose graphs with the current graph
for more accurate state estimation. An overview of VINS-
Mono can be seen in Fig. 1.
VI. C ONCLUSION
Fig. 12. VINS Trajectory Error Result.
In this paper we set out to determine whether or not the use
The graphical representations clarify the discernible chal- of a Kalman Filter could improve the state estimation accuracy
lenges encountered by VINS-Fusion, akin to several visual of VINS-Fusion. Our results demonstrated that VINS-Fusion
SLAM algorithms, particularly evident during rapid turning combined with an Extended Kalman Filter produced similar
motions denoted by the darker red trajectories. This phe- results to VINS-Fusion alone, but did not significantly improve
nomenon contrasts with the smoother and straighter portions accuracy. The use of a UKF actually decreased the state
of the trajectories, characterized by a dark blue hue. With estimation error.
Though our results demonstrate VINS-Fusion to be most
accurate on its own, improvements can be made to better refine
our findings.
For both KFs, a constant covariance matrix was assumed.
We hypothesize this influenced the performance of both VINS-
KF variants. With tuned covariance matrices, the VINS-KF
variants should perform better in comparison to our initial
results.
However, even with tuned covariance matrices, it is not
expected that the VINS-KF variants will perform significantly
better than VINS-Fusion alone. VINS-Fusion has already
proven to a state-of-the-art state estimation algorithm. Further
modifications to it are unlikely to improve its existing perfor-
mance. As a result, even with constant covariance matrices, we
still see our results as being a conclusive comparison between
VINS-Fusion, VINS-Fusion + EKF, and VINS-Fusion + UKF.
R EFERENCES
[1] T. Qin, P. Li, and S. Shen, “Vins-mono: A robust and versatile monocular
visual-inertial state estimator,” IEEE Transactions on Robotics, vol. 34,
no. 4, pp. 1004–1020, 2018.
[2] P. Geneva, K. Eckenhoff, W. Lee, Y. Yang, and G. Huang, “Openvins: A
research platform for visual-inertial estimation,” in 2020 IEEE Interna-
tional Conference on Robotics and Automation (ICRA), 2020, pp. 4666–
4672.
[3] T. Qin, S. Cao, J. Pan, and S. Shen, “A general optimization-based
framework for global pose estimation with multiple sensors,” 2019.
[4] R. E. Michael Kaess, “Extended kalman filter - 16-833 robot localization
and mapping lecture slides,” 2023.
[5] H. S. Chadha, “The unscented kalman filter: Anything
ekf can do i can do it better!” Nov 2019. [Online].
Available: https://towardsdatascience.com/the-unscented-kalman-filter-
anything-ekf-can-do-i-can-do-it-better-ce7c773cf88d