Vision-Based Autonomous Landing for the MPC

Controlled Fixed Wing UAV

Sevinç Günsel

, Şeref Naci Engin

and Mustafa Doğan

Department of Control and Automation Engineering, Yildiz Technical University, Istanbul, Turkey

Department of Control and Automation Engineering, Istanbul Technical University, Istanbul, Turkey

Keywords: Image Segmentation, Kalman Filters, vSLAM, MPC, UAV.

Abstract: This work introduces a novel vision-based autonomous landing system for fixed-wing UAVs optimized for

GPS-denied environments. We combine vSLAM with the linear MPC strategy. A key innovation is to use an

SVD-based Kalman filter in vSLAM, which significantly improves map point update accuracy and efficiency

by reducing noise. The system precisely defines the landing area using image segmentation and Watershed

Transform for real-time vSLAM data, then draws a rotated bounding box. This visual data feeds the linearized

MPC, which computes the optimal control inputs which are longitudinal acceleration, yaw rate, vertical

velocity to guide the UAV along the landing trajectory. Simulation results confirm the robust and effective

performance of our integrated vSLAM-MPC architecture in precisely guiding the UAV to the landing zone.

1 INTRODUCTION

In the last decades, the popularity of unmanned aerial

vehicles is growing. UAVs have been using in many

different fields such as mapping or monitoring areas,

searching and rescuing of people, farming and

military applications (Patruno, 2018). These vehicles

can be in different size, configuration and

characteristics. The most common types are known as

fixed-wing, quadrotor and helicopter. A runway may

be required for take-off and landing for some types of

fixed-wing UAV (Gautam, 2014). Detecting a

runway or a ground target for landing is quietly

challenging part of an autonomous system (Rabah,

2018).

Computer vision techniques are applied to detect

and recognize a runway also positioning the vehicle.

Positioning system depends on GPS sensor. However,

GPS signals may be defective or denied for

environment (Vidal, 2017;Garcia, 2017). Therefore,

computer vision techniques play main role to obtain

environmental and vehicle position informations

(Campoy, 2020). There have been many research on

vision based autonomous landing UAVs. In (Kong,

2014), research and developments on visual based

https://orcid.org/0000-0001-5970-3235

https://orcid.org/0000-0003-2514-9250

https://orcid.org/0000-0001-5215-8887

landing system for both rotor and fixed-wing UAVs

have been examined. Image process with low-

resolution cameras, providing stability of attitude

control, calculation accurate descent rate and ensure

constant orientation and alignment of along runway

axis are challenging points of entire landing process.

Quadrotors are mostly preferred aerial vehicles in

terms of flexible movement capability. As in research

(Liu et al., 2021), at any GPS denied environment, the

quadrotor has been landed on an ArUco pattern that

detected by segmentation and threshold techniques.

The position of the quadrotor has been estimated

using EKF and PID used for movement control.

ArUco markers are more advantageous for vision

algorithms than helipad (Bahera, 2020). Another

study proposed a vision system which estimates

altitude, lateral position and forward speed of the

UAV. Also, visual information has been used to

construct a hierarchical control system (Rondon et al.,

2010). Nevertheless, in case of (Urbanski, 2018)

position of a UAV has been specified using the Haar-

Like classifier so that the classifier become a source

of the vehicle position data for controller. ID and PD

controllers have been proposed (Urbanski, 2018).

Günsel, S., Engin, ¸S. N. and Do

gan, M.

Vision-Based Autonomous Landing for the MPC Controlled Fixed Wing UAV.

DOI: 10.5220/0013710600003982

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 22nd International Conference on Informatics in Control, Automation and Robotics (ICINCO 2025) - Volume 1, pages 227-234

ISBN: 978-989-758-770-2; ISSN: 2184-2809

227

GPS-based systems appear to be simpler and more

reliable, but their accuracy is limited. For this reason,

vision-based landing techniques are more precise

(Gautam, 2014). Research (Campoy, 2020) reveals

that there are basically two main types of algorithms

which are feature tracking and appearance-based

tracking. The filtered outputs of image processing

algorithms are used to control the position and

orientation of the UAV. Visual SLAM algorithms are

critical for three-dimensional mapping and

positioning, especially in uncertain outdoor

environments, i.e. without position information.

However, the performance of these algorithms can

vary depending on environmental factors such as

varying light, vibration and speed.

Visual SLAM algorithms are used in mobile

robots mostly (Riccardo, 2016). Also, these

algorithms can be used in a wide range of applications

like underwater or on air. Currently there is no single

approach that can be applied to use in any case

(Kazerouni et al., 2022). If we look studies (Zhang,

2018; Andert, 2022; Kalay, 2009; Lemaire, 2007),

landmark positions are mostly estimated with EKF.

In this study, unlike other studies, a visual SLAM

algorithm for a fixed-wing UAV is presented in

which map points are updated with an SVD-based

Kalman filter and the runway to be landed is

determined with the obtained map points. In addition,

the linear MPC controller is designed to track the

desired trajectory on the runway.

2 VISUAL SIMULTANEOUS

LOCALIZATION AND

MAPPING

Visual-based navigation is getting significant interest

due to its strong noise resistance, high accuracy, and

economic advantages. In the context of autonomous

landing for UAVs systems equipped with cameras.

Thus, the target and environmental data can be

obtained in real time via on board computers. It offers

position and orientation information for decision-

making and control mechanisms. This enables UAVs

to land autonomously in both fixed and moving

environments, even an area is complex or unknown.

Therefore, visual-based autonomous landing has been

a significant research topic and used in both military

and civilian applications (Xin et al., 2022). Some of

the methods used for image-based landing include

image segmentation, image moments, monocular

vision, and stereo vision (Gautam, 2014). Among

these methods, one of the most commonly used

algorithms for monocular vision is Visual

Simultaneous Localization and Mapping (VSLAM).

In this study, the map points obtained with the

vSLAM algorithm were improved using a singular

value decomposition-based Kalman filter.

vSLAM refers to algorithms that enable robots or

aerial vehicles to simultaneously determine their own

position and map their surroundings while moving

through unknown environments. These algorithms

operate by using images obtained from the vehicle's

image sensors. vSLAM primarily consists of two

main steps: localization and mapping. In the

localization phase, the positions of objects in the

environment are determined, and the vehicle's own

position is calculated by tracking its movement. In the

mapping phase, the vehicle creates a map of its

surroundings by recording the paths it has traversed

and the objects it has observed. Visual odometry

algorithms, which form the basis of the vSLAM

algorithm, estimate the moving vehicle's position

using video frames acquired from a camera

(Amasyali et al., 2010). These algorithms are capable

of not only determining the vehicle's position but also

creating a detailed map of the surroundings.

Especially for aerial vehicles flying at high altitudes

and high speeds, combining SLAM algorithms with

visual odometry allows for more accurate and reliable

results. This combination enables the aerial vehicle to

determine its own position more precisely and

navigate more effectively in complex environments

(Boij,2022). In SLAM algorithms, accurately

determining the position of vehicles and their

surroundings critically depends on calculating

unknown parameters. The two most fundamental

filters for these calculations are the Kalman and

Bayes filters. The Kalman filter is well-suited for use

in linear systems. However, since most real-world

systems are nonlinear, the Extended Kalman Filter

(EKF) is preferred in such cases. The EKF enables the

application of the Kalman filter to nonlinear systems

by linearizing them using a Taylor series expansion.

This allows for reliable vehicle positioning and

mapping operations even in complex and nonlinear

environments (Kazerouni et al., 2022).

2.1 vSLAM Algorithm Overview

We've examined the "Monocular Visual

Simultaneous Localization and Mapping" example

from MATLAB's built-in sample projects. This

particular example was chosen because it's easily

modifiable, offering a flexible code base for

experimentation. The image dataset used in this study

ICINCO 2025 - 22nd International Conference on Informatics in Control, Automation and Robotics

228

was sourced from the California Naval Postgraduate

School's website (Purdue University, 2021).

2.1.1 Map Initialization

The process begins with map initialization, where the

VSLAM system establishes its initial understanding

of the 3D environment and the camera's starting pose.

Since a monocular vision cannot inherently determine

depth, initialization typically requires movement

between at least two frames. The system analyses the

initial images, often using the vehicle's starting

position as a coordinate reference. Feature points

(ORB features), are extracted from the first two

frames and then matched. Geometric relationships are

computed to estimate the relative camera motion.

This relative pose is then used to triangulate the first

set of 3D map points, forming the foundational point

cloud of the environment. A crucial step often

following this is an initial bundle adjustment, which

refines the camera poses and 3D map points by

minimizing re-projection errors, thereby improving

the overall accuracy and consistency of the initial

map (MathWorks, n.d).

2.1.2 Key Frames and Map Points

Key frames are a strategic subset of camera images,

chosen to efficiently represent the camera's path and

environment without processing every single frame,

which would be too computationally demanding.

They are selected when the camera moves

significantly, observes new areas, or if tracking

quality degrades. Each key frame stores its estimated

pose and the features observed within it. Map points

are the 3D points that represent the environment.

They are typically created by combining observations

from multiple key frames. Each map point holds its

3D coordinates and information about which key

frames observe it, along with the corresponding 2D

features in those frames. This structured

representation of key frames and map points forms

the backbone of the VSLAM map, enabling efficient

data storage, optimization, and re-localization.

(MathWorks, n.d).

2.1.3 Place Recognition

Place recognition, also known as loop detection or

loop closure detection, is the process by which a

vSLAM system recognizes that it has returned to a

previously visited location. This is critical for

preventing drift the accumulation of small errors in

pose estimation that can cause the map to become

inconsistent over time. Standard methods often

involve building a database of visual features from

previously visited key frames. When a new key frame

is added, its features are queried against this database.

If a strong match is found with a historical key frame,

it signifies a potential loop closure. This detection

then triggers a global optimization process to correct

the accumulated drift across the entire trajectory and

map (matlab).

2.1.4 Tracking

Tracking is the process of estimating the camera's

current position and orientation as it moves through

an environment. For every new image, the system

tries to match its detected features with existing 3D

map points, often by projecting these 3D points onto

the 2D image to find correspondences. A pose

estimation algorithm then calculates the camera's 3D

pose based on these matches. This operation occurs

frequently, for every incoming frame, and must be

highly efficient to maintain real-time performance. If

tracking fails, the system may attempt to re-localize

itself or cease operation. (MathWorks, n.d).

Our design includes a tracking loop with some

distinct additions from the original algorithm. Our

algorithm reads images and extracts ORB features

within this loop. A significant difference from typical

vSLAM examples is explicit use of an SVD-based

Kalman filter for point update. In standard vSLAM,

state estimation and refinement are primarily handled

by bundle adjustment. The SVD-based Kalman filter

performs point updates, suggests a more continuous,

real-time state estimation approach for map points.

The claim that dimensionality reduction has been

performed thanks to the SVD-based Kalman filter,

thus calculation errors are reduced, numerical

stability and efficiency are improved. This method

provides smoother point trajectories and potentially

more robust tracking in noisy environments. The key

frame control is also part of this tracking loop,

determining when a new key frame should be added

based on specific criteria.

The integration of the SVD-based Kalman filter is

the most unique aspect, suggesting a hybrid approach

combining feature-based geometric methods with

probabilistic filtering for state estimation.

2.1.5 Local Mapping

Local mapping is the process of building and refining

the 3D map in the area around the camera's current

position. This operation is more computationally

demanding than tracking and is performed less

frequently.

Vision-Based Autonomous Landing for the MPC Controlled Fixed Wing UAV

229

When a new key frame is added, local mapping

triangulates new 3D map points from features seen in

the new and nearby key frames, merges redundant

map points to keep the map compact, and conducts a

local bundle adjustment. This local optimization

specifically refines a subset of the map, correcting

errors from tracking and improving the local map's

accuracy without re-optimizing the entire map, thus

keeping the computation manageable (MathWorks,

n.d).

2.1.6 Loop Closure

Loop closure is the final and often most complex

stage of a VSLAM system. Its main purpose is to

correct drift by detecting when the camera revisits a

previously visited location and then performing a

global map correction. After a loop is found, the

system verifies its robustness and determines the

precise relative transformation between the current

and past poses. This leads to a global optimization,

often through pose graph optimization or bundle

adjustment. The resulting loop constraint powerfully

distributes accumulated errors across the entire

estimated trajectory and map, creating a globally

consistent and drift-free map. This computationally

intensive process is performed infrequently.

(MathWorks, n.d).

Our design has a loop closure check section. It

performs a loop closure check after a certain number

of key frames have been created. If a loop closure

candidate is found, it adds loop connections and

performs the loop closure. This indicates that the

algorithm integrates the critical components of loop

closure. The success of this stage relies heavily on the

robustness of the place recognition and the underlying

optimization method used to correct the map once a

loop is identified.

2.2 SVD-Based Kalman Filters

The Kalman filter is a method for estimating the state

variables of a linear stochastic dynamic system that

minimizes the covariance of the prediction error.

When calculating the instantaneous estimate of a state

variable, the predicted value from the previous state

and the measured value are used. Subsequently, the

error value in the new estimate is calculated (Unal,

2021). Singular Value Decomposition (SVD) is a

mathematical method used to factorize a matrix into

three matrices. Mathematically, it can be expressed

as:

A=UΛV



, Λ = 

 (1)

A is an m×n matrix, U is an m×m orthogonal matrix,

and V is an n×n orthogonal matrix. Σ is a matrix

containing the eigenvalues of A. According to

Singular Value Decomposition, the matrix A can also

be expressed as:

A=USU



=UD





(2)

The matrix D is a diagonal matrix. When adapted to

the Kalman filter formula, the resulting P matrix is

shown in the equation below.

𝑃(𝑘) = 𝑈(𝑘)𝐷(𝑘)



𝑈(𝑘)



(3)

The SVD-based Kalman Filter is not affected by such

errors, offering a robust method for numerical

computations. Particularly when dealing with ill-

conditioned matrices, SVD-based approaches

produce more reliable results (Hang et al., 2018). This

is highly important for computing covariance

matrices in Kalman Filter updates and helps reduce

noise originating from both the model and

measurements. SVD-based algorithms are powerful

to discriminate the signal and noise subspaces,

compared to EKF and provide better convergence

(Wang, 1992). Given these advantages, an SVD-

based Kalman filter was preferred in this study to

achieve faster and more accurate results to obtain

more accurate map points in vSLAM.

3 IMAGE PROCESSING AND

SEGMENTATION

Image segmentation is a fundamental and complex

area of digital image processing, using computer

algorithms to divide a digital image into distinct and

meaningful regions. Its primary goal is to simplify the

image's representation for easier analysis by grouping

pixels with similar characteristics. This technique has

wide-ranging applications, including content-based

image retrieval, medical imaging, object detection,

traffic control systems, and video surveillance. Image

segmentation methods are broadly categorized as

either local, focusing on isolating specific regions, or

global, which processes the entire image. These

approaches can also be classified based on the

inherent properties of the images themselves.

(Kaur,

2014). Many methods have been developed to

effectively segment images. Among these,

thresholding-based, edge-based, region-based,

gradient-based, and classification-based approaches

are widely recognized as the most common (Sun,

2017).

ICINCO 2025 - 22nd International Conference on Informatics in Control, Automation and Robotics

230

In this study, Watershed Transform is used for

image segmentation. The process starts by preparing

the image: converting it to grayscale and reducing

noise. Then, it highlights edges by calculating the

image's gradient. Crucially, foreground markers are

generated from both bright areas and location points

from vSLAM. These markers guide the Watershed

Transform to divide the image into distinct regions.

Finally, the algorithm finds the center of each

segment and draws a rotated bounding box around it.

4 LINEAR MPC

IMPLEMENTATION FOR

FIXED WING UAV LANDING

Model Predictive Control (MPC) operates by

calculating an optimized series of control actions at

each sampling interval. This process relies on a

predictive model to forecast the system's future

behaviour. However, these predictions aren't always

perfect due to the real-world imperfections like model

inaccuracies and external disturbances. As opposed to

this, MPC employs a closed-loop approach only the

initial control signals from the calculated sequence

are applied, and then the optimization problem is re-

solved at the each time step to generate a new optimal

input sequence. Consequently, MPC necessitates

solving an optimization problem in every control

cycle (Gavilan et al., 2015).

4.1 Mathematical Model of the UAV

In this work, discrete-time model is proposed for the

fixed-wing UAV. The state and control input vectors

are given as following:

𝑥=[𝑃



𝑃



𝑃



𝜓 𝑉]



(4)

𝑢=[𝛼



𝜔 𝑍



]



(5)

where 𝑃



,𝑃



,𝑃



are presented as coordinates on the

world, 𝜓 is the yaw angle and 𝑉 is the velocity. In

control signal our inputs are longitudinal acceleration

𝛼



, yaw rate 𝜔 which is the angular velocity around

the z-axis and longitudinal velocity 𝑍



. Therefore, the

kinematic equations are given in the form of discrete-

time state-space model as follow:

𝑥

(

𝑘+1

)

𝐴

(𝑘)𝑥

(

𝑘

)

+ 𝐵(𝑘)𝑢(𝑘)

(6)

The system and input matrices can be expressed as

given below:

𝐴=

⎣

⎢

⎡

−𝑉𝑠𝑖𝑛(𝜓)𝑇



𝑉𝑐𝑜𝑠(𝜓)𝑇



𝑐𝑜𝑠(𝜓)𝑇



𝑠𝑖𝑛(𝜓)𝑇



⎦

⎥

⎤

(7)

𝐵=

⎣

⎢

⎡

𝑐𝑜𝑠(𝜓)𝑇



𝑠𝑖𝑛(𝜓)𝑇



𝑇



𝑇



𝑇



⎦

⎥

⎤

(8)

A and 𝐵 matrices are time-varying and linearized

around current state 𝑥(𝑘) and updated for each

iteration step. The sampling time is denoted as T



and

is 0.05 seconds. Our motivation for this study inspired

by (Gavilan & Vazquez & Estaban, 2015) and

(Gavilan et al., 2015). The main differences are

kinematic model and control inputs. In both studies,

airspeed, flight path angle, and bank angle were used

as input as in guidance law. Also, heading angle was

controlled directly by the guidance system. In our

study, the longitudinal dynamics of the UAV is

specifically investigated and trajectory tracking

application is applied by using MPC.

4.2 Model Predictive Control

Algorithm for Reference Tracking

The algorithm starts by converting 2D image

coordinates into 3D world coordinates. Then, initial

and final points are determined from the current

location data to track the landing path. The prediction

horizon (N) defines the number of future time steps

for which the system's behavior is forecast. The

control horizon (M) indicates the number of future

time steps over which the control inputs are

optimized; this number must always be less than or

equal to the prediction horizon.

(Camacho et al.,

1999).

Cost Function: The MPC objective is to minimize a

quadratic cost function that penalizes state deviations

from the reference trajectory and control effort. This

is formulates as a Quadratic Program (MathWorks,

n.d). Minimizing the cost function formula is given as

follows:

𝐽

𝑢





𝐻𝑢



𝑓



𝑢



(9)

where 𝐻 and 𝑓 are the components derived from

augmented prediction matrices and cost weights.

Weight matrices are represented as 𝑄=

𝑑𝑖𝑎𝑔(𝑄





,𝑄





,𝑄





,𝑄



) and 𝑅=

𝑑𝑖𝑎𝑔(𝑅



,𝑅



,𝑅





) While Q penalizes deviation in

reference states, R penalizes control effort. These

Vision-Based Autonomous Landing for the MPC Controlled Fixed Wing UAV

231

weight matrices should be expanded for all horizons.

Then, the matrices become:

𝑄



=𝐼



⊗𝑄

(9)

𝑅



=𝐼



⊗𝑅

(10)

The predicted states during the horizon is:

𝑥





=𝛷𝑥

(

𝑘

)

+𝛤𝑢







(11)

where Φ



=𝐴



and Γ



=𝐴



𝐵 for 𝑗𝑖𝑁 and

𝑗𝑀, otherwise Γ



=0 Quadratic program

matrices are given (MathWorks, n.d):

𝐻=𝛤



𝑄



𝛤+𝑅



(12)

𝑓

=𝛤



𝑄



(𝛷𝑥

(

𝑘

)

−𝑟

(13)

where r is a stacked vector of desired states.

Constraints are needed to design a good MPC and for

the control input it is formulated as:

𝑢



𝑢



𝑢



(14)

Reference Trajectory: The reference trajectory 𝑟 for

the prediction horizon is dynamically generated based

on the current UAV position and the predefined

landing path segment. The current segment vector is:

𝑣



=𝑃



−𝑃



(15)

The progress parameter 𝑠 which is from 0 to 1 is

calculated along the segment.

𝑠



=𝑑𝑜𝑡(𝑃



−𝑃



,𝑣



)/𝑣







(16)

Future reference points are calculated by expanding

this progression according to the desired path speed

and forecast time.

Quadratic Program Solver and State Update: The

Matlab quadprog function solves the QP problem to

obtain the optimal control sequence (MathWorks,

n.d). The next state of the UAV is simulated using a

simple numerical integration of the linearized

dynamics (Euler method).

𝑃

(

𝑥

𝑘+1

)

=𝑃

(

𝑥

𝑘

)

+𝑉𝑐𝑜𝑠 (𝜓)𝑇



(17)

𝑃

(

𝑦

𝑘+1

)

=𝑃

(

𝑦

𝑘

)

+ 𝑉𝑠𝑖𝑛 (𝜓)𝑇



(18)

𝑃

(

𝑧

𝑘+1

)

=𝑃

(

𝑧

𝑘

)

+𝑍



𝑇



(19)

𝜓

(

𝑘+1

)

=𝜓+𝜔𝑇



(20)

𝑉

(

𝑘+1

)

=𝑉+𝛼



𝑇



(21)

This process is repeated by updating the state, re-

linearizing the model, creating a new reference,

solving the QP and applying control until the UAV

reaches the landing target.

5 SIMULATION RESULTS

Map points marked on the image were obtained as

vSLAM output from the MATLAB simulation.

Subsequently, segmentation was applied to the last

read image, dividing it into regions. Utilizing the

segmented regions and current position information,

a bounding box was drawn around the designated

landing area. The outputs obtained are provided

below.

Figure 1: The original image.

The landmarks are printed on all the images, but

the final frame is the interested one at all. Because the

landing area is appeared thoroughly in the last frame.

Figure 2: The land marked image.

Here, the image is segmented into regions by

Watershed method. Each region is shown in a

different colour to discriminate the landing zone

easily.

Figure 3: The segmented image

After the image was segmented into regions, the

landing zone was determined. Then, to define the

landing zone, a bounding box is drawn to cover this

zone properly. The landmarks within the bounding

box boundaries were detected to help to controller to

track a reference trajectory.

ICINCO 2025 - 22nd International Conference on Informatics in Control, Automation and Robotics

232

Figure 4: Bounding box representation.

Finally, according to the reference trajectory over

the landing zone, a start and end point are determined.

Then, the movement of the UAV on this trajectory is

controlled with MPC algorithm.

Figure 5: Trajectory tracking result of MPC.

6 CONCLUSION

This paper successfully demonstrates a robust vision-

based autonomous landing system for a fixed-wing

UAV that integrates a Linear Model Predictive

Control (MPC) strategy with Visual Simultaneous

Localization and Mapping (vSLAM). By leveraging

an SVD-based Kalman filter in the vSLAM

framework, we achieve improved accuracy and

numerical stability in map point updates and reduce

common issues such as noise accumulation and

computational errors. The image processing and

segmentation module using Watershed Transform

and incorporating real-time vSLAM location data

effectively identifies and defines the target landing

area, enabling precise placement of a bounding box.

This visual information is then seamlessly fed to the

linearized MPC controller, which dynamically tracks

a predefined landing trajectory. Simulation results

clearly demonstrate the system's ability to accurately

follow the desired path over the designated landing

zone and validate the effectiveness of our combined

vSLAM-MPC architecture for safe and autonomous

fixed-wing UAV landings.

REFERENCES

Amasyalı, M. F., Münük, A., & Salı, M. (2010). Kamera

görüntülerinden gidilen yolun kestirimi [Estimation of

the path traveled from camera images]. Dicle

Üniversitesi Mühendislik Fakültesi Mühendislik

Dergisi, 1(1), 5–11.

Andert, Franz, and Luis Mejias. “Improving Monocular

SLAM with Altimeter Hints for Fixed-Wing Aircraft

Navigation and Emergency Landing.” 2022

International Conference on Unmanned Aircraft

Systems (ICUAS), 1 June 2015,

Behera, L., Kamath, A. K., & Tripathi, V. K. (2020).

Vision-based autonomous control schemes for

quadrotor unmanned aerial vehicle.

Boij, J. N. (2022). Localization of Combat Aircraft at High

Altitude using Visual Odometry.

Camacho, Eduardo F, and Carlos Bordons. Model

Predictive Control. Advanced Textbooks in Control and

Signal Processing, 1 Jan. 1999.

Campoy, Pascual, et al. “Computer Vision Onboard UAVs

for Civilian Tasks.” Journal of Intelligent and Robotic

Systems, vol. 54, no. 1-3,7Aug.2008,pp.105–135,

https://doi.org/10.1007/s10846-008-9256-z.

García-Pulido, J.A., et al. “Recognition of a Landing

Platform for Unmanned Aerial Vehicles by Using

Computer Vision-Based Techniques.” Expert Systems

with Applications, vol. 76, June 2017, pp. 152–165,

https://doi.org/10.1016/j.eswa.2017.01.017.

Gautam, A., Sujit, P., & Saripalli, S. (2014). A survey of

autonomous landing techniques for UAVs. 2022

International Conference on Unmanned Aircraft

Systems(ICUAS), 1210–1218. https://doi.org/10.

1109/icuas.2014.6842377

Gavilan, F., Vazquez, R., & Esteban, S. (2015). Trajectory

tracking

for

fixed-wing

UAV

using

model

predictive control and adaptive backstepping. IFAC-

PapersOnLine,48(9),132–137.

Gavilan, Francisco, et al. “An Iterative Model Predictive

Control Algorithm for UAV Guidance.” IEEE

Transactions on Aerospace and Electronic Systems, vol.

51, no. 3, July 2015,pp.2406–2419,

Hang, Y., Zhang, T., & Huang, S. (2018, May).

Comparison of EKF based SLAM and optimization

based SLAM algorithms. Paper presented at the 13th

IEEE Conference on Industrial Electronics and

Applications (ICIEA), Wuhan,China.IEEE.

https://doi.org/10.1109/iciea.2018.8397911.

Kalay, Adnan, and İlkay Ulusoy. “Görüntü Tabanlı Eş

Zamanlı Lokalizasyon ve Harita Çıkarma: Stereo ve

Mono SLAM.” IEEE Sinyal İşleme ve İletişim

Uygulamaları Kurultayı, Kocaeli University, 2009.

Kaur, Dilpreet, and Yadwinder Kaur. “Various Image

SegmentationTechniques:AReview.”International

Journal of Computer Science and Mobile Computing,

vol.3, no.5,2014, pp.809–814.

Kazerouni, I. A., Fitzgerald, L., Dooly, G., & Toal, D. (2022).

A survey of state-of-the-art on visual SLAM. Expert

Systems With Applications, 205, 117734. https://doi.

org/10.1016/j.eswa.2022.117734

Kong, W., Zhou, D., Zhang, D., & Zhang, J. (2014,

October). Vision-based autonomous landing system for

unmanned aerial vehicle: A survey. Paper presented at

the 2014 International Conference on Multisensor

Fusion and Information Integration for Intelligent

Vision-Based Autonomous Landing for the MPC Controlled Fixed Wing UAV

233

Systems (MFI),Beijing,China,IEEE. https://doi.org/10.

1109/MFI.2014.6997750

Lemaire, Thomas, et al. “Vision-Based SLAM: Stereo and

Monocular Approaches.” International Journal of

Computer Vision, vol. 74, no. 3, 9 Feb. 2007, pp.343–

364.

Liu, et al. “An Onboard Vision-Based System for

Autonomous Landing of a Low-Cost Quadrotor on a

Novel Landing Pad.” Sensors, vol. 19, no. 21, 29 Oct.

2019, p.4703,

https://doi.org/10.3390/s19214703.

MathWorks. (n.d.). Optimization Problem. MATLAB &

Simulink, from https://www.mathworks.com/help/mpc/

ug/optimization- problem.html

MathWorks. (n.d.). What Is Model Predictive Control

from https://www.mathworks.com/help/mpc/gs/what-is-

mpc.html

MathWorks. (n.d.).Monocular Visual Simultaneous

Localization and Mapping. MATLAB & Simulink.

https://www.mathworks.com/help/vision/ug/mo

nocular-visual-simultaneous-localization-and-

mapping.html

Patruno, C., Nitti, M., Petitti, A., Stella, E., & D’Orazio, T.

(2018). A Vision-Based approach for unmanned aerial

vehicle landing. Journal of Intelligent &

RoboticSystems,95(2),645–664. https://doi.org/

10.1007/s10846-018-0933-2

Purdue University. (2021). UAV_Dataset home page.

https://engineering.purdue.edu/~bouman/UAV_

Dataset/

Rabah, M., Rohan, A., Talha, M., Nam, K., & Kim, S. H.

(2018). Autonomous vision-based target detection and

safe landing for UAV. International Journal of Control

Automation and Systems,16(6),3013–3025. https://doi.

org/10.1007/s12555-018-0017-x

Riccardo Giubilato, et al. A Comparison of Monocular and

Stereo Visual FastSLAM Implementations. 1 June

2016,pp.227–232, https://doi.org/10.1109/

metroaerospace.2016.75 73217..

Rondon, Eduardo, et al. “Vision-Based Altitude, Position

and Speed Regulation of a Quadrotor Rotorcraft.” 2011

IEEE/RSJ International Conference on Intelligent

Robots and Systems, 1Oct.2010,pp.628–633,

https://doi.org/10.1109/iros.2010.5652745.

Sun, Da-Wen. Computer Vision Technology for Food

Quality Evaluaton. Boston, Ma, Elsevier, 2017.

Unal, G. (2021). Visual target detection and tracking based

on Kalman filter. Journal of Aeronautics and Space

Technologies, 14(2), 251–259.

Urbanski, Konrad. “Control of the Quadcopter Position

Using Visual Feedback.” International Conference on

Mechatronics - Mechatronika, 1 Dec. 2018.

Vidal, V. F., Honório, L. M., Santos, M. F., Silva, M. F.,

Cerqueira, A. S., & Oliveira, E. J. (2017, May 1).

UAV vision aided positioning system for location and

landing. IEEEXplore.

https://doi.org/10.1109/CarpathianCC.2017.

797 0402

Wang, L., Libert, G., & Manneback, P. (1992, December).

Kalman filter algorithm based on singular value

decomposition. Paper presented at the 31st IEEE

Conference on Decision and Control, Tucson,

AZ,USA.IEEE.

Xin, L., Tang, Z., Gai, W., & Liu, H. (2022). Vision-Based

Autonomous Landing for the UAV: A Review.

Aerospace, 9

(11), 634.

Zhang, Yanhao, et al. “Comparison of EKF Based SLAM

and Optimization Based SLAM Algorithms.” OPUS -

Open Publications of UTS Scholars (University of

Technology Sydney), 1May2018.

ICINCO 2025 - 22nd International Conference on Informatics in Control, Automation and Robotics

234