Real-Time UAV Navigation Using Machine Learning, Satellite Images, and Continuously Operating High-Speed Cameras with Enhanced Image Matching and Robust Feature Tracking

  • NAVIGATION: Journal of the Institute of Navigation
  • June 2026,
  • 73
  • navi.775;
  • DOI: https://doi.org/10.33012/navi.775

Abstract

Machine-learning-based navigation systems have become crucial for unmanned aerial vehicles (UAVs), particularly in environments affected by signal jamming and laser pointing. Traditional sensors such as global navigation satellite system sensors, lasers, and radar provide essential navigation data but have limitations, including the need for continuous pointing and vulnerability to interference. Advancements in imaging technology have rendered cameras a viable alternative. This article presents a real-time navigation solution for high-speed UAVs that minimize disruption from environmental dynamics. Key methodologies include (i) creating a virtual environment for model training, (ii) extracting UAV flight dynamics for simulation, (iii) applying and validating image-matching techniques, and (iv) developing a visual tracking algorithm for navigation between successive image-matching steps. The approach utilizes UAV-captured images alongside satellite images of the flight region. Experimental results and analyses demonstrate the effectiveness of various image-matching methods and visual tracking algorithms, showcasing the capabilities of the proposed solution compared with available techniques.

Keywords

1 INTRODUCTION

This section outlines the problem description, overviews related methods in the literature, and highlights the novel contributions of the proposed method.

1.1 Problem Description

Radar, laser, and global navigation satellite system (GNSS) sensors are frequently employed in navigation solutions for specific targets. However, these sensors exhibit several disadvantages that can significantly impact their performance. For GNSS sensors, critical factors include signal jaming and the changing positions of satellites. In the case of laser systems, continuous marking from a certain distance and signal reflection are important considerations. Radar systems face challenges related to signal jaming, signal reflection, and the radar cross-section of the target. These limitations influence the overall effectiveness of laser, radar, and GNSS sensors.

In systems relying solely on an inertial measurement unit (IMU), the positioning error grows over time and can reach the kilometer scale. In contrast, a hybrid system integrating a camera, IMU, and barometer reduced the error to approximately 300 m in subsonic flight simulations (Luo et al., 2020). An IMU used for altitude estimation allowed the detection of a target region when an unmanned aerial vehicle (UAV) was stationary or moving slowly (Sun et al., 2023). Consequently, this IMU is suitable primarily for low-speed or near-stationary UAV operations.

Although IMUs are effective in hybrid operations with a camera during short-duration, low-altitude flights, the error margin increases significantly during long-duration or high-altitude operations, adversely affecting navigation performance. Additionally, high-performance IMU sensors suitable for high-altitude, high-speed flights are more expensive than the camera itself, further increasing the overall system cost.

In contrast, image-based navigation systems, while useful, are limited by their two-dimensional nature, which hinders the acquisition of precise depth information and leads to high error margins. These systems may be suitable for ground vehicles but exhibit elevated error rates when applied to UAVs operating in three-dimensional environments, primarily owing to the impact of perspective angular differences on accuracy.

Although infrared (IR) and stereo cameras have been employed in camera-based navigation solutions, they are not well-suited to the problem addressed in this study. IR cameras suffer from several limitations, including prominent heat signatures, high cost, loss of visible environmental features, a lack of available IR reference imagery for the target area, and relatively low resolution. These factors lead to inevitable performance degradation in image-matching tasks at high altitudes. While stereo cameras enable simultaneous observation and depth estimation using multiple sensors, they present notable disadvantages, such as increased system cost, potential adverse effects on UAV aerodynamics, and unreliable depth estimation at high altitudes.

Navigation can target specific objects or areas, relying on object detection models that require sufficient data. However, in scenarios with multiple objects, identifying a primary target can complicate autonomous motion. Thus, object-based navigation is inadequate when specific objects are absent or when training data are limited.

This article develops novel image-matching-based models that do not depend on object-based detection. The primary challenge addressed is optimal route navigation for UAVs, such as missiles or drones, operating at high altitudes and supersonic speeds, independent of signal jaming and reflection. The proposed method matches satellite images with UAV-captured images, enhancing navigation robustness by visually tracking features. A novel algorithm corrects matched features and minimizes errors, while an adaptive anomaly detection filter addresses inaccuracies based on sequential camera movement, autonomously selecting reference images for the UAV path.

The developed algorithm generates outputs suitable for use as inputs to localization and mapping algorithms. However, conventional simultaneous localization and mapping (SLAM) approaches do not apply to the scenario considered in this context. Standard SLAM algorithms require a relatively static background while foreground elements change, meaning that at least one of the XYZ axes should remain approximately constant. In the present problem, the images are captured as the camera directly faces the ground, producing background-dominated frames. Because the UAV moves at high speed along all three axes (X, Y, and Z), the background continuously changes, rendering standard SLAM methods ineffective in this context.

Although high accuracy can be achieved at low altitudes and over short distances, performance degrades at higher altitudes, where the surface appears increasingly planar in captured images, leading to errors in regions with elevation variation. In addition, SLAM algorithms inherently suffer from a cumulative error that increases with distance. The proposed method addresses these limitations by comparing incoming images with reference images acquired at multiple altitudes. During each matching cycle, error correction, delay compensation, and robust feature tracking are performed, resulting in improved performance compared with conventional SLAM approaches reported in the literature (Navard & Yilmaz, 2024).

SLAM models do not require an external reference image; instead, they estimate pose by sequentially matching frames acquired from onboard cameras. However, for a UAV operating at high speed and descending from high altitude toward a designated target, these models may produce erroneous estimates, particularly during trajectory correction maneuvers. Reliable mapping and localization depend on accurate initialization, including a precise knowledge of the system's starting position and the intrinsic parameters of the camera.

Consequently, the performance of SLAM algorithms is highly sensitive to uncertainties in both the initial pose and camera calibration. To mitigate these limitations, hybrid approaches that integrate GNSS and/or IMU measurements are typically required to enhance robustness and accuracy.

In this study, Unreal Engine (Epic Games, 2019) was utilized to simulate the movements of a UAV and to prepare a visual data set. AirSim (Shah et al., 2018) facilitated the transfer of images from the game engine to the Python environment (Van Rossum & Drake, 2009). For the UAV movements and dynamics, OpenRocket (Niskanen, 2013) and AirSim's drone modules were employed. The three-dimensional environments in the game engine include both ready-made virtual environments and environments created using Google Earth (Gorelick et al., 2017) and geographical landforms. Satellite images used as references were sourced from the Google Earth engine (Gorelick et al., 2017) for real environments and from the game engine for virtual environments. Images of virtual environments, mimicking satellites and incorporating relevant camera parameters, were obtained and used as reference images within the created simulation environment. These simulation images were then fed to the proposed algorithm in real time.

Satellite images of the specified target region at different zoom levels were matched with one another to autonomously select reference images. From the features identified through this matching process, the features that could provide the most accurate alignment along a trajectory that could be followed by the UAV camera were determined. A new, highly accurate method was developed using reference images at various zoom levels.

The method developed in this article is primarily designed for UAVs but can also be adapted for missiles under certain conditions. A key consideration for missiles is the dynamic change in the center of gravity (Cg) and center of pressure (Cp) during flight. In contrast, the UAV simulated in this study is less affected by such changes owing to its smaller dimensions and relatively wider wingspan. For stable missile flight, the Cp is typically positioned below the Cg; a larger gap increases stability but reduces responsiveness to navigation commands. In unstable conditions, a missile may deviate from its trajectory because of environmental factors, independent of navigation inputs.

Furthermore, as the missile consumes fuel and loses weight, the Cg shifts during flight. To ensure both stability and effective navigation, a pre-flight analysis based on trajectory and range is required. The method proposed in this study addresses this challenge by dynamically weighting the vector toward the target point according to changes in the Cg–Cp balance over time, thereby maintaining stability while optimizing responsiveness to navigation outputs.

The proposed approach addresses the matching of satellite imagery with UAV camera images, acknowledging the angular, environmental, and perspective differences inherent in the data. These discrepancies can lead to performance degradation and incorrect matching results. Given the UAV's high-speed descent from altitude, real-time navigation is critical, necessitating the use of high-performance, low-power computing solutions resilient to environmental conditions. The objective was to develop a robust, low-cost, and highly accurate camera-based navigation system for a high-speed UAV operating continuously toward dynamic or static targets.

Feature extraction and image matching involved estimating the camera's motion in a two-dimensional plane using only camera images. By calculating the average shift and rotation for all features, the change in features from the image center to the edges was determined, providing insight into potential pixel changes due to camera movement. This information is crucial for eliminating incorrectly matched features.

1.2 Related Work

This section provides a literature review of methods relevant to this article: (i) image matching, (ii) visual point-based tracking, (iii) visual boundary box-based tracking, and (iv) image-matching navigation. Thus far, the addressed problem has been underexplored, largely because of insufficient image data from high-speed UAVs at high altitudes. Moreover, previous studies have not concurrently tackled critical issues, including data set preparation using Unreal Engine (Epic Games, 2019), UAV kinematics under varying conditions, and real-time navigation based solely on imagery. The simultaneous correction of features, elimination of faulty matches, and autonomous selection of optimal reference satellite images for enhanced navigation accuracy are the key novel contributions of this article. Thus, the methods developed herein significantly advance UAV navigation methodologies.

Given the high-speed movement of UAVs, real-time target detection is essential for accurate navigation. To address this, a visual tracking algorithm with a shorter cycle time was developed. Traditional tracking algorithms (Ahmine et al., 2019; Fiaz et al., 2019) can operate efficiently in central processing unit (CPU) environments but suffer from increased pixel error due to environmental effects and dynamic camera motion. Deep-learning models (Held et al., 2016; Karaev et al., 2024; Wojke et al., 2017) yield good results but require powerful graphics processing units (GPUs). Although the methods of Held et al. (2016) and Wojke et al. (2017) approach real-time performance, unfortunately, the pixel error escalates with each tracking iteration. The approach proposed in this paper both produces satisfactory results and can be used in real time.

1.2.1 Image Matching

The methods presented herein overcome the previously mentioned challenges of image-based navigation by using an image-matching method that extracts and matches features from images. Traditional algorithms, such as the scale-invariant feature transform (Tareen & Saleem, 2018), oriented FAST and rotated BRIEF (ORB) algorithm (Tareen & Saleem, 2018), and accelerated KAZE algorithm (Tareen & Saleem, 2018), can match sequential video images with minimal errors but exhibit limitations when faced with angular and perspective differences, resulting in high error rates, especially between satellite and UAV images.

In contrast, deep-learning models enhance accuracy despite image discrepancies. The combination of SuperPoint and SuperGlue (Xu et al., 2024) has shown significant improvements, with LightGlue (Xu et al., 2024) reducing the cycle time by 30%. While models such as the local feature transformer (LoFTR), dense kernelized feature matching (DKM), RoMa (Xu et al., 2024), and DeDoDe (Edstedt et al., 2024) models improve accuracy, they still exhibit performance degradation under varying conditions. Additionally, issues with incorrect feature matches underscore the need for effective anomaly detection, as existing models like AffNet (Mishkin et al., 2018) and AdaLAM (Cavalli et al., 2020) have not adequately addressed these challenges.

1.2.2 Visual Point-Based Tracking

This method involves altering features or pixel points across sequential images, offering advantages over frame-based tracking owing to multiple trackable states. Traditional approaches, such as the Lucas-Kanade approach (Ahmine et al., 2019) and its improved variant (Ahmine et al., 2019), enhance performance but reduce algorithm speed. Notable deep-learning models include the cell-keypoints model (Chan et al., 2017) and CoTracker (Karaev et al., 2024), which excel in indoor and close-up outdoor environments. However, edge tracking in high-altitude and rural areas remains challenging, yielding suboptimal results. Recent models capable of tracking any point impose a high processing load (Karaev et al., 2024), hindering real-time performance even with advanced GPUs, particularly as point counts increase (Güven & Yetik, 2024).

1.2.3 Visual Boundary Box-Based Tracking

This method involves defining a specific frame in an image and tracking it across sequential images, resulting in fewer states compared with point-based tracking. However, this approach is highly dependent on frame size and pattern, leading to increased error rates in sequential images. Traditional algorithms, such as the kernelized correlation filter (KCF) (Fiaz et al., 2019), multiple instance learning (MIL) (Fiaz et al., 2019), and channel and spatial reliability tracker (CSRT) (Fiaz et al., 2019), operate quickly in CPU environments but are sensitive to camera perspective changes. Recently developed deep-learning models, including generic object tracking using regression networks (GOTURN) (Held et al., 2016), simple online and real-time tracking with a deep association metric (DeepSORT) (Wojke et al., 2017), and Track to Detect and Segment (Wu et al., 2021), yield good results in small frames but require GPU environments for real-time performance, particularly in high-speed, high-altitude, and rural settings.

1.2.4 Image-Matching-Based Navigation

Image-based navigation has recently gained traction for drones and UAVs, yet the UAV characteristics considered in this article differ significantly from those in the existing literature. Notably, no navigation solution identifies the optimal route based solely on images for a UAV flying continuously at high speed from a high altitude. Previous studies, such as the work by Liu and Liu (2021), focused on object tracking with a moving camera, which lacked mobility and did not track three-dimensional movement. Additionally, Amiri and Moradi (2016) conducted image matching and homography estimation without providing a navigation solution or specific target searches.

Image-matching-based navigation enables target region detection with minimal reference images. These reference images are matched with those captured from the UAV. However, both traditional and contemporary image-matching models are significantly affected by angular differences and environmental variables. Notably, angular and perspective discrepancies between the reference and target images can severely impair performance. Recent advancements in image matching include models such as those described by Potje et al. (2024) and Xu et al. (2024). Traditional methods, including that of Tareen and Saleem (2018), can operate efficiently on CPUs without GPUs but are highly sensitive to variations in angle, perspective, and environmental conditions. In contrast, deep-learning-based methods offer improved robustness against these challenges; however, such methods are constrained by the data sets and architectures used for training.

A study by Kinnari et al. (2023) demonstrated GNSS-independent and seasonally robust localization over a wide area; however, the localization error increased with larger viewpoint discrepancies between query images and reference images. In addition, the substantial memory requirements of the reference image database and the extensive search space led to long processing times, precluding real-time operation. The problem considered in this work involves less challenging conditions than those addressed in the referenced study, as it does not entail rapid or continuous motion.

In another study, Zhao et al. (2023) applied a three-dimensional virtual simulation environment, similar to that employed in this work. However, owing to the absence of a game engine, the simulation lacked the level of complexity and realism achieved in the present study. Moreover, the proposed navigation system was not designed for fast, continuously moving UAVs. Rather, their study focused on a highly maneuverable UAV operating at low altitude and low speed and is therefore not applicable to the high-altitude, high-speed scenarios with limited maneuverability and strict timing constraints considered in this article.

More recent matching approaches, such as that presented by Zhang et al. (2025), emphasize computational efficiency and real-time performance by incorporating lightweight attention mechanisms and enhanced feature representations. Nevertheless, the increased pipeline complexity and limited validation in dynamic or cross-domain environments hinder the applicability of these approaches to continuously operating high-speed UAVs. This paper seeks to address key challenges in visual navigation for high-speed, continuously operating UAVs, including scalability, generalization, and real-world deployment.

The closest relevant studies, i.e., those of Güven and Yetik (2024) and Zhang et al. (2021), explored matching satellite images with UAV images but did not prioritize time constraints or high-speed trajectory tracking, which are key requirements. While Zhang et al. (2021) developed a navigation solution, this approach fell short in accuracy and error minimization. Conversely, Güven and Yetik (2024) created a data set for high-altitude images but did not address UAV kinematics.

This article proposes a navigation solution that optimally considers factors such as wind speed and environmental conditions, aiming for centimeter-level accuracy in feature error, a significant advancement over previous image-matching-based navigation studies, and thus fills a gap in this area.

1.3 Contributions

Inspired by our previous machine-learning work on feature tracking (Güven & Yetik, 2024), a robust point-based tracking algorithm was developed. Our previous method addresses high-speed camera scenarios; however, it relies on fixed orbits in the data set, limiting software adaptability. Additionally, the navigation solution and the issue of incorrect feature returns are not adequately explored.

Tracking algorithm errors normally accumulate across sequential images, hindering continuous tracking. To enhance robustness, image-matching and tracking algorithms are executed in parallel, mutually triggering each other. In our proposed solution, the corrected image-matching results are integrated into the tracking algorithm, improving the overall performance.

The major contributions of this study are as follows:

  • A simulation environment is established for a UAV operating at high altitudes and supersonic speeds, accompanied by a prepared data set.

  • A new algorithm is created for the autonomous selection of reference satellite images for matching.

  • Faulty features from the image-matching model are detected, and filters for outliers are developed.

  • A robust visual tracking model is designed for features identified through image matching.

  • An image-based navigation solution is developed to track the optimal route to a target in real time for UAVs at high altitudes and supersonic speeds.

These enhancements enable the real-time navigation of a continuously operating high-speed UAV with negligible error.

2 METHODOLOGY

This section presents the models utilized for the developed method, the application of these models, and the creation of the simulation environment. The flow of the algorithm presented in this article includes the following steps: (i) application of image-matching models, (ii) outlier detection in features, (iii) autonomous reference image selection, (iv) robust feature tracking, (v) navigation output for the UAV, and (vi) data set preparation. The steps of the methods employed within this framework are explained in detail below.

2.1 Real-Time Correction and Visual Tracking of Features from Satellite and Continuously Operating High-Speed UAV Images Along an Optimal Trajectory

This section presents and explains the algorithm flow developed as a solution to the problem addressed in this article. Figure 1 illustrates the workflow of the algorithm. The method initiates with preprocessing, which involves the autonomous selection of satellite images. Subsequently, the main process is executed, where image matching and its subsequent component, anomaly detection, operate in parallel with homography estimation (Luo et al., 2023) and robust feature tracking. A detailed description of each process is provided in Section 2.2.

Workflow of the proposed algorithm
FIGURE 1

Workflow of the proposed algorithm

2.2 Blocks of the Developed System

This section outlines the algorithms employed in the innovative method described in this article.

2.2.1 Autonomous Selection of Reference Satellite Images

Selecting reference images with the appropriate order and orientation is essential for optimal image-matching performance. The orientation of reference satellite images must closely align with the camera orbit; otherwise, significant performance degradation occurs in image-matching models. Widely used models, including SuperGlue (Xu et al., 2024), LightGlue (Xu et al., 2024), LoFTR (Xu et al., 2024), and DKM (Xu et al., 2024), have shown performance degradation due to angular differences, with roll angle limits measured at 40°, 35°, 45°, and 45° and perspective angle limits at 35°, 30°, 40°, and 40°, respectively.

Reference images are sourced from Google Earth (Gorelick et al., 2017). Discrepancies exist between camera and satellite specifications, influenced by lens options. The default field of view (FOV) for Google Earth (Gorelick et al., 2017) is 60°, whereas the FOV for the camera used in this article is 50°. Three reference images, selected based on image change time and device performance, include the first reference image chosen according to the maximum altitude reached by the air vehicle, with the first zoom level determined by Equation (1):

Zi=hitan⁡(50∘/2)tan⁡(60∘/2)1

where Zi represents the zoom level for the reference image and hi denotes the vehicle's altitude. For an altitude of 6000 m, the first reference image should be captured at approximately 4800 m above ground.

All reference images are obtained at a resolution of 1920 × 1080, matching the camera's specifications, through web scraping in Python (Van Rossum & Drake, 2009). Zoom levels are determined using Equation (1), resulting in a zoom level of 4800 m for an altitude of 6000 m, 2800 m for an altitude of 3500 m, and 1000 m for an altitude of 1300 m. The final zoom level is capped at 1000 m, as Google Earth (Gorelick et al., 2017) does not provide further zoom options. After the zoom levels have been established, images are autonomously selected using Equation (2):

Ii={ϵ(Sk2),ifi=1ϵ(Si,Sk),otherwise2

In Equation (2), the first satellite image S1 is selected from an analysis of each dual satellite image at the specified zoom level, focusing on the route and target point. This analysis is conducted using the image-matching model described in Section 2.2.2. The symbol ϵ denotes the target point error in pixels between images. For subsequent reference images, Si is chosen by comparing the matching error with the previous image.

2.2.2 Implementation of Image-Matching Models

Multiple models and algorithms were evaluated in the development of an image-matching solution suitable for our problem. The performances of the traditional and deep-learning algorithms tested are presented in Section 3. Both runtime and matching accuracy were critical for the solution. Examination revealed that the LightGlue model (Xu et al., 2024) outperformed other deep-learning models and traditional algorithms. Similar to other models, LightGlue has hypothetical constraints for matching images. Training and comparisons with the data set indicated effective operation within ± 40° for angular differences due to rotation and ± 30° for perspective differences, but poor or no results were observed in other scenarios. To reduce the runtime, images were downsampled to a width of 640 pixels, considering the aspect ratio, the capacity of the Nvidia Jetson Xavier AGX card, and the model's runtime.

OpenCV (Joshi, 2015) provides several image-resizing algorithms. Among these, the AREA algorithm (Joshi, 2015) offers superior performance in preserving image quality during downsampling, resulting in reduced distortion compared with other resizing approaches (Moukthika, 2025). Minimizing image distortion is critical for accurate feature extraction and enables the detection of a greater number of reliable features than alternative downsampling methods. Here, we have the following:

ID=AREA(I)3

The input image is first converted to grayscale format to reduce the processing time of the deep-learning model. Then, in Equation (3), the input image is downsampled by the AREA (Joshi, 2015) method, preserving the original aspect ratio of the image:

F,D=E(ID)4

where E denotes the feature extraction deep-learning model, which is SuperPoint (Xu et al., 2024), for a given image I. F and D are the outputs of the model, corresponding to the extracted features and their descriptors, respectively:

R=M(DR,DT)5

The matching model, denoted as M, utilizes LightGlue (Xu et al., 2024) with the descriptors of the reference image (DR) and target image (DT) as inputs, producing R as the output to obtain matching features. LightGlue (Xu et al., 2024) produces mutually consistent feature matches by using a lightweight, attention-based architecture that dynamically adapts its computation to the difficulty of the matching task. By prioritizing speed, memory efficiency, and scalability, LightGlue reframes feature matching as a dynamically pruned, attention-based bipartite graph problem, enabling high accuracy at low computational cost. For real-time processing, the image-matching model operates in parallel, triggering a separate parallel tracking process. The target point in the reference image facilitates locating the intended point in the camera view. Consequently, the point-tracking model developed in this article can track the points of interest in real time.

The space sets formed by UAV camera images and satellite images are represented as Cj and Ii, respectively. The features matched through image matching, along with their descriptors, are denoted as F and D. The camera images are represented as FT(j) and DT(j), whereas the reference satellite images are denoted as FR(i) and DR(i). Because the results are determined within the image space of the camera, the expression Rj is utilized.

In Algorithm 1, reference images are initially downsampled and converted to grayscale. Features and descriptors for each reference image are then extracted and stored prior to the main process. If the camera frame number exceeds the matching process count, the next camera frame is obtained and preprocessed similarly to the reference images. Reference images vary with the altitude of the air vehicle. Despite the capabilities of deep-learning models, they may not match images with significant perspective differences, such as altitude variations. If the current altitude is lower than that of a reference image, matching will proceed with the next reference image.

Algorithm 1 Implementation of Image-Matching Models

Require: ∃Ii ∈ A and ∃CJ ∈ G and ∃hi ∈ H

Ensure: i ∈ ℤ and 0 ≤ j ≤ ∞, j ∈ ℤ

 IiD←AREA(Ii)// Downsample the reference images

 FR(i),DR(i)←E(IiD)// Extract and assign features and descriptors from reference images

 n ← j

 while j ≠ ∞ do

  if n ≤ j then

   if hi ≤ ALT then // Check the estimated altitude

    Update DR(i)// Update if the estimated altitude is less than the reference image altitude

   end if

   CjD←AREA(Cj)// Downsample the camera image

   FR(i),DT(i)←E(CiD)// Extract and assign features and descriptors from camera images

   Rj←M(DR(i),DT(j))// Match descriptors from the reference and camera images

  else

   Rj ← None

  end if

 end while

2.2.3 Detection of Anomalies Among Matching Features

The matching performances of features from existing models and filtering methods in the literature are compared in Section 3. The comparison indicates that most features are incorrectly matched for the problem addressed in this paper, resulting in meaningless matching scores. The high number of incorrect matches renders traditional anomaly replacement methods, such as random sample consensus (Raguram et al., 2008), ineffective, whereas deep-learning models like AdaLAM (Cavalli et al., 2020) also prove insufficient.

A model based on camera movement and edge variability is proposed to achieve more accurate and robust results. Features inconsistent with camera movement are eliminated. Homography is estimated from matched points around specific seed points in the reference image, considering local correct matching probabilities (Luo et al., 2023). The homography matrices for each seed point are fused to determine the target point's location in the reference satellite image in pixel coordinates. Homography is preferred over affine transformation (Luo et al., 2023) because of the three-dimensional camera movement and pronounced perspective differences.

In Figure 2, five fixed seed boxes are selected, with the middle box indicating the target region. The other four boxes are positioned with a fixed gap of 50 pixels from the target box, each covering a square of 100 pixels in width. Homography estimation is performed for the middle box if it contains at least four matched points with the camera frame. Each estimated point is then used for target point estimation, resulting in a more stable and robust outcome by considering the local match consensus.

Anomaly detection with local seed points
FIGURE 2

Anomaly detection with local seed points

The target point, represented as pixel coordinates tx and ty, is selected by the user, with the assumption that these points are near the center of the image. Subsequently, four additional target points are generated in the reference image, as indicated in Equation (6). Here, Mca denotes the number of matches within the relevant box of the shifted target point. In the camera frame, corresponding target points are identified based on the reference image. Finally, homography estimations are obtained by using the matches surrounding the target points that exhibit the highest matching scores from the deep-learning model in Section 2.2.2:

txa,tya={tx−150,ty, if a =left,Mca≥4tx,ty−150, if a =up,Mca≥4tx+150,ty, if a =right,Mca≥4tx,ty+150, if a =bottom,Mca≥46

2.2.4 Robust Visual Tracking of Features

In machine-learning models, the extraction of descriptive features is crucial. The camera used operates at 30 frames per second (fps), and the high-speed UAV is expected to travel at least 10 m per frame. Considering the 50° FOV of the camera, the tolerable frame-miss margin is approximately 100 m, depending on the altitude at which navigation is initiated. Based on this margin, a feature space consisting of 10 consecutive frames was constructed. A total of 33 features were extracted to facilitate point tracking on the images. These features include the following:

  • The first feature quantifies the extent of change in the points to be tracked, calculated as the difference between the tracked frames. The maximum value for this feature is 10, as it incorporates information from frames up to 10 frames prior.

  • The next two features represent the X and Y coordinates of the tracked points within a coordinate system, with the midpoint of the image designated as the center.

  • The subsequent 10 features correspond to the extent of rotation change of the previous 10 frames, determined through homography estimation (Luo et al., 2023).

  • The following 20 features represent the X and Y values of translation changes from the previous 10 frames, as determined through homography estimation (Luo et al., 2023).

Figure 3 displays the feature-importance distribution of the feature-tracking model developed using XGBoost. The most influential features are the X and Y pixel coordinates of the tracked points, whereas the importance of the remaining features progressively decreases across the subsequent 10 frames. Features corresponding to frames further in the past exhibit the lowest importance.

Feature-importance graph of the feature-tracking model developed using the XGBoost algorithm
FIGURE 3

Feature-importance graph of the feature-tracking model developed using the XGBoost algorithm

Here, we apply the following:

χ=F(ρ,τ)7

Φ=H•χ8

κ=Φ/ϕ9

The symbol F in Equation (7) denotes the 3 × 4 homography matrix utilized for detecting changes in the image's boundary corners. To achieve this, a 4 × 2 matrix representing the tracked points, τ, is expanded with a 4 × 1 vector of ones; the transpose is then taken to obtain a matrix with the same dimensions as χ. The 3 × 4 matrix Φ in Equation (8) results from multiplication of the homography matrix H(3×3) and the matrix χ. In Equation (9), the 3 × 4 matrix κ is transposed, and its last column is removed. The symbol ϕ represents the previous row of the matrix Φ, yielding a 4 × 2 matrix that indicates the new positions of the latest image relative to the previous image, with translation and rotation changes derived from the new and old corner and center points. Figure 4 presents an example of homography estimation along with the associated point correspondences between the matched images.

The machine-learning model employed is XGBoost (Chen et al., 2015), selected because of the lack of a discernible pattern in the changes within the sample space of the data set, as well as XGBoost's reduced sensitivity to fluctuations. XGBoost is computationally efficient and facilitates rapid training on large data sets. The model outputs the coordinates of points in the current frame (two values in the X and Y planes) by estimating the changes in pixel coordinates through supervised learning. Prior to actual training, the model parameters were optimized through 100 trials using the Optuna (Akiba et al., 2019) library. The training was conducted using the XGBoost (Chen et al., 2015) and Scikit-Learn (Kramer & Kramer, 2016) libraries. The optimal model parameters and the search spaces for real numbers utilized in parameter optimization are presented in Table 1.

Illustration of homography estimation and corresponding point alignment between matched images
FIGURE 4

Illustration of homography estimation and corresponding point alignment between matched images

View this table:
TABLE 1 Optimal Parameter Values and Search Spaces for the Machine-Learning Model

2.2.5 Navigation of a High-Speed Aerial Vehicle

In the problem addressed in this article, the navigation of a high-speed UAV is defined as reaching the target point in the reference satellite image via the optimal route. The navigation values consist of the pixel coordinates of the target in the camera image. Achieving high-accuracy navigation is challenging when relying solely on a two-dimensional image plane without depth information for a three-dimensional moving vehicle. Consequently, direction is determined based on the kinematic capacity of the UAV.

Under normal operating conditions, as the vehicle approaches a target, it continuously decreases altitude, resulting in a corresponding reduction in scaling relative to the previous frame. Under this assumption, the diagonal points of the frames in Section 2.2.4 will lie within the previous diagonal points, as shown in Equation (10):

PDt⊂PDt−110

Here, the set of corner points for the current frame is denoted as PDt, whereas the corner points of the previous frame are represented as PDt−1. The forward orientation vector (in the X direction) of the vehicle in the XYZ coordinate system is calculated in pixels by averaging the differences between the corner points in sets PDt and PDt−1. Translation and rotation information derived from the homography matrix (Luo et al., 2023) is utilized for the Y and Z directions, resulting in pixel unit values within the XYZ coordinate system.

Figure 5 illustrates the Euler angles, with theta θ (pitch) and psi Ψ (yaw) being critical for navigation and guidance. In this article, the θ and Ψ angles are employed for navigation and rescaling with delayed components. These delayed components are essential for guidance, as the tracking solution experiences delays despite operating in real time (runtime of less than 30 ms). Here, we have the following:

Demonstration of Euler angles on an aerial vehicle
FIGURE 5

Demonstration of Euler angles on an aerial vehicle

tan⁡(Δθ)=(Pzt−Pzt−1)(Pxt−Pxt−1)2+(Pyt−Pyt−1)211

tan⁡(ΔΨ)=Pyt−Pyt−1Pxt−Pxt−112

Equations (11) and (12) describe the conversion of Cartesian coordinates (XYZ) to Euler angles (θ and Ψ). The results represent the changes between frames at times t and t – 1.

2.2.6 Data Set Preparation

Simulation environments were prepared using a game engine (Epic Games, 2019). Two images were generated from the Unreal game engine's maps, while two real-world images were obtained from Google Earth (Gorelick et al., 2017). High-speed UAV platforms were integrated with OpenRocket (Niskanen, 2013) and AirSim (Shah et al., 2018) software. A navigation solution was implemented with five target points under time-varying environmental conditions across four environments, as shown in Figure 8. The data set consists of 800 videos, each 25 s long, at 30 fps, with a resolution of 1920 × 1080 and a 50° FOV angle. Features in the videos were labeled in the simulation environment, enabling robust feature tracking through a learning-based machine-learning model. A key distinction of this data set is the real-time input of the navigation solution for route correction. The UAV's maximum capacities include an altitude of 8000 m and a ground speed of 500 m/s. The equations necessary for the UAV simulation are presented below. The workflow for preparing the simulation data is illustrated in Figure 6.

Workflow for preparing simulation data A three-dimensional virtual environment is constructed using Unreal Engine (Epic Games, 2019), in which simulated UAV camera images are generated. These images are streamed in real time to the onboard processor via AirSim (Shah et al., 2018). The device hosts the reference image database, image-matching models, and feature-tracking modules, thereby enabling fully ondevice processing and navigation support.
FIGURE 6

Workflow for preparing simulation data

A three-dimensional virtual environment is constructed using Unreal Engine (Epic Games, 2019), in which simulated UAV camera images are generated. These images are streamed in real time to the onboard processor via AirSim (Shah et al., 2018). The device hosts the reference image database, image-matching models, and feature-tracking modules, thereby enabling fully ondevice processing and navigation support.

Various forces act on a UAV, including gravity, thrust, friction, and normal force. The normal force varies based on the UAV's angle of attack, air friction, and thrust duration. Changes in normal force influence the UAV's pitch, yaw, and roll moments. Figure 7 illustrates the forces and axes for a representative UAV.

The pitch, yaw, and roll directions of a UAV are influenced by the angle of attack α, gravity G, motor thrust T, drag D, and normal force N. This image was sourced from Niskanen (2013).
FIGURE 7

The pitch, yaw, and roll directions of a UAV are influenced by the angle of attack α, gravity G, motor thrust T, drag D, and normal force N. This image was sourced from Niskanen (2013).

The optimal and actual trajectories shown in Figure 8 are generated from the time step at which the target first enters the camera FOV and navigation is initiated. For each camera frame in which the target is visible, the optimal trajectory is determined by navigating based on the ground truth position of the target. The trajectory-based navigation error is computed as the difference between corresponding steps of the optimal and actual trajectories, whereas the target-based navigation error is defined as the difference between the target position at the end of the flight and the UAV landing point.

Simulation sample images The top row presents artificially prepared environments captured under ideal weather conditions. The second row displays the cloudy states of these artificially prepared environments. The third row showcases sample route outputs under ideal conditions, while the fourth row presents sample route outputs under windy conditions. The last row depicts the optimal and actual trajectories for the corresponding environments. The optimal trajectory is marked by black dots, whereas the actual trajectory is shown as a green line.
FIGURE 8

Simulation sample images

The top row presents artificially prepared environments captured under ideal weather conditions. The second row displays the cloudy states of these artificially prepared environments. The third row showcases sample route outputs under ideal conditions, while the fourth row presents sample route outputs under windy conditions. The last row depicts the optimal and actual trajectories for the corresponding environments. The optimal trajectory is marked by black dots, whereas the actual trajectory is shown as a green line.

We calculate the following:

CN=NqAref13

where N is the normal force, q is the dynamic pressure, and Aref is the characteristic area of a cylinder body. We obtain the following:

CN=KAplan Arefsin2⁡α14

where K is a constant with an approximate value of K≈1.1,α represents the angle of attack, and Aplan denotes the planform area of a cylinder. For small changes in the angle of attack, sin2⁡α≈α2. By combining Equations (13) and (14), the normal force (N) can be derived (Niskanen, 2013):

m(x)=xN(x)15

where m(x) denotes the pitch moment along the X axis, through the center of the air vehicle to its nose (Niskanen, 2013):

DqAwet=11.5ln⁡(R)−5.6216

Here, D represents the drag force, Awet denotes the surface area interacting with the working fluid, and R is the Reynolds number (Mosetti, 1984). Consequently, a UAV can be simulated at supersonic speeds using these equations.

3 EXPERIMENTAL RESULTS

This section compares the results of the method developed in this article with those of existing methods, as well as navigation outcomes under different environmental conditions in the applied simulations. More than 200 distinct simulations were conducted for each environment designed in the game engine, with variations in conditions such as weather and brightness. Results were derived from these simulations.

To compare the tracking algorithms within a specific frame with point-based tracking algorithms, such as the innovative robust feature-tracking method developed in this study, a score metric (Güven & Yetik, 2024) was designed to evaluate the performance of both methods, as shown in Equation (17). The running times of the algorithms were also compared. A minimum bounding box containing points identified by point-based tracking was defined, ensuring that at least two corners corresponded to tracked points for error calculations. To accurately assess the error, 30 initial images were randomly selected from the tested videos, followed by tracking of the subsequent 10 images:

S=∑n=1N|α(n)⊝β(n)|17

In Equation (17), α(n) denotes the real and labeled pixel coordinates of the points to be tracked, while β(n) represents the points identified by the algorithms. The ⊝ symbol denotes pixel-wise subtraction. The limit value N indicates the total number of points tracked in the video images. The scoring metric, which reflects the algorithm performance, is the sum of the absolute differences between α(n) and β(n) (Güven & Yetik, 2024). Table 2 presents the average errors and cycle times obtained from simulation images on the same device (Nvidia Jetson AGX Xavier) for each model.

View this table:
TABLE 2

Comparison of Image-Matching Deep-Learning Models

ALIKED: a lighter keypoint and descriptor extraction network via deformable transformation; DISK: discrete keypoints

In Table 2, the average calculated error represents the average of the relative distance errors between the detected location of the target point in the simulation environment and the actual location of the target point. RIO denotes the reference image optimization method developed in this paper, whereas FAD refers to the feature anomaly detection method introduced herein. This table compares error values with and without the application of the FAD and RIO methods. The optimal solution for overall navigation of a UAV is achieved with SuperPoint and LightGlue (Xu et al., 2024) using the RIO and FAD models.

In Figure 9, the proposed method demonstrates performance comparable to state-of-the-art (SOTA) models while maintaining real-time processing times. The SOTA models operated in offline mode, processing frames sequentially after the inference time of the models, and could not achieve real-time performance. The results are detailed in Table 3.

Test images (Balloon, 2015) for comparing the developed method with SOTA models In the first row, the ground truths of the features are shown as red dots in the first, third, fifth, seventh, and tenth images obtained after image matching. The second row provides an offline display of the feature-tracking outputs of the TAPTRv2 (tracking any point with transformers) (Li et al., 2024) model in the corresponding images. The third row provides an offline display of the feature-tracking outputs of the CoTracker (Karaev et al., 2024) model in the relevant images. Finally, the last row provides the outputs of the method developed in this article.
FIGURE 9

Test images (Balloon, 2015) for comparing the developed method with SOTA models

In the first row, the ground truths of the features are shown as red dots in the first, third, fifth, seventh, and tenth images obtained after image matching. The second row provides an offline display of the feature-tracking outputs of the TAPTRv2 (tracking any point with transformers) (Li et al., 2024) model in the corresponding images. The third row provides an offline display of the feature-tracking outputs of the CoTracker (Karaev et al., 2024) model in the relevant images. Finally, the last row provides the outputs of the method developed in this article.

View this table:
TABLE 3 Visual Feature-Tracking Results of Custom Algorithms, Deep-Learning Models, and the Proposed Model MSE: mean square error

The results in Table 3 indicate that the proposed method does not exhibit the best tracking performance based on pixel error, but is very close to the best-performing method with a considerably lower computation time, enabling real-time operation, unlike existing methods. Thus, the proposed method is the most suitable model for the problem addressed in this article, owing to its emphasis on minimal error and

real-time processing. The proposed model demonstrates significant performance improvements in inference time and competitive results in pixel errors. All models were tested on the Nvidia Jetson AGX Xavier module, with the pixel error percentage parameter (δP) set at 10 pixels. In Table 3, p-values of less than 0.05 provide statistically significant evidence to reject the null hypothesis and support the conclusion that the mean of the error values obtained from the generated scenarios is lower than the parameter (δP).

In Table 4, device running-time performance results are presented in milliseconds. For this problem, the hardware must be an internet of things (IoT) device owing to constraints related to size and low power consumption.

View this table:
TABLE 4

Device Performance Comparison in Milliseconds

*: unable to yield results

All results reported in Table 5 were obtained using an Nvidia Jetson AGX Xavier platform. For the SLAM-based methods, it is assumed that both the initial pose and the camera intrinsic parameters are known a priori. In contrast, the proposed method does not require any prior parameter initialization beyond the provision of reference images.

View this table:
TABLE 5

Navigation Results of SOTA SLAM Models and the Proposed Model

DPVS: deep patch visual SLAM; DROID: differentiable recurrent optimization-inspired design

Table 5 presents a comparison between SOTA SLAM approaches and the method proposed in this study. The proposed method demonstrates superior performance relative to conventional SLAM models. As SLAM algorithms rely on sequential frame matching and feature tracking, their performance is highly dependent on the availability of distinctive visual features. In the scenario considered in this work, the imagery predominantly consists of low-texture background regions, leading to increased feature mismatches and tracking failures and consequently higher error rates.

4 CONCLUSION

This paper addressed the challenge of providing navigation to a target point within a referenced satellite image for a continuously operating high-speed UAV. Producing a stable, robust, real-time solution is challenging, given the real-world constraints of the problem. Previous studies have focused on navigation using cameras in relatively slow UAVs and aerial vehicles capable of hovering. Similar concepts have been applied in mapping studies involving fixed-wing unmanned or manned aerial vehicles. However, this study distinguishes itself by addressing high-speed UAVs with limited maneuverability and offering innovative robust visual feature tracking. Existing camera solutions in the literature typically require secondary sensors, such as IMUs. The innovation presented in this article enables navigation using only cameras, enhancing the feasibility of the approach.

SLAM models were evaluated and found to be insufficient for the target problem when used in isolation. Their strong dependence on accurate initial pose estimates and precise intrinsic camera parameters necessitates a hybrid approach. In this study, a novel method was proposed that incorporates GNSS uncertainty into the initialization process and enables a comparative analysis between SLAM-based estimates and simulated GNSS-derived initial positions.

The results demonstrate that the proposed method achieves improved performance relative to standalone SLAM approaches. The problem is addressed with a careful consideration of cost, system performance, and real-time operational constraints. Furthermore, a key contribution of this work is the development of a framework that maintains reliable operation in GNSS-denied environments, overcoming a critical limitation of conventional SLAM systems that rely on GNSS support for robust initialization.

The preparation of the data set illustrating the problem is also an important part of this article. A game engine motor was utilized to generate the data set, resulting in realistic images. Experimental results indicated that the employed method yielded superior outcomes compared with other approaches, facilitating navigation to the specified target point with acceptable errors. An online hardware environment was established and tested for the problem's applicability, with suitable hardware identified, listed, and compared against alternatives.

Simulations were conducted under constraints in which the UAV altitude did not exceed 8,000 m and the ground speed was limited to 500 m/s. Reference images, along with target coordinate information, matching features, and feature descriptors, were preloaded onto the onboard device, thereby reducing the runtime during flight.

The Nvidia Jetson AGX Xavier was determined to be the most suitable hardware with regard to size, weight, and performance for image-matching and visual feature-tracking operations. The image-matching model was deemed unsuitable on non-Nvidia cards because the expected maximum latency was exceeded. The innovative aspects of the method developed in this article include finding appropriate reference images through an autonomous optimization algorithm, comparing image-matching models on a problem-specific basis, and ensuring parallel communication among models, ultimately providing a novel navigation solution using only three autonomously determined reference satellite images.

4.1 Future Work

The errors encountered in this study can be attributed to imprecise predictions of camera movements and angular, perspective, and environmental differences between the reference image and the images captured by the UAV camera. As the UAV transitions from high altitude to ground level, target detection in the images was deemed acceptable, with an error of 10 pixels. Future studies will utilize depth information from reference satellite images to minimize this error, leveraging the landforms and geographical heights of the region. While targets are considered fixed in the image-based navigation solution presented herein, alternative solutions using laser and radar do not require fixed targets. Consequently, future methods have the potential to also detect and track objects moving in a specific pattern around an initially fixed target, leading to the creation of a model with online training independent of traditional object detection model training.

HOW TO CITE THIS ARTICLE:

Güven, A., & Yetik, I.S. (2026). Real-time UAV navigation using machine learning, satellite images, and continuously operating high-speed cameras with enhanced image matching and robust feature tracking. NAVIGATION, 73. https://doi.org/10.33012/navi.775

This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited.

REFERENCES

  1. ↵
    Ahmine, Y., Caron, G., Mouaddib, E. M., & Chouireb, F. (2019, August). Adaptive Lucas-Kanade tracking. Image and Vision Computing, 88, 1–8. https://doi.org/10.1016/j.imavis.2019.04.004
    Google ScholarOpenURL
  2. ↵
    Akiba, T., Sano, S., Yanase, T., Ohta, T., & Koyama, M. (2019, July). Optuna: A next-generation hyperparameter optimization framework. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (pp. 2623–2631). https://doi.org/10.1145/3292500.3330701
    Google Scholar
  3. ↵
    Amiri, A. J., & Moradi, H. (2016, October). Real-time video stabilization and mosaicking for monitoring and surveillance. In 2016 4th International Conference on Robotics and Mechatronics (ICROM) (pp. 613–618). https://doi.org/10.1109/ICRoM.2016.7886813
    Google Scholar
  4. ↵
    Campos, C., Elvira, R., Gomez, J. J., Montiel, J. M. M., & Tardós, J. D. (2021). ORB-SLAM3: An accurate open-source library for visual, visual-inertial and multi-map SLAM. IEEE Transactions on Robotics, 37(6), 1874–1890.
    Google ScholarOpenURL
  5. ↵
    Cavalli, L., Larsson, V., Oswald, M. R., Sattler, T., & Pollefeys, M. (2020, August). Handcrafted outlier detection revisited. In European Conference on Computer Vision (pp. 770–787). https://doi.org/10.1007/978-3-030-58529-7_45
    Google Scholar
  6. ↵
    Chan, S., Zhou, X., & Chen, S. (2017). Robust adaptive fusion tracking based on complex cells and keypoints. IEEE Access, 5, 20985–21001.
    Google ScholarOpenURL
  7. ↵
    Chen, T., He, T., Benesty, M., Khotilovich, V., Tang, Y., Cho, H., Chen, K., Mitchell, R., Cano, I., Zhou, T., Li, M., Xie, J., Lin, M., Geng, Y., Li, Y., Yuan, J., & Cortes, D. (2015, September). Xgboost: Extreme gradient boosting. R package version 0.4-2, 1(4), 1–4. https://doi.org/10.32614/cran.package.xgboost
    Google ScholarOpenURL
  8. ↵
    Edstedt, J., Bökman, G., & Zhao, Z. (2024, June). DeDoDe v2: Analyzing and improving the DeDoDe keypoint detector. In IEEE/CVF Computer Society Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) (pp. 4245–4253). https://doi.org/10.1109/CVPRW63382.2024.00428
    Google Scholar
  9. ↵
    Epic Games. (2019, April 25). Unreal engine (Version 4.22.1). https://www.unrealengine.com
    Google Scholar
  10. ↵
    Fiaz, M., Mahmood, A., Javed, S., & Jung, S. K. (2019). Handcrafted and deep trackers: Recent visual object tracking approaches and trends. ACM Computing Surveys (CSUR), 52(2), 1–44.
    Google ScholarOpenURL
  11. ↵
    Gorelick, N., Hancher, M., Dixon, M., Ilyushchenko, S., Thau, D., & Moore, R. (2017, December). Google earth engine: Planetary-scale geospatial analysis for everyone. Remote Sensing of Environment, 202, 18–27. https://doi.org/10.1016/j.rse.2017.06.031
    Google ScholarOpenURLCrossRef
  12. ↵
    Güven, A., & Yetik, İ. Ş. (2024, May). Robust point tracking based on image matching and machine learning on video images taken from fast moving camera. In 2024 32nd Signal Processing and Communications Applications Conference (SIU) (pp. 1–4). https://doi.org/10.1109/SIU61531.2024.10601068
    Google Scholar
  13. ↵
    Held, D., Thrun, S., & Savarese, S. (2016, August). Learning to track at 100 fps with deep regression networks. In Computer Vision–ECCV 2016: 14th European Conference (pp. 749–765). https://doi.org/10.48550/arXiv.1604.01802
    Google Scholar
  14. ↵
    High Altitude Balloon. (2015, May). Sub orbital: High altitude weather balloon [Video]. Retrieved November 11, 2024 from YouTube. https://www.youtube.com/watch?v=GD_F4dy_Bfc
    Google Scholar
  15. ↵
    Joshi, P. (2015). OpenCV with Python by example. Packt Publishing Ltd.
    Google Scholar
  16. ↵
    Karaev, N., Rocco, I., Graham, B., Neverova, N., Vedaldi, A., & Rupprecht, C. (2024, September). Cotracker: It is better to track together. In European conference on computer vision (pp. 18–35). Cham: Springer Nature Switzerland. https://doi.org/10.48550/arXiv.2307.07635
    Google Scholar
  17. ↵
    Kinnari, J., Renzulli, R., Verdoja, F., & Kyrki, V. (2023). LSVL: Large-scale season-invariant visual localization for UAVs. Robotics and Autonomous Systems, 168, 104497. https://doi.org/10.1016/j.robot.2023.104497
    Google ScholarOpenURL
  18. ↵
    Kramer, O., & Kramer, O. (2016, May). Scikit-learn. Machine Learning for Evolution Strategies, 20, 45–53. https://doi.org/10.1007/978-3-319-33383-0_5
    Google ScholarOpenURL
  19. ↵
    Li, H., Zhang, H., Liu, S., Zeng, Z., Li, F., Ren, T., Li, B., & Zhang, L. (2024, December). TAPTRv2: Attention-based position update improves tracking any point. Advances in Neural Information Processing Systems, 37, 101074–101095, https://doi.org/10.52202/079017-3205
    Google ScholarOpenURL
  20. ↵
    Lipson, L., Teed, Z., & Deng, J. (2024, October). Deep patch visual SLAM. In European Conference on Computer Vision (pp. 424–440). https://doi.org/10.1007/978-3-031-72627-9_24
    Google Scholar
  21. ↵
    Liu, T., & Liu, Y. (2021). Moving camera-based object tracking using adaptive ground plane estimation and constrained multiple kernels. Journal of Advanced Transportation, 2021(1), 8153474.
    Google ScholarOpenURL
  22. ↵
    Luo, X.-l., Lv, J.-h., & Sun, G. (2020). A visual-inertial navigation method for high-speed unmanned aerial vehicles. arXiv. https://arxiv.org/abs/2002.04791
    Google Scholar
  23. ↵
    Luo, Y., Wang, X., Liao, Y., Fu, Q., Shu, C., Wu, Y., & He, Y. (2023). A review of homography estimation: Advances and challenges. Electronics, 12(24), 4977.
    Google ScholarOpenURL
  24. ↵
    Mishkin, D., Radenovic, F., & Matas, J. (2018, September). Repeatability is not enough: learning discriminative affine regions via discriminability. In Proceedings of the European Conference on Computer Vision (ECCV) (pp. 284–300). https://doi.org/10.48550/arXiv.1711.06704
    Google Scholar
  25. ↵
    Mosetti, F. (1984). Reynold’s number. In Beaches and coastal geology (pp. 686–688). Springer US. https://doi.org/10.1007/0-387-30843-1_366
    Google Scholar
  26. ↵
    Moukthika. (2025). Resizing and rescaling images with OpenCV. Retrieved April 15, 2025, from https://opencv.org/blog/resizing-and-rescaling-images-with-opencv/
    Google Scholar
  27. ↵
    Navard, P., & Yilmaz, A. (2024). A probabilistic-based drift correction module for visual inertial slams. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, XLVIII-2–2024, 297–300. https://doi.org/10.5194/isprs-archivesxlviii-2-2024-297-2024
    Google Scholar
  28. ↵
    Niskanen, S. (2013). Openrocket technical documentation. Development of an Open Source model rocket simulation software, 11–13.
    Google Scholar
  29. ↵
    Potje, G., Cadar, F., Araujo, A., Martins, R., & Nascimento, E. R. (2024, June). Xfeat: Accelerated features for lightweight image matching. In 2024 IEEE / CVF Computer Vision and Pattern Recognition (CVPR) (pp. 2682–2691). https://doi.org/10.48550/arXiv.2404.19174
    Google Scholar
  30. ↵
    Raguram, R., Frahm, J.-M., & Pollefeys, M. (2008). A comparative analysis of ransac techniques leading to adaptive real-time random sample consensus. In D. Forsyth, P. Torr, & A. Zisserman (Eds.), Computer vision – eccv 2008 (pp. 500–513). Springer Berlin Heidelberg.
    Google Scholar
  31. ↵
    Shah, S., Dey, D., Lovett, C., & Kapoor, A. (2018, January). Airsim: High-fidelity visual and physical simulation for autonomous vehicles. In Field and Service Robotics: Results of the 11th International Conference (pp. 621–635). https://doi.org/10.1007/978-3-319-67361-5_40
    Google Scholar
  32. ↵
    Sun, X., Zhang, C., Zou, L., & Li, S. (2023, November). Real-time optimal states estimation with inertial and delayed visual measurements for unmanned aerial vehicles. Sensors, 23(22), 1–20. https://doi.org/10.3390/s23229074
    Google ScholarOpenURL
  33. ↵
    Tareen, S. A. K., & Saleem, Z. (2018, March). A comparative analysis of sift, surf, kaze, akaze, orb, and brisk. In 2018 International Conference on Computing, Mathematics and Engineering Technologies (iCoMET) (pp. 1–10). https://doi.org/10.1109/ICOMET.2018.8346440
    Google Scholar
  34. ↵
    Teed, Z., & Deng, J. (2021, August). DROID-SLAM: Deep visual SLAM for monocular, stereo, and RGB-D cameras. Advances in Neural Information Processing Systems, 34, 16558–16569. https://doi.org/10.48550/arXiv.2108.10869
    Google ScholarOpenURL
  35. ↵
    Van Rossum, G., & Drake, F. L. (2009). Python 3 reference manual. CreateSpace.
    Google Scholar
  36. ↵
    Wojke, N., Bewley, A., & Paulus, D. (2017, September). Simple online and realtime tracking with a deep association metric. In 2017 IEEE International Conference on Image Processing (ICIP) (pp. 3645–3649). https://doi.org/10.1109/ICIP.2017.8296962
    Google Scholar
  37. ↵
    Wu, J., Cao, J., Song, L., Wang, Y., Yang, M., & Yuan, J. (2021, June). Track to detect and segment: An online multi-object tracker. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 12352–12361). https://doi.org/10.48550/arXiv.2103.08808
    Google Scholar
  38. ↵
    Xu, S., Chen, S., Xu, R., Wang, C., Lu, P., & Guo, L. (2024, March). Local feature matching using deep learning: A survey. Information Fusion, 107, 102344. https://doi.org/10.1016/j.inffus.2024.102344
    Google ScholarOpenURL
  39. ↵
    Zhang, H., He, H., Zhou, Y., & Lei, F. (2025, October). Efficient image matching for UAV visual navigation via dalglue. Scientific Reports, 15, 1–13. https://doi.org/10.1038/s41598-025-21602-5
    Google ScholarOpenURL
  40. ↵
    Zhang, X., He, Z., Ma, Z., Wang, Z., & Wang, L. (2021, November). Life: A novel learning local features extraction for UAV navigation based on infrared aerial image and satellite reference image matching. Remote Sensing, 13(22), 1–24. https://doi.org/10.3390/rs13224618
    Google ScholarOpenURL
  41. ↵
    Zhao, C., Wu, D., He, J., & Dai, C. (2023, August). A visual positioning method of UAV in a largescale outdoor environment. Sensors, 23(15), 1–16. https://doi.org/10.3390/s23156941
    Google ScholarOpenURL
  42. ↵
    Zhao, X., Wu, X., Chen, W., Chen, P. C. Y., Xu, Q., & Li, Z. (2023, April). Aliked: A lighter keypoint and descriptor extraction network via deformable transformation. IEEE Transactions on Instrumentation Measurement, 72, 1–16. https://doi.org/10.1109/TIM.2023.3271000
    Google ScholarOpenURL
Loading
Loading
Loading
Loading
PDF
Back to top

In this Issue

NAVIGATION: Journal of the Institute of Navigation

NAVIGATION: Journal of the Institute of Navigation

Vol. 73

June 2026

Table of Contents Index by author
  • Share

    Share This Article

    Real-Time UAV Navigation Using Machine Learning, Satellite Images, and Continuously Operating High-Speed Cameras with Enhanced Image Matching and Robust Feature Tracking
    • Institute of Navigation
    • Jun 2026,
    • 73
    • navi.775;
    • DOI: 10.33012/navi.775
    Twitter logoFacebook logoMendeley logo
  • Bookmark this Article

    SAVE ITEM IN SELECTED FOLDER.

Jump to section

Keywords

feature extraction, image homography and pose estimation, image matching, image-point tracking, machine learning, unmanned aerial vehicle navigation