Georeferenced UAV Localization in Mountainous Terrain Under GNSS-Denied Conditions

Abstract

In Global Navigation Satellite System (GNSS)-denied environments, unmanned aerial vehicles (UAVs) relying on Vision-Based Navigation (VBN) in high-altitude, mountainous terrain face severe challenges due to geometric distortions in aerial imagery. This paper proposes a georeferenced localization framework that integrates orthorectified aerial imagery with Scene Matching (SM) to achieve robust positioning. The method employs a camera projection model combined with Digital Elevation Model (DEM) to orthorectify UAV images, thereby mitigating distortions from central projection and terrain relief. Pre-processing steps enhance consistency with reference orthophoto maps, after which template matching is performed using normalized cross-correlation (NCC). Sensor fusion is achieved through extended Kalman filters (EKFs) incorporating Inertial Navigation System (INS), GNSS (when available), barometric altimeter, and SM outputs. The framework was validated through flight tests with an aircraft over 45 km trajectories at altitudes of 2.5 km and 3.5 km in mountainous terrain. The results demonstrate that orthorectification improves image similarity and significantly reduces localization error, yielding lower 2D RMSE compared to conventional rectification. The proposed approach enhances VBN by mitigating terrain-induced distortions, providing a practical solution for UAV localization in GNSS-denied scenarios.

1. Introduction

Unmanned aerial vehicles (UAVs) are pivotal in applications such as exploration, disaster response, and surveillance [1]. Traditionally, an integrated navigation system that combines Inertial Navigation System (INS) and Global Navigation Satellite System (GNSS) has provided reliable positioning for UAVs. However, GNSS is vulnerable to signal blockage and interference, leading to GNSS-denied environments such as urban canyons, indoor settings, or areas with adversarial jamming [2,3]. In such scenarios, INS errors accumulate rapidly, leading to significant positional drift and necessitating robust alternative navigation techniques to ensure stable and accurate localization.

Vision-Based Navigation (VBN) has emerged as a promising approach for GNSS-denied environments, driven by advances in camera miniaturization, image sensor technology, computing technology, and image processing algorithms [4]. VBN techniques can be categorized into three primary approaches based on their use of maps: Visual Odometry (VO), which incrementally estimates motion without relying on maps; Visual Simultaneous Localization and Mapping (VSLAM), which concurrently builds and uses environmental maps for localization; and Scene Matching (SM), which determines position by comparing current images with pre-built maps [5].

VO estimates UAV position and orientation in real time by analyzing visual data from consecutive image frames. To improve robustness, VO is often integrated with Inertial Measurement Unit (IMU) data, a technique referred to as Visual Inertial Odometry (VIO). Due to its cost-effectiveness, VO has been widely adopted in commercial UAVs [6,7]. However, VO relies on relative motion estimation through techniques such as optical flow, leading to inevitable error accumulation over time [8]. VSLAM mitigates this by simultaneously estimating position and constructing maps, using techniques like loop closure and bundle adjustment to correct errors [9,10]. Nevertheless, VSLAM faces challenges in expansive outdoor environments where loop closure opportunities are limited, resulting in gradual error accumulation.

In contrast, SM provides absolute positioning by comparing aerial images with pre-built map data—such as satellite imagery, orthophoto maps, or 3D models—thereby avoiding cumulative errors inherent in VO and VSLAM [11]. By leveraging absolute coordinates, SM maintains consistent global positioning when high-quality map data is available. Its robustness in texture-rich environments and compatibility with wide fields of view at high altitudes make it particularly effective for GNSS-denied navigation in outdoor UAV operations.

The most challenging aspect of SM-based localization is the need to match heterogeneous images captured under varying conditions, such as different viewpoints, illumination, sensors, and resolutions. Previous studies have explored several approaches to address these difficulties. Traditional template matching methods have employed diverse similarity metrics between reference images and aerial imagery. Conte et al. [12] proposed a VBN architecture that integrates inertial sensors, VO, and image registration to a georeferenced image. Their system, using normalized cross-correlation (NCC) as the matching metric, reported a maximum positional error of 8 m in a 1 km closed-loop trajectory at 60 m altitude. Yol et al. [13] applied Mutual Information (MI)-based matching, achieving RMSE values of 6.56 m (latitude), 8.02 m (longitude), and 7.44 m (altitude) in a 695 m flight test at 150 m altitude. Sim et al. [14] proposed an integrated system that estimates aircraft position and velocity using sequential aerial images. They employed Robust Oriented Hausdorff Measure (ROHM) for the image matching, and flight tests with helicopters and aircraft at altitudes up to 1.8 km and distances up to 124 km showed errors on the order of hundreds of meters. Wan et al. [15] introduced an illumination-invariant Phase Correlation (PC) method to match aerial images with reference satellite imagery. In their study, aerial images captured by a UAV flying at an average altitude of 350 m were roughly rectified to a nadir view, and the average positioning errors were reported as 32.2 m along the x-axis and 32.46 m along the y-axis.

Recent advances in UAV localization under GNSS-denied conditions have introduced deep learning and feature-matching techniques. To handle differences in viewpoint and seasonal appearance, Goforth et al. [16] employed CNN features trained on satellite data with temporal optimization, achieving an average localization error of less than 8 m for an 850 m flight at 200 m altitude, but their approach was constrained by limited robustness in diverse terrains. Gao et al. [17] combined SuperPoint [18] and SuperGlue [19] for cross-source registration, reaching sub-pixel accuracy (0.356 m) with satellite imagery of 0.3 m resolution, although their evaluation did not include high-altitude mountainous terrain, leaving applicability in such environments uncertain. Sun et al. [20] proposed the Local Feature Transformer (LoFTR) framework, which leverages Transformers for detector-free dense matching, enabling robust correspondences even in low-texture or repetitive-pattern regions. Nevertheless, its direct applicability to UAV-to-satellite matching in mountainous terrain remains unverified. Hikosaka et al. [21] extracted road networks from aerial and satellite images using a U-Net model followed by template matching, achieving precise alignment in road-rich areas, but their method is unsuitable for mountainous regions with sparse man-made structures and oblique UAV imagery.

As shown in Figure 1, distortions occur in aerial images depending on the camera’s viewpoint and terrain elevation. To improve localization accuracy affected by viewpoint differences when observing terrain, several studies have been conducted. Woo et al. [22] proposed a method for estimating UAV position and orientation by matching oblique views of mountain peaks with Digital Elevation Models (DEMs), but validation was limited to simulations, restricting its generalizability. Kinnari et al. [23] performed orthorectification of UAV images to match them with reference images; however, this approach assumes local planarity of the environment, converting images to a nadir view, which limits its effectiveness in rugged terrains with significant elevation changes, such as mountainous areas. Chiu et al. [24] used a 3D georeferenced model to render reference images in an oblique view similar to the UAV images for matching, reporting RMSE errors of 9.83 m over 38.9 km; however, information about the map and flight altitude was not specified. Ye et al. [25] employed a coarse-to-fine approach for oblique-view images, but their tests were conducted with images captured by a UAV flying at 150 m altitude over a university campus with many buildings, limiting applicability to natural terrains.

In common with prior studies on VBN, our work leverages SM to improve UAV localization accuracy in GNSS-denied environments. However, most existing studies concentrated on low-altitude flights in flat or texture-rich areas and provided only limited analysis of geometric distortions caused by significant terrain variations at high altitudes (above 2 km) [26]. To address these gaps, this study analyzes the impact of terrain-induced geometric distortions on localization accuracy in high-altitude UAV imagery and proposes a VBN architecture using a novel SM method that integrates orthorectification to enhance consistency with reference maps. By incorporating DEM-based orthorectification into the SM process, we extend the applicability of image-based localization methods to challenging mountainous environments that have not been adequately addressed in the literature. The proposed approach is validated through real flight experiments in rugged terrain, demonstrating its effectiveness for stable and accurate absolute navigation.

The main contributions of this work are as follows:

• A VBN architecture and SM technique efficient for high-altitude UAV localization in mountainous terrain.
• Orthorectification of aerial imagery using a projection model and DEM to mitigate geometric distortions, thereby improving matching accuracy with orthophoto maps.
• Validation in real flight experiments over mountainous regions.

The paper is organized as follows: Section 2 details the proposed VBN algorithm. Section 3 presents real-world experiments validating the approach. Section 4 offers discussion, and Section 5 provides concluding remarks.

2. Methods

This section provides an overview of the proposed VBN algorithm for high-altitude UAV position estimation in mountainous terrain, followed by detailed descriptions of its components. The algorithm integrates data from an IMU, a GNSS receiver, a barometric altimeter, and a camera to achieve robust geolocation, particularly in GNSS-denied environments. Orthorectification and an extended Kalman filter (EKF) are employed to mitigate terrain-induced distortions and fuse sensor measurements, respectively.

2.1. Overview

The proposed method combines INS, GNSS, barometric altimetry, and SM to estimate the UAV’s position, velocity, and attitude. Figure 2 presents a block diagram of the algorithm, illustrating the data flow and processing stages. The key components are as follows:

• Input Data: Measurements from the IMU, GNSS, barometric altimeter, and camera, along with reference data including orthophoto maps and DEMs. When GNSS is available, the GNSS navigation solution is used to correct INS errors; otherwise, the SM result is used.
• INS Mechanization: The IMU outputs are processed to compute the UAV’s position, velocity, and attitude using INS mechanization.
• Aerial Scene Matching: Aerial images are compared with the georeferenced orthophoto map to estimate the UAV’s position. Orthorectification is applied to compensate for terrain-induced geometric distortions using DEM data. Aerial images undergo image processing to achieve consistent resolution, rotation, and illumination with georeferenced images.
• Sensor Fusion: The proposed method employs three EKF-based sub-modules for sensor fusion: a 13-state EKF for horizontal navigation errors, attitude errors, and sensor biases; a 2-state EKF for vertical channel (altitude) stabilization; and a 2-state error estimator for the barometric altimeter. These sub-modules collectively produce a corrected navigation solution.

2.2. Aerial Scene Matching

SM is a fundamental technique for estimating the absolute position of UAVs by registering aerial imagery with georeferenced images. This process enables the computation of the UAV’s position. This subsection details the key components of SM, including the role of georeferenced images, the orthorectification for UAV imagery, and image processing techniques using template matching. The flow chart of SM is shown in Figure 3.

2.2.1. Georeferenced Image

Georeferenced images are essential for SM, serving as the reference dataset to which UAV-captured imagery is registered. These images are embedded with geographic coordinates in a standard Coordinate Reference System (CRS), such as WGS84 (EPSG:4326) or Universal Transverse Mercator (UTM), enabling pixel-to-geospatial coordinate mapping. To maximize the SM accuracy for UAV localization, georeferenced images should be ideally in the form of orthophotos, which are orthorectified to eliminate geometric distortions caused by terrain relief, camera tilt, and lens imperfections. This ensures that each pixel in the georeferenced image corresponds precisely to the geographic location of the actual terrain.

2.2.2. Orthorectification

Accurate and precise UAV localization requires transforming UAV-captured images into a form consistent with georeferenced orthophoto maps. Geometric distortions induced by terrain relief are a primary factor degrading the accuracy of SM for UAV position estimation. These distortions arise from the central projection of the camera and the three-dimensional characteristics of the terrain. To mitigate these effects, this subsection proposes an orthorectification technique applied to UAV imagery, transforming it into a form consistent with reference orthophoto maps, thereby enhancing SM performance and minimizing geolocation errors.

The orthorectification process for UAV images is summarized as follows:

• Projection Model Development: A projection model is created using the camera’s intrinsic (e.g., focal length) and extrinsic (e.g., position and attitude) parameters [27]. The camera position and attitude information determining the projection model uses aided navigation information estimated through the sensor fusion framework (Section 2.3).
• Pixel-to-Terrain Projection: The projection path for each pixel in the image sensor is computed based on the projection model.
• Pixel Georeferencing: Using the DEM, the intersection between each pixel’s projection line and the terrain is determined, yielding the actual geographic coordinates corresponding to each image pixel.
• Reprojection and Compensation: The image is reprojected to compensate for terrain relief displacements, aligning each pixel with its actual geographic location.

Orthorectification Using Camera Projection Model

The following describes the detailed procedure for calculating the actual geographic coordinates of terrain points corresponding to each pixel in a UAV-captured image. The method is formulated using the camera projection model and DEM. For clarity, Figure 4 illustrates a cross-sectional view ($y^w = 0$) under the assumption of a pitch angle $\theta = 0$. Boldface symbols are used to represent vectors.

The world coordinate system (W-frame), UAV body coordinate system (B-frame), and camera coordinate system (C-frame) are depicted in Figure 4, where the superscripts of the origin and the x, y, z axes indicate the corresponding frame. For example, ${\mathit{O}}^w$ and $x^w$, $y^w$, $z^w$ denote the origin and the axes of the W-frame, respectively. The rotation matrix ($C_b^w$) from the B-frame to the W-frame, defined by Euler angles, is given as:

$$\begin{array}{c}{C}_b^w=C_z\left(\psi \right)C_x\left(\theta \right)C_y\left(\varphi \right)\hfill \\ =\left[\begin{array}{ccc}\mathrm{c}\mathrm{o}\mathrm{s}\psi & \mathrm{s}\mathrm{i}\mathrm{n}\psi & 0\\ -\mathrm{s}\mathrm{i}\mathrm{n}\psi & \mathrm{c}\mathrm{o}\mathrm{s}\psi & 0\\ 0& 0& 1\end{array}\right]\left[\begin{array}{ccc}1& 0& 0\\ 0& \mathrm{c}\mathrm{o}\mathrm{s}\theta & \mathrm{s}\mathrm{i}\mathrm{n}\theta \\ 0& -\mathrm{s}\mathrm{i}\mathrm{n}\theta & \mathrm{c}\mathrm{o}\mathrm{s}\theta \end{array}\right]\left[\begin{array}{ccc}\mathrm{c}\mathrm{o}\mathrm{s}\varphi & 0& -\mathrm{s}\mathrm{i}\mathrm{n}\varphi \\ 0& 1& 0\\ \mathrm{s}\mathrm{i}\mathrm{n}\varphi & 0& \mathrm{c}\mathrm{o}\mathrm{s}\varphi \end{array}\right]\hfill \\ =\left[\begin{array}{ccc}\mathrm{c}\mathrm{o}\mathrm{s}\psi \mathrm{c}\mathrm{o}\mathrm{s}\varphi -\mathrm{s}\mathrm{i}\mathrm{n}\psi \mathrm{s}\mathrm{i}\mathrm{n}\theta \mathrm{s}\mathrm{i}\mathrm{n}\varphi & \mathrm{c}\mathrm{o}\mathrm{s}\psi \mathrm{c}\mathrm{o}\mathrm{s}\varphi -\mathrm{s}\mathrm{i}\mathrm{n}\psi \mathrm{s}\mathrm{i}\mathrm{n}\theta \mathrm{s}\mathrm{i}\mathrm{n}\varphi & \mathrm{c}\mathrm{o}\mathrm{s}\psi \mathrm{s}\mathrm{i}\mathrm{n}\varphi +\mathrm{s}\mathrm{i}\mathrm{n}\psi \mathrm{s}\mathrm{i}\mathrm{n}\theta \mathrm{c}\mathrm{o}\mathrm{s}\varphi \\ \mathrm{s}\mathrm{i}\mathrm{n}\psi \mathrm{c}\mathrm{o}\mathrm{s}\varphi +\mathrm{c}\mathrm{o}\mathrm{s}\psi \mathrm{s}\mathrm{i}\mathrm{n}\theta \mathrm{s}\mathrm{i}\mathrm{n}\varphi & \mathrm{c}\mathrm{o}\mathrm{s}\psi \mathrm{c}\mathrm{o}\mathrm{s}\theta & \mathrm{s}\mathrm{i}\mathrm{n}\psi \mathrm{s}\mathrm{i}\mathrm{n}\varphi -\mathrm{c}\mathrm{o}\mathrm{s}\psi \mathrm{s}\mathrm{i}\mathrm{n}\theta \mathrm{c}\mathrm{o}\mathrm{s}\varphi \\ -\mathrm{c}\mathrm{o}\mathrm{s}\theta \mathrm{s}\mathrm{i}\mathrm{n}\varphi & \mathrm{s}\mathrm{i}\mathrm{n}\theta & \mathrm{c}\mathrm{o}\mathrm{s}\theta \mathrm{c}\mathrm{o}\mathrm{s}\varphi \end{array}\right]\hfill \end{array}$$

(1)

where $\varphi$, $\theta$, $\psi$ represent roll, pitch, and yaw, respectively.

The rotation matrix ($C_b^c$) from the B-frame to the C-frame is:

$$C_b^c=\left[\begin{array}{ccc}1& 0& 0\\ 0& -1& 0\\ 0& 0& -1\end{array}\right]$$

(2)

Thus, the rotation matrix ($C_c^w$) from the C-frame to the W-frame is obtained as:

$$C_c^w=C_b^w{C}_c^b$$

(3)

The origins of the C-frame and B-frame are assumed to coincide with the focal point of the camera. ${\mathit{O}}^c$ and ${\mathit{O}}^b$ are located at ${(0, 0, h)}^w$ in the W-frame, where $h$ denotes the UAV altitude. The image plane of the camera is positioned at a distance equal to the focal length ($f$) from ${\mathit{O}}^c$, with the camera oriented according to the UAV’s attitude.

Orthorectification reprojects each pixel (index $(i,j)$) in the original UAV image to a corresponding pixel in the orthorectified image, thereby compensating for terrain-induced displacements. The terrain elevation $D\left(\mathit{x}\right)$ is obtained from the DEM.

The coordinates of the $(i,j)$-th pixel in the C-frame are given by:

$${\mathit{x}}_{i,j}^c={\left[\begin{array}{ccc}{p}_i& p_j& f\end{array}\right]}^T, where p_i=\left(i-{\displaystyle \frac{W}{2}}\right)\mu , { p}_j=\left(\mathrm{j}-{\displaystyle \frac{H}{2}}\right)\mu$$

(4)

where $p_i$ and $p_j$ are the image plane coordinates, $W$ and $H$ denote the image width and height in pixels, and μ is the pixel pitch.

The coordinates in the W-frame are obtained by applying the rotation matrix ($C_c^w$) and a translation by altitude:

$${\mathit{x}}_{i,j}^w=C_c^w{\mathit{x}}_{i,j}^c+{\left[\begin{array}{ccc}0& 0& h\end{array}\right]}^T$$

(5)

The projection line for pixel $(i, j)$ passes through ${\mathit{x}}_{i,j}^w$ and ${(0, 0, h)}^w$ in the W-frame. The line equation can be expressed in vector (parametric) form as:

$$\overrightarrow{{\mathit{r}}^w}\left(\mathrm{t}\right)=\left[\begin{array}{c}0\\ 0\\ h\end{array}\right]+\mathrm{t}\left[\begin{array}{c}{\mathit{x}}_{i,j}^w\left(x\right)\\ {\mathit{x}}_{i,j}^w\left(y\right)\\ {\mathit{x}}_{i,j}^w\left(z\right)-h\end{array}\right],\hspace{1em}\mathrm{t}\in \mathbb{R}$$

(6)

where ${\mathit{x}}_{i,j}^w(x)$, ${\mathit{x}}_{i,j}^w(y)$, and ${\mathit{x}}_{i,j}^w(z)$ are the x-, y- and z-components of ${\mathit{x}}_{i,j}^w$, and $t$ is a real-valued parameter.

The intersection ${\mathit{y}}_{i,j}^w$ between the projection line and the terrain elevation $D\left(\mathit{x}\right)$ is computed iteratively, as described in Algorithm 1 and illustrated in Figure 4. The horizontal coordinate of ${\mathit{y}}_{i,j}^w$, denoted ${\mathit{d}}_{i,j}$, represents the actual geographic position corresponding to the $(i, j)$-th pixel in the orthorectified image. Once the terrain intersections for all UAV image pixels are determined using Algorithm 1, the pixels are reprojected onto a regular grid, thereby generating the orthorectified image.

Algorithm 1. Fixed-Point Iteration for Finding Intersections of a 3D Line and a Discretized Data Surface.

Require: 3D line $(x, y, z)$ = $\left(at, bt, h_0+t\left(c-h\right)\right)$ *, discretized data ${\left\{\left(x_i, y_i, z_i\right)\right\}}_{i = 1}^n$, initial height $z_0 = 0$, initial height offset $h$, tolerance $\epsilon$, maximum iterations $N$

Ensure: Intersection point $(x, y, z)$ where $z\approx D(x, y)$ and $(x, y, z)$ lies on the line

$1: \mathrm{Initialize} t\leftarrow {\displaystyle \frac{-h}{c-h}}$$, n\leftarrow 0$         $▷ \mathrm{From} z_0=h+t\left(c-h\right) = 0$

$2: \mathrm{Compute} x_n\leftarrow at$$, y_n\leftarrow bt$

$3: \mathbf{while} n<N$ do

$4: n\leftarrow n+1$

$5: \mathrm{Compute} z_n\leftarrow D(x_n, y_n)$ using 2D interpolation (e.g., bilinear)

$6: \mathrm{Compute} t_{new} \leftarrow {\displaystyle \frac{z_n-h}{c-h}}$

$7: \mathbf{if} \left|t_{new}-t\right|<\epsilon$ then

$8: \mathbf{return} \left(a{t}_{new}, b{t}_{new}, h+t_{new}\left(c-h\right)\right)$

9:               end if

$10: t\leftarrow t_{new}$

$11: \mathrm{Compute} x_n\leftarrow at$$, y_n\leftarrow bt$

12: end while

13: return Failure: Did not converge within N iterations

* The line is defined by the camera image pixel position $(a,b,c)$ and the reference point (0, 0, h) in the W-frame.

Orthorectification Results

The effectiveness of the proposed orthorectification method is demonstrated in Figure 5. A representative case was selected for this demonstration because it includes both moderate oblique viewing angles and significant terrain elevation variations, which clearly highlight the impact of terrain-induced distortions on image registration. In this section, the orthorectification effect on the aerial image is illustrated in detail, while the improvement in localization accuracy of SM is further analyzed in Section 3.2.

Figure 5a shows the original aerial image acquired at an altitude of 3.5 km, with the camera oriented at a roll angle of 5° and a pitch angle of 3°. The corresponding DEM of the imaged terrain is presented in Figure 5b, where the elevation varies within a range of 100–420 m. Figure 5c,d are the rectified image and the orthorectified image, respectively. The rectified image is obtained by applying a homography transformation to align the image plane parallel to the ground surface, effectively converting the oblique view into a nadir view. Figure 5d shows the orthorectified image generated using the method proposed in this study.

A pixel-wise comparison between the rectified and orthorectified images is provided in Figure 5e,f. Figure 5e presents a red–cyan composite of the two images, while Figure 5f illustrates the difference mask, where pixels with an intensity discrepancy exceeding a tolerance of 50 are highlighted in white. The correspondence between the white regions in Figure 5f and areas of higher or lower terrain elevation in the DEM confirms that terrain-induced displacements are strongly correlated with terrain elevation.

Image matching was conducted by registering both the rectified and orthorectified images with the reference orthophoto map (Figure 5g,h). For the rectified image (Figure 5c), template matching using NCC yielded a maximum similarity score of 0.303. In contrast, the orthorectified image (Figure 5d) achieved a score of 0.480, corresponding to an absolute increase of 0.177 in similarity with the reference map. Furthermore, the localization error obtained from template matching was reduced from 15.8 m to 9.5 m. A detailed comparison of the geolocation errors estimated through template matching is provided in Section 3.2. These results demonstrate that the proposed orthorectification method effectively mitigates terrain-induced distortions, thereby improving the accuracy of SM and geolocation in mountainous environments.

2.2.3. Image Processing and Matching

The primary difficulty in SM lies in registering two heterogeneous images with differing characteristics, such as variations in illumination, resolution, and perspective. At high altitudes (e.g., several kilometers), the distinctiveness of keypoints, such as building corners or road intersections, diminishes due to their smaller apparent size, making traditional feature-based methods like Scale-Invariant Feature Transform (SIFT) or Speeded-Up Robust Features (SURF) less effective. In mountainous regions, where distinct geometric structures like roads or buildings are scarce, feature-based matching is particularly challenging. Consequently, template matching, which relies on global pattern comparison, is employed as a more suitable approach for SM in such environments.

To achieve consistent image registration between UAV imagery and georeferenced images, the following pipeline is implemented, as illustrated in Figure 3:

• Illumination Normalization: Histogram equalization or Contrast-Limited Adaptive Histogram Equalization (CLAHE) [28] is applied to both UAV images and georeferenced images to mitigate variations in lighting and contrast, ensuring robustness across diverse environmental conditions.
• Lens Distortion Correction: UAV images are corrected for lens-induced distortions using pre-calibrated intrinsic camera parameters and distortion coefficients, ensuring accurate spatial geometry.
• Resolution Adjustment: To ensure spatial consistency between the UAV image and the georeferenced map, the UAV image is rescaled based on the aided altitude. The scaling factors in the $X$ and $Y$ directions are computed as:

$${\mathrm{Scale}}_X={\mathrm{Scale}}_Y={\displaystyle \frac{H}{f\cdot R_{map}}}$$

(7)

where $H$ is the flight altitude above ground, $R_{map}$ is the reference map resolution, and $f$ is the focal length.

• Orthorectification: Orthorectification is performed to remove geometric distortions in aerial imagery caused by camera position, attitude, and terrain elevation variations, thereby ensuring consistency with the georeferenced orthophoto maps (Section 2.2.2).
• Rotational Alignment: Before template matching, the georeferenced image and the aerial image should be rotationally aligned using the UAV’s attitude data. The orthorectified aerial image shows a ground footprint rotated by the UAV’s heading. Since template matching is carried out by sliding a rectangular template over the georeferenced image, the orthorectified image would be excessively cropped without alignment. To minimize this effect, both the orthorectified image and the georeferenced image are rotated by the UAV’s heading angle, thus achieving rotational alignment. This rotation is implemented using a 2D affine transformation matrix defined as

$$M=\left[\begin{array}{ccc}\mathrm{c}\mathrm{o}\mathrm{s}\theta & \mathrm{s}\mathrm{i}\mathrm{n}\theta & \left(1-\mathrm{c}\mathrm{o}\mathrm{s}\theta \right)\cdot c_x-\mathrm{s}\mathrm{i}\mathrm{n}\theta \cdot c_y\\ -\mathrm{s}\mathrm{i}\mathrm{n}\theta & \mathrm{c}\mathrm{o}\mathrm{s}\theta & \mathrm{s}\mathrm{i}\mathrm{n}\theta \cdot c_x+\left(1-\mathrm{c}\mathrm{o}\mathrm{s}\theta \right)\cdot c_y\end{array}\right]$$

(8)

where $\theta$ denotes the UAV heading angle and $(c_x, c_y)$ is the rotation center.

• Template Matching: Following the preprocessing, template matching is employed to estimate the UAV’s position by correlating the UAV image (template) with a reference image. The similarity between the template $T$ and a region of the reference image $I$is measured using NCC [29], which is known for robustness against linear illumination variations. NCC computes the normalized correlation coefficient between the template and a sliding window in the reference image.

The NCC at position $(x,y)$ in the reference image is defined as:

$$\mathrm{N}\mathrm{C}\mathrm{C}\left(x,y\right)={\displaystyle \frac{{\sum }_{i,j}\left[\left(I\left(x+i, y+j\right)-{\overline{I}}_{x,y}\right)·\left(T\left(i,j\right)-\overline{T}\right)\right]}{\sqrt{{\sum }_{i,j}{\left(I\left(x+i, y+j\right)-{\overline{I}}_{x,y}\right)}^2·{\sum }_{i,j}{\left(T\left(i,j\right)-\overline{T}\right)}^2}}}$$

(9)

where$i$ and $j$ are indices spanning the template dimensions, $I\left(x+i,y+j\right)$ is pixel intensity in the reference image at position $\left(x+i, y+j\right)$, relative to the top-left corner of the window at $(x, y)$, $T(i,j)$ is pixel intensity in the template at position $(i, j)$, ${\overline{I}}_{x,y}$ is mean intensity of the reference image window with top-left corner at $(x,y)$, and $\overline{T}$ is mean intensity of the template.

The estimated UAV position $\left(\widehat{x},\widehat{y}\right)$ is obtained as:

$$\left(\widehat{x},\widehat{y}\right)=\mathrm{a}\mathrm{r}\mathrm{g}\underset{x,y}{\mathrm{m}\mathrm{a}\mathrm{x}} \mathrm{N}\mathrm{C}\mathrm{C}(x,y)$$

(10)

The step-by-step application of the signal processing described above is illustrated in Figure 6.

2.3. Sensor Fusion

The EKF [30] integrates inertial navigation outputs with external measurements to refine localization accuracy (Figure 7). The framework consists of three components: a horizontal channel filter, a vertical channel filter, and a barometric altimeter error estimator.

2.3.1. Horizontal Channel EKF

The horizontal-channel EKF fuses INS navigation states with external measurements from GNSS or SM. The state vector is defined as:

$$x_k={\left[\delta \phi \delta \lambda {\delta V}_e {\delta V}_n {\psi }_e {\psi }_n {\psi }_u {\alpha }_x {\alpha }_y {\alpha }_z {\beta }_x {\beta }_y {\beta }_z\right]}^T$$

(11)

where $\delta \phi$, $\delta \lambda$ denote latitude and longitude errors, ${\delta V}_e$, ${\delta V}_n$ are velocity errors in the east and north directions, ${\psi }_e$, ${\psi }_n$, ${\psi }_u$ are attitude (pitch, roll, yaw), and ${\alpha }_i$ and ${\beta }_i$ represent accelerometer and gyro biases along $i$-axis, respectively.

The measurement vector consists of the differences between the INS position and those obtained from GNSS or SM. For synchronization, the navigation states are stored in the INS buffer at 300 Hz. Time synchronization between the INS and GNSS is achieved using the Pulse Per Second (PPS) signal, whereas synchronization with SM is performed based on the image acquisition time. In this process, the latency between aerial image capture and its delivery to the filter is pre-measured and compensated. For instance, in the flight tests described in Section 3, the latency of the aerial image was measured to be approximately 170 ms.

Measurement updates in the EKF are performed based on a priority scheme, where GNSS measurements have higher priority than SM measurements. That is, when GNSS measurements are available, they are used to update the EKF, whereas in GNSS-denied environments, SM measurements are employed for the update. The measurement models are:

$$Z_{GNSS,k}=\left[\begin{array}{c}{\phi }_{INS,k}-{\phi }_{GNSS,k}\\ {\lambda }_{INS,k}-{\lambda }_{GNSS,k}\end{array}\right]=H{x}_k+v_{GNSS, k}$$

(12)

$$Z_{SM,k}=\left[\begin{array}{c}{\phi }_{INS,k}-{\phi }_{SM,k}\\ {\lambda }_{INS,k}-{\lambda }_{SM,k}\end{array}\right]=H{x}_k+v_{SM,k}$$

(13)

where $\phi$ and $\lambda$ denote latitude and longitude, respectively.

To prevent incorrect measurements from being incorporated into the EKF, a Mahalanobis distance-based validation is applied. For each SM measurement, the Mahalanobis distance between the predicted state and the measurement is computed. Measurements exceeding a predefined threshold are rejected as outliers, ensuring robust EKF updates and mitigating the effect of false matches on localization accuracy.

The Mahalanobis distance is defined as

$$d_M=\sqrt{{\left(z-\widehat{z}\right)}^T{S}^{-1}\left(z-\widehat{z}\right)}$$

(14)

where $z$ is the observed measurement, $\widehat{z}$ is the predicted measurement, and S is the innovation covariance.

2.3.2. Vertical Channel EKF

Errors in the vertical channel of the INS diverge rapidly compared to the horizontal channel, necessitating external altitude stabilization. In this study, altitude stabilization is achieved by using GNSS altitude whenever available and barometric altitude otherwise. When GNSS altitude is available, it is also used to estimate the bias and scale-factor error of the barometer. This ensures consistency of barometric altitude with the WGS-84 reference, even during GNSS outages.

The state vector for the vertical channel is:

$$x_k={\left[\begin{array}{cc}\delta h& \delta V_u\end{array}\right]}^T$$

(15)

where $\delta h$ and $\delta V_u$ denote altitude and vertical velocity errors, respectively.

2.3.3. Barometric Altimeter Error Estimator

The barometric altimeter error estimator [31] operates jointly with the vertical channel filter. When GNSS altitude is available, it estimates the barometer bias and scale factor error, enabling correction of barometric altitude during GNSS outages. The state vector is:

$$x_k={\left[\begin{array}{cc}{B}_{Bias}& B_{SF}\end{array}\right]}^T$$

(16)

and the measurement model is defined as:

$$z_k=\left[\begin{array}{c}{h}_{baro,k}-(1+B_{SF})h_{GNSS,k}\\ {\displaystyle \frac{h_{baro,k}}{h_{GNSS,k}}}-1\end{array}\right]=H{x}_k+v_{k,baro}$$

(17)

The corrected barometric altitude is then computed as:

$$h_{baro,corrected,k}=\left(1+B_{SF}\right)h_{baro,k}+B_{Bias}$$

(18)

Which ensures alignment with GNSS altitude in the WGS-84 datum.

3. Experimental Setups and Results

This section describes the experimental setup and the geolocation performance evaluation of the proposed VBN algorithm, specifically designed for UAV position estimation in mountainous terrain.

3.1. Experiment on Real Flight Aerial Image Dataset

Flight tests were conducted in April 2024 using a Cessna 208B aircraft, equipped with onboard sensors including an IMU, a GNSS receiver, a downward-facing camera, a barometric altimeter, and a signal processing computer (Figure 8). This configuration enabled the collection of a comprehensive dataset for evaluating the performance of both VBN and SM.

In general, clear imagery of the terrain is essential for achieving high accuracy in SM. When the terrain is obscured by dense fog, snow, or shadows, the accuracy of matching is significantly reduced, and false matches may occur. Therefore, the flight scenarios in this study assumed favorable weather conditions, and the aerial images were collected under stable and clear weather.

The reference orthophoto maps and DEM used for SM were provided by the Korean government. The orthophoto maps were originally produced in 2022 with a resolution of 1 m, and for performance analysis in this study, the maps were downsampled to resolutions of 2 m, 4 m, and 8 m. The DEM was produced in 2019 with a resolution of 5 m.

3.1.1. Sensor Specification

The specifications of the onboard sensors are summarized in

Table 1. Characteristics of the onboard sensors used in the flight tests.

Sensor	Output Rate	Resolution	FOV	Focal Length	Bias
Camera	1 Hz	4096 × 3000 pixels *	37.84° (H), 29.19° (V)	16 mm	—
Gyroscope	300 Hz	—	—	—	0.04°/h
Accelerometer	300 Hz	—	—	—	60 µg
Barometer	5 Hz	—	—	—	—

* The aerial imagery was downsampled by 1/4 to create an image dataset of 1024 × 750 pixels.

. The camera was equipped with a 16 mm focal length lens, providing a horizontal field of view (HFOV) of 37.84° and a vertical field of view (VFOV) of 29.19°. The original images were captured at a resolution of 4096 × 3000 pixels, but they were downsampled by 1/4 to create an image dataset of 1024 × 750 pixels. The camera’s frame rate is 1 Hz. At an altitude of 3 km, this corresponds to a ground sample distance (GSD) of approximately 2 m. The IMU included accelerometers with a 300 Hz output rate and a bias of 60 µg, and gyroscopes with a 300 Hz output rate and a bias of 0.04°/h. The barometer provided altitude measurements at a 5 Hz output rate, serving as an external source to constrain divergence in the vertical channel of the INS.

3.1.2. Flight Scenarios and Dataset Characteristics

The dataset was collected over five flight paths across two distinct mountainous regions with significant terrain elevation variations, as illustrated in Figure 9 and summarized in

Table 2. Characteristics of the flight dataset for different trajectories, including heading, altitude, duration, length, and speed. The GSD of the aerial images is calculated using flight altitude and camera parameters.

Path ID	Heading (°)	Altitude (km)	Duration (min)	Length (km)	Speed (km/h)	GSD (m)
1	−172.9	3.5	8.7	44.5	304	2.3
2	−8.4	3.5	8.5	44.6	311	2.3
3	−176.9	2.5	8.9	44.6	300	1.5
4	6.6	2.5	9.4	45.4	289	1.5
5	−165.4	2.5	9.8	45.4	278	1.5

. Paths 1 to 3, with headings −172.9° (north-to-south), −8.4° (south-to-north), and −176.9° (north-to-south), covered one region at altitudes of 3.5 km (Paths 1 and 2) and 2.5 km (Path 3). Paths 4 and 5, with headings 6.6° (south-to-north) and −165.4° (north-to-south), covered another region at 2.5 km altitude. Each path spanned approximately 45 km over 9 min at 300 km/h. Aerial images captured by the camera yielded a GSD of approximately 2.3 m at 3.5 km altitude and 1.5 m at 2.5 km altitude. This design enabled evaluation of the algorithm’s robustness under diverse flight directions, altitudes, and terrain conditions.

In addition, the flight scenarios were assumed to start under GNSS-available conditions, before transitioning to GNSS-denied conditions. Accordingly, at the beginning of each dataset, the position, velocity, and attitude errors of the INS were considered to be well-corrected by the EKF using GNSS updates, resulting in a negligible initial navigation error. This assumption ensures that the performance evaluation focuses on the degradation and recovery of navigation accuracy after the transition to GNSS-denied conditions.

3.1.3. Data Processing

The collected dataset was processed in post-flight simulations to evaluate the proposed algorithm. Aerial images were orthorectified using aided navigation information to correct terrain-induced geometric distortions, ensuring geometric consistency with the reference orthophoto maps. The algorithms were implemented in Python 3.8. OpenCV was employed for reading aerial images and applying image preprocessing operations, while NumPy and SciPy were used for numerical computations. The simulations were conducted on a standard PC running Windows 10 with an Intel i7 CPU and 32 GB RAM.

Carrier-phase Differential GNSS (CDGNSS) data provided high-precision ground truth for validating horizontal position. The reference trajectory was obtained using NovAtel’s CDGNSS post-processing software GrafNav 8.90, which provides sub-meter horizontal accuracy.

3.2. Localization Accuracy

Figure 10 illustrates the localization errors obtained from SM between the captured aerial imagery and georeferenced orthophoto maps. Figure 10a shows the results obtained using conventional rectification, whereas Figure 10b presents the results using the proposed orthorectification method.

In Figure 10a, a substantial increase in localization error is observed in areas with significant terrain elevation changes (as shown in Figure 9c). For example, when the image numbers are approximately 140, 280, and 440, the localization error rises sharply, which coincides with the steep terrain variations in Flight Path 2 shown in Figure 9c. Furthermore, improving the map resolution from 8 m to 2 m provides only marginal benefits in position accuracy when conventional rectification is applied. In contrast, Figure 10b demonstrates a considerable reduction in localization error compared to Figure 10a. Notably, as the reference map resolution improves from 8 m to 2 m, localization accuracy shows a marked enhancement, underscoring the effectiveness of the proposed orthorectification approach in mitigating terrain-induced errors.

The performance of the proposed VBN algorithm—integrating SM with orthorectification and the sensor fusion framework—was evaluated under GNSS-denied conditions. Localization accuracy was assessed using the two-dimensional root-mean-square error (2D RMSE), with results summarized in

Table 3. Localization accuracy summarized by 2D RMSE for rectification (Rect.) and orthorectification (Ortho.) across different map resolutions (2, 4, and 8 m) and flight paths (Path ID 1–5).

Path ID	Rect. 8 m	Rect. 4 m	Rect. 2 m	Ortho. 8 m	Ortho. 4 m	Ortho. 2 m
1	11.9	10.3	10.3	9.1	4.5	4.6
2	10.3	9.1	8.4	7.5	5.3	3.4
3	9.4	7.8	6.4	8.8	3.6	2.6
4	9.4	5.9	6.0	10.4	5.6	3.4
5	15.3	10.1	9.6	8.4	5.4	3.5

and Figure 11. The results are organized according to flight paths, map resolution, and the use of orthorectification.

Table 3 confirms that applying orthorectification consistently improves localization accuracy across various flight paths and map resolutions. The improvement is particularly pronounced for Flight Path 5. Flight Paths 4 and 5 involve steeper terrain and lower altitudes (approximately 1 km lower than other paths), presenting more challenging conditions for SM. These conditions lead to a higher incidence of false matching and increased distortion in the aerial imagery. Nevertheless, the proposed orthorectification method effectively mitigated these errors, yielding substantial accuracy gains even in the most challenging scenarios.

As illustrated in Figure 11, localization error distributions shift markedly when orthorectification is applied. Whereas the benefits of rectification tend to saturate as the map resolution increases, orthorectification exhibits an almost linear reduction in RMSE with finer resolutions. For example, with the 2 m resolution map, localization errors were reduced by approximately 43–64% compared with rectification. Under the same conditions, rectification alone produced errors equivalent to 3–5.2 pixels, while the proposed method achieved an average error of only 1.4 pixels across all tested scenarios.

These results confirm both the robustness and the effectiveness of the proposed VBN framework in enhancing UAV navigation performance in GNSS-denied environments.

Figure 12 presents the navigation errors for Flight Path 2 in GNSS-denied environment, utilizing a 4 m resolution orthophoto map. The proposed VBN algorithm is compared with INS and VBN with rectified aerial images. INS exhibits a gradual increase in localization error over time, accumulating a position error of approximately 300 m over an 8.5 min flight duration. The VBN with conventional rectification yields a final position error of 10.2 m, with errors remaining within 30 m. By contrast, the proposed method achieved a final position error of 2.9 m, maintaining errors within 10 m throughout the flight. These results were obtained under the same test conditions (clear weather, mountainous terrain, and a 4 m resolution orthophoto map) and indicate that the proposed VBN algorithm is a promising alternative to GNSS in GNSS-denied environments. However, the performance may degrade depending on the quality of the reference map and environmental factors such as clouds, heavy fog, or significant temporal appearance changes. Under these test conditions, the algorithm demonstrated a significant reduction in localization errors for UAVs operating at high altitudes over mountainous terrain.

4. Discussion

4.1. Effect of VBN Errors on SM

The proposed orthorectification method relies on the UAV’s position and attitude information. In this process, the UAV’s attitude determines the geometry of camera projection rays, while horizontal position errors induce inaccuracies in the elevation retrieved from the DEM. Additionally, altitude errors affect the projection distance, influencing the scaling of the captured imagery.

The experiments in Section 3 assumed a scenario where the UAV initially operates in a GNSS-available environment before transitioning to a GNSS-denied condition, with SM outputs used to update the EKF. This implies that INS position, velocity, and attitude errors are well-corrected in GNSS-available conditions, resulting in minimal navigation errors at the beginning of GNSS-denied operation.

This section provides an additional analysis of how aided navigation errors affect the performance of the proposed SM. Simulations were conducted for three cases summarized in

Table 4. Simulation cases for SM performance under navigation errors.

Case	Roll/Pitch Error (°)	Yaw Error (°)	Latitude/Longitude Error * (m)	Altitude Error (m)
1	0.03	0.5	10	20
2	0.05	1	20	40
3	0.1	1.5	30	60

* Latitude and longitude errors were converted into meters using the WGS-84 ellipsoid, where 1° latitude ≈ 111.32 km and 1° longitude ≈ 111.32 × cos (latitude) km at the latitude of the test region (Korea).

. As described in Section 3.1.1, the IMU used in this study has a gyro bias of 0.04°/h and an accelerometer bias of 60 µg. With such specifications, the expected alignment performance corresponds to approximately 0.003° in horizontal attitude error and 0.2° in yaw error [30]. To reflect more challenging conditions, the attitude errors were set to values more than ten times larger: roll/pitch errors of 0.03/0.05/0.1°, and yaw errors of 0.5/1.0/1.5°. The position errors were defined as 10/20/30 m in latitude and longitude, and 20/40/60 m in altitude, reflecting typical GNSS receiver accuracies. Simulations were carried out using the Flight Path 2 dataset with 4 m resolution georeferenced maps, and the results are presented in Figure 13 and

Table 5. VBN localization accuracy for navigation error cases.

Case	2D RMSE (m)
	VBN with Orthorectification	VBN with Rectification
Normal	5.32	9.08
1	5.52	9.40
2	6.58	9.19
3	9.97	10.80

.

Figure 13a illustrates the SM results with orthorectification applied, while Figure 13b corresponds to conventional rectification. In Figure 13a, the SM position error gradually increases as the navigation errors grow. In Case 3, where the largest errors are introduced, terrain-induced distortions are not fully compensated, resulting in errors comparable to those observed with rectification (Figure 13b). Table 5 summarizes the performance of the VBN algorithm when using the SM outputs as measurements. Although these errors lead to a slight degradation in performance, the impact remains within tolerable range.

These findings demonstrate that the proposed SM and VBN method maintains reliable localization accuracy even in the presence of moderate navigation errors, highlighting its robustness for practical UAV operations in GNSS-denied environments.

4.2. Effect of Flight Altitude and Shadows on SM

Because SM estimates UAV position by registering aerial images to reference maps, its accuracy is inherently dependent on the quality of both the reference maps and the captured imagery. Distinct terrain features are especially beneficial for improving localization accuracy. Since the proposed method employs an area-based template matching approach, a larger footprint increases the likelihood of including distinctive features, whereas smaller footprints reduce reliability and increase the risk of false correspondences. With a fixed FOV camera, low-altitude imagery produces narrower footprints, limiting the available terrain features for matching.

Figure 14, Figure 15 and Figure 16 illustrate the influence of footprint size and shadow on SM performance. Figure 14 shows the template matching result for image number 41 in Flight Path 2 (altitude: 3.5 km). The processed aerial image (Figure 14b) covers approximately 4.1 km² and is correctly matched to the rectangular region highlighted in the reference map (Figure 14a). In contrast, Figure 15 presents the result for image number 298 in Flight Path 4 (altitude: 2.5 km), where the footprint coverage is only about 1.7 km², approximately 40% of that in Figure 14b. Due to the smaller footprint and the widespread shadows across the mountainous terrain in Figure 15a, a false match occurred, producing a localization error of 0.9 km.

Figure 16 further shows SM localization errors along Flight Path 4, where the combined effects of reduced footprint and terrain shadows produced abrupt error spikes. Although orthorectification reduced the overall error and the frequency of false matches compared with conventional rectification, intermittent mismatches caused by narrow footprints were still observed. While such outliers can be identified and rejected in the EKF update process, persistent mismatches could significantly degrade VBN performance. Therefore, the proposed method is considered most effective when operated above a certain altitude, where sufficient terrain coverage is ensured and the risk of shadow-induced mismatches is reduced.

5. Conclusions

This study proposed a VBN framework that integrates orthorectification-based SM with INS- and barometer-aided sensor fusion for UAV operations in GNSS-denied environments. By compensating for terrain-induced distortions, the method consistently improved localization accuracy across different flight paths and reference map resolutions.

The experimental results using real aerial flight data demonstrated that the proposed framework achieves an average localization accuracy of 1.4 pixels. The findings confirm that orthorectification is a critical step for enhancing the reliability of SM, especially in mountainous terrain where geometric distortions are significant. In particular, the framework maintained stable performance in challenging cases such as steep relief and high-altitude imagery, where conventional rectification methods typically produce large errors. These contributions establish the proposed framework as a viable alternative to GNSS in environments where satellite signals are unavailable, degraded, or intentionally denied.

Despite these promising results, several limitations remain. The flight trajectories used in this study were limited to straight paths at altitudes of 2.5–3.5 km. At lower altitudes over mountainous terrain, the reduced image footprint is expected to make matching more difficult. It should be noted that the proposed method is sensitive to the quality of both reference maps and captured images, and may be affected by severe occlusions due to shadows, fog, or clouds. The resolution and quality of the DEM and georeferenced images are expected to influence the overall localization performance; therefore, high-precision datasets provided by the Korean government were employed in this study to focus on verifying the fundamental effectiveness of the proposed framework. Moreover, although the dataset was systematically constructed through real flight tests with synchronized INS, camera, GNSS, and barometer measurements, the validation was performed in an offline environment rather than in real time. Finally, the IMU employed was of navigation-grade, which may not fully represent the performance achievable with lower-cost sensors.

Future work will address these limitations in several directions: (i) extending validation to diverse flight trajectories and altitudes, including low-altitude UAV flights in mountainous terrain; (ii) optimizing and accelerating the algorithm for real-time onboard implementation on resource-constrained UAV hardware; and (iii) investigating the applicability of the framework with low-cost MEMS-grade IMUs to broaden its practicality for small UAVs.