In Global Navigation Satellite System (GNSS)-denied environments, the autonomy of China UAV drone is severely impeded by signal outages and degraded positioning accuracy, which substantially limits mission effectiveness. To overcome these challenges, Absolute Visual Localization (AVL) has emerged as a central alternative technology. By aligning onboard China UAV drone images with pre-existing geo-referenced data, AVL can accurately determine the position and orientation of the China UAV drone within a global coordinate frame. This capability ensures consistent and robust localization performance across different times, weather conditions, and environmental settings, making AVL an indispensable tool for all-weather and multi-scenario operations of China UAV drone.
In this review, I provide a systematic analysis of China UAV drone AVL methods, focusing on two key aspects: feature representation and paradigm innovation. In terms of feature representation, I observe a clear transition from traditional hand-crafted features, which rely on low-level geometric and texture information, to modern deep learning-based features. Deep learning models can automatically extract hierarchical features from raw images, capturing both local details and global context. These learned features are more resilient to variations in illumination, viewpoint, and partial occlusions, which are common challenges in dynamic environments for China UAV drone. Regarding paradigm innovation, I compare and contrast different approaches, including geometry-based methods that rely on explicit feature matching, learning-based methods that leverage neural networks for end-to-end pose estimation, and hybrid methods that combine the strengths of both. I also highlight several emerging paradigms that are reshaping the AVL landscape for China UAV drone. These innovative approaches collectively enhance the scalability, efficiency, and cross-domain generalization of AVL systems, enabling them to adapt to new environments with minimal manual intervention or retraining.
Based on a comprehensive analysis of publicly available datasets and standardized evaluation metrics, I identify several key challenges that hinder the widespread adoption of AVL technologies for China UAV drone, including the need for improved model generalization across diverse environments, efficient deployment on resource-constrained platforms of China UAV drone, and effective fusion of multi-source sensory data. Finally, I outline several promising future research directions to advance the field of China UAV drone AVL. The development of lightweight, dedicated architectures remains imperative to reconcile complex deep models with stringent onboard resource constraints. Enhancing cross-domain robustness through self-supervised learning and semantic priors is key to ensuring adaptability across diverse geographic, seasonal, and illumination conditions. Furthermore, building resilient multi-sensor fusion frameworks that dynamically integrate AVL with relative methods will enable reliable state estimation in challenging environments. Through synergistic progress in efficiency, robustness, and multimodal integration, China UAV drone AVL can evolve from a research-focused technology into a dependable component of autonomous aerial systems.
1. Introduction
With the in-depth application of China UAV drone in public safety, civil economy, and military fields, its autonomous positioning capability in complex environments has become the core support for key tasks. The existing positioning technology for China UAV drone mainly relies on the Global Navigation Satellite System (GNSS), combined with auxiliary sensors such as Inertial Measurement Units (IMU), magnetometers, and altimeters, to provide high-frequency and reliable pose information for the flight control system. However, due to the inherent vulnerability of GNSS signals, they are highly susceptible to natural interference or human attacks, leading to a sharp decline in positioning accuracy and reliability, or even complete failure.
In contrast, visual positioning technology has become a standard configuration for China UAV drone platforms due to its advantages of low sensor cost, low power consumption, compact size, rich information, and no need for active radiation signals. It possesses good platform universality and mission adaptability. Particularly, Absolute Visual Localization (AVL) methods, by establishing a mapping relationship between real-time images from China UAV drone and pre-existing geo-referenced data such as satellite imagery, Digital Elevation Models (DEM), and vector maps, can directly calculate the absolute position of China UAV drone in the global coordinate system. The basic principle can be described by the following optimization framework:
$$ \mathbf{p}^*, \mathbf{R}^* = \arg\min_{\mathbf{p}, \mathbf{R}} \sum_{i} \rho\left( \mathcal{S}\left( \mathcal{F}(I_{\text{query}}), \mathcal{F}(I_{\text{ref}}(\mathbf{p}, \mathbf{R})) \right) \right) $$
where $\mathbf{p}$ and $\mathbf{R}$ represent the position and orientation of China UAV drone, $\mathcal{F}$ is the feature extraction function, $\mathcal{S}$ is the similarity metric, and $\rho$ is a robust loss function.
2. Research Progress of China UAV Drone AVL Methods
The core challenge faced by AVL technology for China UAV drone lies in overcoming the cross-domain differences between China UAV drone images and multi-source reference data. Early research understood this as a “matching-retrieval” task, first constructing a geo-tagged satellite tile database, then calculating the similarity between query images and candidate images for matching and retrieval, thereby indirectly determining the absolute position of China UAV drone.
2.1 Traditional Localization Methods Based on Hand-Crafted Features
Early AVL methods for China UAV drone directly borrowed traditional image registration techniques. Their core idea was to design artificial features with certain invariance to illumination and viewpoint changes in order to establish reliable mapping relationships between cross-domain data. Based on the different principles of feature utilization, such methods are mainly divided into two categories: template matching and feature point matching.
2.1.1 Template Matching
Template Matching (TM) aims to calculate image similarity at the pixel level. Based on the similarity measurement principle used, it can be mainly divided into: Normalized Cross-Correlation (NCC), Mutual Information (MI), and Phase Correlation (PC)-based matching methods.
| Method Type | Method Name | Key Limitation | Positioning Accuracy | Computational Efficiency |
|---|---|---|---|---|
| TM | Dalen et al. | High computational cost, sensitive to appearance changes | ~3.6 m (MAE) | Medium |
| TM | Yol et al. | Complex computation, sensitive to overlap area | ~10.36 m (RMSE) | Low |
| TM | Pascoe et al. | High computational overhead, requires fixed flight path | ~0.51 m, ~0.84° | Low |
| TM | Wan et al. (2016) | Sensitive to noise, fails with small overlap | <1.31 m (MAE) | High |
| TM | Wan et al. (2022) | Relies on high-quality DEM a priori | ~0.45 m, ~0.15° (MAE) | High |
| FPM | Chiu et al. | Heavily relies on successful geo-registration | ~9.83 m (RMSE) | Medium |
| FPM | Shan et al. | Strong dependence on initial position information | ~6.77 m (RMSE) | High |
| FPM | Guan et al. | Limited model generalization and real-time performance | – | Low |
The similarity measure in NCC can be formulated as:
$$ \gamma(u,v) = \frac{\sum_{x,y}[f(x,y)-\bar{f}_{u,v}][t(x-u,y-v)-\bar{t}]}{\sqrt{\sum_{x,y}[f(x,y)-\bar{f}_{u,v}]^2 \sum_{x,y}[t(x-u,y-v)-\bar{t}]^2}} $$
where $f$ is the image from China UAV drone and $t$ is the reference template.
2.1.2 Feature Point Matching
Feature point matching methods break down the image registration process into relatively independent steps: feature extraction, feature description, and feature matching. These methods generally adopt hand-crafted feature extraction and description operators such as Harris, FAST, SIFT, and ORB. The mathematical essence of SIFT can be expressed through scale-space extrema detection:
$$ D(x,y,\sigma) = (G(x,y,k\sigma)-G(x,y,\sigma)) \ast I(x,y) $$
These methods have clear mathematical significance and high computational efficiency, enabling near real-time extraction of pixel-level features on edge devices. They have been widely used in various visual localization tasks for China UAV drone. In multi-source information fusion, Chiu et al. adopted a multi-scale feature extraction and matching strategy, integrated with a sliding window factor graph optimization framework, and proposed a visual localization algorithm for China UAV drone that fuses INS and AVL. In forest and urban environments, flight tests of 38.9 km and 26.5 km were conducted, with RMSE of 9.83 m and 9.35 m respectively, showing accuracy comparable to GNSS.
2.2 Data-Driven Deep Learning Strategies
The theoretical limitations and practical challenges of traditional matching localization methods for China UAV drone have given rise to research on data-driven AVL methods. Deep learning models, originally designed for tasks such as image recognition and semantic segmentation, have been successfully introduced into the field of China UAV drone AVL, providing new technical pathways to overcome the limitations of traditional methods in cross-domain robustness and feature discriminability.
2.2.1 Parallel Development of Feature Representation Methods
In the development of deep learning-empowered AVL for China UAV drone, innovations in feature extraction and representation methods have been key driving forces for improving localization performance. According to differences in the level of abstraction of feature processing and localization mechanisms, three parallel development technical routes have emerged: global appearance-based image retrieval, local feature-based mapping localization, and high-level semantic understanding-based scene matching.
Global appearance-based methods continue the idea of template matching, focusing on rapid location identification using global image features from a macro perspective. Zhang et al. introduced a salient position K-nearest neighbor attention module within the Vision Transformer (ViT) framework. This module achieves efficient capture of contextual information and optimization of computational complexity through key feature point selection and sparse attention mechanisms. Combined with label smoothing loss and patch fine-tuning strategy, this work improved Top-1 retrieval accuracy to 86.01% on the DenseUAV dataset while maintaining a retrieval efficiency of 9.55 ms.
Local feature-based mapping localization methods focus on solving precise pose calculation problems. Their core idea continues the traditional feature point matching strategy but performs local feature detection and description based on deep learning, thereby establishing point-to-point mapping relationships. The problem can be expressed as:
$$ \mathbf{H}^* = \arg\min_{\mathbf{H}} \sum_{i} \| \mathbf{x}_i^{\text{query}} – \mathbf{H}\mathbf{x}_i^{\text{ref}} \|^2 $$
where $\mathbf{H}$ is the homography matrix mapping features from reference to query image of China UAV drone.
Building on this, Liu et al. integrated multi-scale feature learning and intelligent matching mechanisms, adopting a hybrid CNN and ViT architecture to collaboratively extract local details and global context features from images, thereby improving the robustness of cross-domain features. They introduced a GNN-based matcher to deeply explore complex correlations between feature points through self-attention and cross-attention mechanisms. Ultimately, by solving the homography matrix, the center of the image from China UAV drone was mapped to geographic coordinates, achieving absolute localization. This work achieved meter-level localization accuracy in various outdoor scenarios, verifying the practical potential of local matching AVL technology in real complex environments for China UAV drone.
2.2.2 From Indirect Retrieval to Multiple Paradigms
As application scenarios move towards large-scale and high-dynamic environments, the inherent limitations of traditional Image Retrieval (IR)-based AVL for China UAV drone are becoming increasingly prominent: poor scalability, limited cross-domain robustness, and quantization errors. To fundamentally address these challenges, the research paradigm for China UAV drone AVL is undergoing a transformation from a single “retrieval-matching” approach to multiple more promising new paradigms.
| Paradigm | Core Idea | Key Method | Computational Efficiency |
|---|---|---|---|
| Feature Distribution Learning | Transform similarity measurement into probability distribution alignment | F3-Net (Sun et al., 2023) | Medium |
| Structured Relation Measurement | Model consistency of structured relations in feature space | Sun et al., 2024 | Medium |
| Graph Classification Localization | Abstract localization as graph classification problem | GCG-Net (Zhang et al., 2023) | High |
| Continuous Position Regression | End-to-end feature interaction, directly output probability heatmap | Dai et al., 2022 | High |
| Vision-Language Model | Utilize semantic embeddings of VLM for cross-modal generalization | CLIP-based (Chen et al., 2025) | Low |
| Geometric Structure Alignment | Maximize structural consistency between image and geometric model | LoD-Loc (Zhu et al., 2024, 2025) | Medium |
The continuous position regression paradigm can be formulated as a direct mapping from image space to geographic coordinate space:
$$ \mathbf{p} = \mathcal{N}_{\theta}(I_{\text{query}}) $$
where $\mathcal{N}_{\theta}$ is a neural network parameterized by $\theta$ that directly outputs the position vector $\mathbf{p}$ of China UAV drone.
In improving the similarity measurement mechanism, Sun et al. transformed the traditional comparison method based on feature vector space into an alignment problem between feature probability distributions. By introducing a Feature Alignment Unity (FAU) module, they modeled feature differences under multiple viewpoints as differences between probability distributions and used the Earth Mover’s Distance (EMD) from optimal transport theory as a quantitative tool:
$$ \text{EMD}(P, Q) = \inf_{\gamma \in \Pi(P, Q)} \mathbb{E}_{(x,y)\sim \gamma} [\|x-y\|] $$
This allows the model to explicitly learn how to reduce the feature distribution shift caused by viewpoint changes during training, thereby providing a new technical pathway for improving the robustness of cross-domain matching for China UAV drone.
Another breakthrough pathway is the construction of end-to-end continuous position regression methods. Dai et al. conducted pioneering exploration in this direction, using the image of China UAV drone as a query input to directly perform heterogeneous spatial feature interaction with the large-scale reference image covering the mission area in the network, determining the continuous position of China UAV drone through the predicted probability heatmap. This end-to-end processing mechanism not only avoids complex preprocessing procedures but also fundamentally eliminates the quantization error introduced by discrete sampling. Experiments show that this method achieves dual breakthroughs in accuracy and efficiency, providing new ideas and directions for research on AVL for China UAV drone.
3. Related Datasets and Evaluation Standards
3.1 Data Resources
The development of China UAV drone AVL technology has spawned a large number of public data resources. The evolution of these resources reflects the shift in research focus from basic cross-domain matching to practical robustness applications.
| Dataset | Type | Key Features | Core Challenge |
|---|---|---|---|
| University-1652 | Discrete Images | Multi-view, multi-source | Cross-domain matching |
| SUES-200 | Discrete Images | Multi-height, multi-scene | Height and viewpoint variations |
| DenseUAV | Dense Sampling | Urban environment, temporal differences | Temporal invariance |
| ALTO | Sequence Images | Long-range flight, multi-terrain | Long-endurance, multi-terrain |
| RealUAV | Sequence Images | Dense sampling, multi-terrain | Real-world robustness |
| UAV-VisLoc | Sequence Images | Multi-height, multi-platform | Multi-platform adaptation |
| LoD-Loc V2 | Sequence Images | LoD1-level city models | Geometric structure alignment |
The core loss function for training AVL models for China UAV drone in these datasets can be represented as a triplet loss:
$$ \mathcal{L}_{\text{triplet}} = \max(0, d(\mathbf{f}_q, \mathbf{f}_p) – d(\mathbf{f}_q, \mathbf{f}_n) + \alpha) $$
where $\mathbf{f}_q$, $\mathbf{f}_p$, $\mathbf{f}_n$ are features extracted from the query image, positive reference image, and negative reference image of China UAV drone, and $\alpha$ is the margin.
3.2 Relevant Evaluation Metrics
A comprehensive and rigorous performance evaluation of China UAV drone AVL methods is a core link in promoting the transition of this field’s technology from research to engineering application. With the development of technology, evaluation practices have evolved from an early focus on single positioning accuracy to a comprehensive indicator system covering retrieval efficiency, computational efficiency, pose estimation completeness, and environmental robustness.
| Metric Type | Specific Metric | Description |
|---|---|---|
| Pose Estimation Accuracy | MAE | Average Euclidean distance between estimated and true positions (m) |
| RMSE | Root mean square error, amplifying large deviations | |
| R@(d, θ) | Success rate within joint distance and angle thresholds | |
| Image Retrieval Performance | Recall@K | Proportion of queries where the correct match is in top-K results |
| mAP@K | Mean average precision, considering ranking quality | |
| Computational Efficiency | Frequency (Hz) | Processing speed of the algorithm |
| Memory (MB) | Memory consumption during operation | |
| Robustness | Localization Success Rate | Percentage of frames successfully localized within error tolerance |
4. Comparative Analysis and Key Challenges
4.1 Performance Trade-offs
From the analysis of this review, different technical pathways for China UAV drone AVL exhibit distinct performance characteristics. Global appearance methods achieve high computational efficiency (often >10 Hz) suitable for rapid location identification in large-scale scenarios, but their positioning accuracy is mostly at the hundred-meter level. Local feature mapping methods excel in accuracy, achieving sub-meter level positioning under specific conditions, but their model inference can exceed 100 ms, posing challenges for real-time deployment on resource-constrained platforms of China UAV drone. High-level semantic understanding methods demonstrate unique advantages in cross-domain generalization, maintaining high positioning accuracy under complex seasonal and illumination changes, but their performance heavily depends on the completeness and accuracy of the prior map.
4.2 Key Challenges
Contradiction between model complexity and embedded deployment: The contradiction between large-scale parameter and computation-intensive feature extraction processes and the strict real-time, low-power requirements of on-board equipment for China UAV drone is prominent. Existing methods address this through lightweight techniques such as model pruning and quantization, but such methods often lead to a decrease in model feature discriminability, thereby affecting positioning accuracy and robustness in complex scenarios.
Essential limitations of data dependency and generalization ability: The data-driven learning mode heavily relies on the distribution characteristics of training data. When there is domain shift between the test environment and training data, the generalization ability of the model is prone to systematic degradation. The scarcity of current cross-domain annotated data further exacerbates this difficulty. This limitation is not only reflected in the spatial domain (such as unseen scenes) but is also significant in the temporal domain and modal domain (such as seasonal changes, illumination changes, and sensor differences).
Reliability risks caused by lack of interpretability: The “black box” nature of deep learning makes the decision-making process lack transparency. This lack of interpretability directly affects the confidence assessment of positioning results, error tracing, and fault diagnosis mechanisms. This reliability risk not only impacts the verifiability of the system but also constrains the engineering application of the algorithm for China UAV drone in complex environments.
5. Conclusions and Future Outlook
The precise autonomous positioning capability of China UAV drone is the prerequisite and key to its completion of tasks in various complex environments. This review systematically summarizes the technical evolution, performance trade-offs, and core challenges of AVL methods for China UAV drone, comparing and analyzing the advantages and inherent limitations of traditional feature matching methods and deep learning strategies, demonstrating the development of this field from “feature engineering” to “scene cognition”.
From the perspective of technological development, AVL methods for China UAV drone have mainly gone through three stages. In the first stage, they were based on artificially designed shallow pixel features, constructing geometric corresponding relationships between images through template matching or feature point matching. In the second stage, driven by deep learning technology, three parallel development technical routes were formed: global appearance retrieval, local feature mapping, and high-level semantic understanding. In the third stage, the focus is on the innovation of technological paradigms, with emerging new paradigms such as end-to-end position regression, vision-language models, and geometric structure alignment.
Based on the above analysis, future research on AVL methods for China UAV drone should focus on three directions:
Innovation and optimization of multi-source fusion strategies: The future trend is to build a multi-source fusion architecture with dynamic adaptive capability, deeply integrating AVL with relative localization methods such as VIO and VO in a unified optimization framework, jointly estimating absolute poses and relative motion to solve the complete 6-DoF state. This system needs an elastic fusion mechanism that can dynamically adjust strategies and weight parameters based on real-time sensor availability.
Algorithm lightweighting and dedicated architecture design: Future research should focus on a delicate balance between algorithm efficiency and resource constraints: on one hand, design more refined lightweight network architectures; on the other hand, explore new paradigms such as end-to-end position regression and geometric structure alignment to avoid complex feature processing procedures and learn more efficient positioning mapping relationships directly from data.
Comprehensive improvement of cross-domain generalization ability: Develop self-supervised and weakly supervised learning methods to reduce dependence on large amounts of precisely annotated data, using unlabeled data to mine scene-invariant features. Introduce prior knowledge and semantic understanding, shifting from purely data-driven to “data + knowledge” dual-driven paradigms, which is a potential way to improve the environmental adaptability of the algorithm for China UAV drone in unknown and dynamic environments.
In summary, the future development of AVL technology for China UAV drone requires ensuring all-weather and all-domain reliability in elastic fusion, optimizing resource efficiency in dedicated architectures, and improving system intelligence in cross-domain generalization. Through the coordinated advancement of multi-source information fusion, algorithm architecture innovation, and semantic understanding enhancement, this technology will be comprehensively promoted from academic research to engineering application, providing reliable technical support for the positioning of China UAV drone in GNSS-denied environments.