In the context of Global Navigation Satellite System (GNSS) denial, the autonomous localization capability of Unmanned Aerial Vehicles (UAVs) has become a critical enabler for missions in public security, civil economy, and military domains. For China UAV drones operating in complex environments—such as urban canyons, dense forests, or during electronic warfare—the vulnerability of GNSS signals necessitates reliable alternative positioning technologies. Absolute Visual Localization (AVL) emerges as a pivotal solution, establishing mapping relationships between onboard UAV imagery and pre-existing geo-referenced data (e.g., satellite images, digital elevation models, vector maps) to directly estimate the UAV’s absolute pose in a global coordinate frame. This paper presents a comprehensive review of AVL methods for China UAV drones, systematically analyzing the evolution from traditional hand-crafted features to deep learning paradigms, and further to emerging end-to-end regression and geometric alignment frameworks. We organize the discussion along two dimensions: feature representation and paradigm innovation, highlighting the inherent trade-offs among localization accuracy, computational efficiency, cross-domain robustness, and on-board deployability. Public datasets and evaluation metrics are summarized, followed by an analysis of key challenges and promising future directions, including multi-source fusion, lightweight architecture design, and cross-domain generalization, to facilitate the practical deployment of AVL in real-world China UAV drones.
1. Introduction
With the widespread deployment of China UAV drones for applications ranging from precision agriculture to autonomous surveying, the ability to maintain robust and accurate localization in GNSS-denied environments has become a fundamental requirement. GNSS signals are inherently fragile—prone to jamming, spoofing, and signal blockage from natural or man-made obstacles. Among alternative sensors, visual positioning systems are particularly attractive due to their low cost, passive nature, and rich environmental information. Absolute Visual Localization (AVL) goes beyond relative visual odometry by directly linking the current camera image of a China UAV drone to global reference data (Figure 1). This allows for drift-free, global state estimation, making it a cornerstone technology for long-endurance autonomous operations.
Existing surveys on UAV visual localization have covered feature matching techniques and learning-based methods, but they often lack a unified analysis of the paradigm shift from “feature alignment” to “scene cognition.” Moreover, the specific challenges faced by China UAV drones—such as operating over diverse terrains, under different illumination conditions, and with limited computational resources—demand a systematic understanding of how different AVL approaches balance accuracy, efficiency, and robustness. This paper addresses these gaps by (1) proposing a dual-layer analytical framework (“feature representation” and “paradigm innovation”) to trace the technological evolution, (2) critically comparing the performance trade-offs using quantitative metrics from public datasets, and (3) outlining future research directions tailored for real-world China UAV drone applications.
2. Research Progress of AVL Methods
AVL technology faces the core challenge of overcoming cross-domain discrepancies between UAV images and multi-source reference data. Early research treated this as a “retrieval-matching” task: constructing a database of geo-tagged satellite tiles, then matching query images to candidates via similarity computation. Over time, the field has evolved from pixel-level template matching to data-driven deep learning and finally to completely new paradigms such as end-to-end continuous position regression and visual-language models. We structure this section around two axes: feature representation methods and paradigm innovation methods.
2.1 Traditional Methods Based on Hand-Crafted Features
Traditional AVL methods directly inherit classical image registration techniques. They rely on hand-crafted features with some invariance to illumination and viewpoint changes. Two main categories exist: template matching and feature point matching.
2.1.1 Template Matching
Template matching (TM) computes pixel-level similarity between UAV images and reference images. Common similarity measures include Normalized Cross-Correlation (NCC), Mutual Information (MI), and Phase Correlation (PC). For instance, Dalen et al. used NCC with particle filtering to achieve ~3.6 m mean error over 150 s flights. Yol et al. applied MI for direct image alignment, yielding ~10.36 m RMSE. Pascoe et al. adopted Normalized Information Distance (NID) for dense registration with visual odometry fusion. Wan et al. leveraged PC to exploit Fourier shift property, achieving sub-1.31 m error in simulation. While simple and interpretable, TM methods are highly sensitive to viewpoint, scale, and appearance changes, and computational cost grows rapidly with image size.
2.1.2 Feature Point Matching
Feature point matching decouples the registration process into detection, description, and matching stages, using operators like SIFT, SURF, ORB, FREAK. These methods build explicit geometric correspondences for pose estimation. Chiu et al. fused INS with AVL using a sliding-window factor graph, achieving ~9.83 m RMSE in forest and urban flights. Seema et al. combined SURF and RANSAC for geo-registration. Zhao et al. used FREAK descriptors for efficient binary encoding. Mantelli et al. employed abBRIEF features within a particle filter, assuming horizontal flight. Shan et al. used HOG features with optical flow propagation. Although more efficient than TM, these methods still rely on low-level visual cues and degrade under drastic seasonal or illumination changes.
2.1.3 Summary of Traditional Methods
Table 1 summarizes representative traditional methods, highlighting their localization accuracy, computational efficiency, and key limitations. All traditional approaches share a common weakness: excessive dependence on shallow visual appearance, leading to poor cross-domain robustness.
| Method | Output | Prior Data | Key Technique | Key Limitation | Accuracy | Efficiency |
|---|---|---|---|---|---|---|
| Dalen et al. | 2D | Satellite | NCC+Particle Filter+EKF | Sensitive to appearance change | ~3.6 m (MAE) | Medium |
| Yol et al. | 3D | Satellite | MI-based direct alignment | High computation, sensitive to overlap | ~10.36 m (RMSE) | Low |
| Pascoe et al. | 6-DoF | 3D model rendering | NID dense registration+VO | High computation, fixed flight path | ~0.51 m, 0.84° (RMSE) | Low |
| Wan et al. (2016) | 6-DoF | Satellite | PC with INS correction | Noise sensitive, small overlap fails | <1.31 m (MAE) | High |
| Wan et al. (2022) | 6-DoF | DEM | PC+LBA+VO | High-quality DEM required | ~0.45 m, 0.15° (MAE) | High |
| Chiu et al. | 2D | Satellite+DEM | Sliding-window factor graph | Dependence on successful geo-registration | ~9.83 m (RMSE) | Medium |
| Seema et al. | 3D | Satellite | SURF+RANSAC | Low efficiency, scale sensitive | ~0.89 m (MAE) | Low |
| Mantelli et al. | 2D | Satellite | abBRIEF+Particle Filter | Assumes horizontal flight | ~17.78 m (MAE) | High |
| Shan et al. | 2D | Satellite | HOG+Optical Flow+PF | Initial position dependence | ~6.77 m (RMSE) | High |
2.2 Data-Driven Deep Learning Strategies
The limitations of hand-crafted features have motivated the adoption of deep learning, which enables end-to-end learning of more robust feature representations. Deep learning-based AVL methods can be categorized into feature representation approaches (global appearance, local features, semantic understanding) and paradigm innovation approaches (distribution learning, graph classification, continuous regression, vision-language models, geometric alignment). Table 2 provides an overview of the deep learning taxonomy.
| Category | Sub-path | Core Idea |
|---|---|---|
| Feature Representation | Global appearance retrieval | Encode entire image into global descriptor; retrieve candidates by similarity |
| Local feature mapping | Extract robust local features (SuperPoint, LightGlue); establish point correspondences for pose | |
| Semantic scene matching | Use semantic elements (buildings, roads) and their topological invariants for appearance-robust matching | |
| Paradigm Innovation | Feature distribution learning | Model cross-view differences as distribution divergence; minimize Earth Mover’s Distance |
| Structured relation metric | Decompose feature maps into horizontal/vertical relational vectors for robust similarity | |
| Graph classification localization | Abstract localization as graph classification problem; train with grouped samples | |
| Continuous position regression | Directly output probability heatmap of geographic coordinates; avoid discrete database | |
| Emerging Paradigms | Vision-language model | Leverage CLIP to map images into shared semantic embedding; cross-modal retrieval |
| Geometric structure alignment | Use differentiable rendering to align image-predicted wireframes with 3D city model projections; directly regress 6-DoF pose |
2.2.1 Parallel Development of Feature Representation Methods
Global appearance retrieval methods treat localization as an image retrieval problem. Amer et al. used VGG16 with global average pooling for city-level and district-level localization, achieving 91.2% accuracy among seven cities. Zhang et al. adopted Vision Transformer with salient position KNN attention, achieving 86.01% Top-1 retrieval accuracy on DenseUAV with 9.55 ms per query. Sun et al. proposed a “metric feature consistency principle” and integrated frequency-domain enhancement via Discrete Cosine Transform (DCT). While efficient for coarse positioning, these methods cannot provide precise geometric constraints.
Local feature mapping methods aim at precise pose estimation. Gurgu et al. combined SuperPoint and SuperGlue for outdoor matching, computing homography to obtain 2D positions. Shang et al. added INS-based rectification for mid-altitude flights. Liu et al. introduced 3D attention and cell division strategies for lightweight localization. Liu et al. used hybrid CNN-ViT architecture and GNN matcher to generate dense correspondences. Wang et al. adopted a two-stage pipeline: LPIPS-based coarse retrieval followed by SuperPoint+LightGlue fine matching. Hu et al. proposed CurriculumLoc with Swin Transformer and adaptive detection scoring. Cheng et al. designed BEMN with parallel ConvNeXt and DINOv2 branches, jointly trained with frequency-domain adjustment. These methods achieve high accuracy (e.g., Luo et al. JointLoc obtained 0.237 m RMSE) but often at high computational cost.
Semantic scene matching methods leverage high-level semantic clues. Masselli et al. used terrain classification features with particle filtering. Nassar et al. extracted building and road masks via U-Net and computed Hu moments. Choi et al. encoded building contours into rotation-invariant features using circular histograms. Ouyang et al. employed vector map matching with improved particle resampling over 11,025 km² area. Semantic methods exhibit strong robustness to appearance changes but are sensitive to segmentation quality and map completeness.
2.2.2 From Indirect Retrieval to Multiple Paradigms
To overcome the fundamental limitations of retrieval-based frameworks—poor scalability, quantization errors, and lack of cross-domain interaction—researchers have proposed innovative paradigms.
Feature distribution learning: Sun et al. (F3-Net) transformed feature similarity measurement into distribution alignment using Earth Mover’s Distance. Structured relation metric: Sun et al. further decomposed features along horizontal/vertical axes to model spatial layout relations. Graph classification localization: Zhang et al. (GCG-Net) used GNN to classify images into geographic groups, reducing computational overhead. Continuous position regression: Dai et al. introduced Drone Referring Localization, where the query image interacts with a large reference image via heterogeneous spatial feature interaction to directly output a probability heatmap. This eliminates the need for a discrete database and avoids quantization errors, though resolution limits remain. Vision-language models: Chen et al. used CLIP to generate structured text descriptions for both UAV and satellite views, enabling cross-modal matching in a shared semantic space. Geometric structure alignment: Zhu et al. (LoD-Loc) proposed a differentiable neural wireframe alignment framework that maximizes consistency between predicted building line segments and projections from a Level-of-Detail city model. This approach bypasses texture dependence and directly regresses 6-DoF pose. LoD-Loc V2 extended this to low-detail city models using SAM2 for semantic segmentation and particle filtering.
Table 3 summarizes the performance of representative deep learning methods across different categories, showing the trade-offs among accuracy, recall, and efficiency.
| Category | Method | Output | Accuracy / Recall | Efficiency |
|---|---|---|---|---|
| Global appearance | Zhang et al. (2025) | Retrieval | R@1: 86.01% | High |
| Global appearance | Sun et al. (2024) | Retrieval | R@1: 77.33% | Low |
| Local feature | Liu et al. (2024) | 2D | 5.4–14.3 m (MAE) | High |
| Local feature | Hu et al. (2024) | 2D | MA@20: 94.5% | Low |
| Semantic | Ouyang et al. (2024) | 2D | ~6.6 m (MAE) | High |
| Continuous regression | Dai et al. (2022) | 2D | MA@20: 83.70% | High |
| Vision-language | Chen et al. (2025) | Retrieval | R@1: 88.18% | Low |
| Geometric alignment | Zhu et al. (2024) | 6-DoF | R@2m‑2°: 84.41% | Medium |
| Geometric alignment | Zhu et al. (2025) | 6-DoF | R@2m‑2°: 54.20% | Medium |
3. Datasets and Evaluation Standards
High-quality datasets and standardized evaluation metrics are essential for advancing AVL research. Table 4 lists major publicly available datasets that support AVL for China UAV drones. These datasets range from early cross-view matching benchmarks (University-1652, SUES-200) to trajectory-based datasets (ALTO, RealUAV, UAV-VisLoc) and specialized datasets (Boson-nighttime for thermal infrared, JointLoc for planetary, LoD-Loc for geometric alignment). Notably, datasets like RealUAV and UAV-VisLoc provide sequential flight data with IMU and GNSS ground truth, directly enabling the development and evaluation of fusion-based AVL algorithms.
| Dataset | Data Type | GNSS / IMU | Core Challenge | Size |
|---|---|---|---|---|
| University-1652 | Discrete images | No | Cross-view matching | 8.6 GB |
| SUES-200 | Discrete images | No | Multi-height, multi-scene | 5.3 GB |
| DenseUAV | Dense sampled images | No | Urban dense sampling | 16 GB |
| ALTO | Sequential images | Yes | Long-range, multi-terrain | 29 GB |
| RealUAV | Sequential images | Yes | Real-world flight, multi-terrain | – |
| UAV-VisLoc | Sequential images | Yes | Multi-height, multi-UAV | 16.4 GB |
| Boson-nighttime | Thermal infrared | No | Night/low-light, cross-modal | 69 GB |
| JointLoc | Sequential images (sim+real) | Yes | Planetary, self-similar terrain | 2.43 GB |
| LoD-Loc V1 | Sequential images | Yes (6-DoF) | Geometric alignment with LoD2/3 | 91.58 GB |
| LoD-Loc V2 | Sequential images | Yes (6-DoF) | Geometric alignment with LoD1 | – |
Evaluation metrics for AVL include: mean absolute error (MAE), root mean square error (RMSE), recall@K, mean average precision@K, and 6-DoF pose success rate (e.g., R@2m‑2°). For China UAV drone applications, computational efficiency (frame rate) and memory footprint are critical due to on-board resource constraints. Future benchmarks should integrate challenging conditions such as seasonal changes, dynamic obstacles, and long-duration flights to better reflect real-world operational scenarios.
4. Conclusion and Future Prospects
This review has systematically traced the evolution of absolute visual localization methods for China UAV drones, from traditional hand-crafted features to deep learning and emerging paradigms. Despite significant progress in accuracy, efficiency, and robustness, several key challenges remain:
1) Balance among accuracy, efficiency, and robustness: Traditional methods are efficient but sensitive to appearance changes; deep local feature methods achieve high accuracy but require heavy computation; semantic methods offer robustness but depend on prior map completeness and quality.
2) Model complexity vs. on-board resource constraints: Deep neural networks with millions of parameters are challenging to deploy on resource-limited China UAV drones. Lightweight architecture design (e.g., attention pruning, cell division) and novel paradigms (e.g., continuous regression, geometric alignment) offer potential solutions.
3) Data dependency and generalization: Current models suffer from domain shift when faced with unseen geographic areas, seasonal changes, or sensor variations. Self-supervised learning, domain adaptation, and knowledge-driven approaches (e.g., vision-language models) can mitigate this issue.
4) Fragility of single visual sensing in extreme environments: In texture-less, dynamic, or low-light conditions, visual-only methods fail. Multi-source fusion combining AVL with inertial, LiDAR, and barometric sensors is essential for reliable 6-DoF state estimation.
Moving forward, future research on AVL for China UAV drones should focus on three directions:
➤ Multi-source fusion integration: Develop adaptive fusion frameworks (e.g., factor graphs) that tightly couple AVL with VIO/VO, dynamically weighting constraints based on sensor reliability. This enables robust full-state estimation even when visual matching degrades.
➤ Lightweight and specialized architectures: Design efficient network structures (e.g., 3D attention, mobile-friendly transformers) and explore new problem paradigms that bypass expensive retrieval pipelines, such as end-to-end coordinate regression and geometric structure alignment.
➤ Cross-domain generalization: Leverage self-supervised learning to mine invariant features from unlabeled data, incorporate semantic priors via visual-language models or topological constraints from vector maps, and adopt “pre-train then fine-tune” strategies to adapt to new environments with minimal annotation effort.
Through synergistic advances in efficiency, robustness, and multi-modal integration, AVL technology for China UAV drones will evolve from a research-oriented tool into a dependable core component for autonomous aerial systems operating in GNSS-denied environments.
We acknowledge that this review covers a wide spectrum of methods, but we have not exhaustively discussed every contribution. The dynamic progress in this field suggests that future breakthroughs will likely come from hybrid approaches that combine the strengths of multiple paradigms, ultimately enabling China UAV drones to navigate reliably under any condition.