Jiayu Yuan

Satellite signal interruptions disable unmanned aerial vehicle (UAV) navigation, making visual localization a vital alternative. To improve robustness in oblique flight paths and complex environments, we propose the UAV visual geo-localization network with transformer and hypergraph feature aggregation enhancement (HTGL). First, we enhanced feature extraction capabilities by optimizing the downsampling strategy and attention allocation mechanism in the backbone network. Second, we designed the Hypergraph Feature Aggregation Enhancement (HFAE) module based on hypergraph-based adaptive correlation enhancement (HyperACE) to improve the model’s ability to capture higher-order correlations. Furthermore, we constructed the Complex Scene Rotation dataset (CSR10) and proposed a method for simulating winter scenes, thereby overcoming the limitations of existing research in terms of scenes, flight directions, and seasons. Additionally, two evaluation metrics—pixel distance root mean square error (PD-RMSE) and geographic distance root mean square error (GD-RMSE)—were introduced to enable a comprehensive assessment of algorithm performance. Experimental results show that HTGL achieved RDS scores of 85.95% (+1.4%), 83.64% (+4.1%), and 91.52% (+1.49%) on the UL14, UL14_ROT, and CSR10 datasets, respectively, demonstrating strong robustness in rotated and complex scenes. Actual deployment and flight tests on the Jetson Orin NX platform further validated the model’s excellent engineering practicality.

Achieving long-term, continuous, and accurate localization for Unmanned Aerial Vehicles (UAVs) in outdoor GNSS-denied environments where pre-existing reference maps are available is challenging. To this end, this paper proposes a Cross-view Particle Filter Localization (CPFL) framework. Unlike existing particle filter approaches that rely on inertial sensors for state propagation or sparse semantic labels for observation updates, CPFL is a vision-driven solution. This framework introduces specific adaptations into the two core stages of particle filtering: In the motion propagation stage, it achieves visual state transition by calculating a feature-based inter-frame homography mapping to estimate the 2D global relative motion components, eliminating the dependency on inertial priors; in the observation correction stage, a Dual-Granularity Adaptive Gating (DGAG) cross-view network is designed to mitigate perceptual aliasing and generate discriminative absolute position weights for the particles. By fusing these two stages through a filter mechanism, the framework transforms unbounded cumulative drift into bounded absolute localization errors. Furthermore, addressing the measurement deficiencies of traditional single-frame metrics, this paper also proposes a Trajectory Continuity Index (TCI@d) tailored for continuous localization tasks. Experiments on the real-world MAFS dataset confirm that this framework achieves a mean localization error of 5.28 m and a localization success rate of 89.7% under a 10-m threshold. Compared with mainstream vision-only algorithms and IMU-fusion baselines, this framework demonstrates lower mean errors and improved trajectory continuity, validating its effectiveness for long-term robustness.

Heterogeneous data from diverse sensor platforms can complement each other, providing more comprehensive and richer information for the analysis of complex scenarios. To effectively utilise this information for UAV geo-location tasks, accurate feature extraction and fusion methods are crucial. To address this challenge, we propose the Multi-Modal Geo Location Transformer Network (MMGLT), leveraging attention mechanisms to fuse cross-modal features. Unlike traditional dual modal fusion frameworks integrating only drone downward-view and satellite images, MMGLT innovatively uses a triple-input structure to process RGB images, LiDAR-derived auxiliary data, and satellite images. It also designs a tailored cross-attention mechanism to link visual semantics with remote sensing features f inely, enabling more robust multi-source feature learning. Subsequently, MMGLT employs a Global Attention Upsampling Head that refines features layer by layer via attention and adaptive up sampling; by combining dynamic feature re-allocation with multi scale pyramid fusion, it mitigates feature degradation and boosts representational power. With the aim of facilitating subsequent applications, we have successfully deployed the MMGLT model on the NVIDIA Jetson Orin NX compact computer. To evaluate the performance of MMGLT in the task, this paper constructs the CrossView-Van dataset. To the best of our knowledge, this is the first end-to-end geo-location dataset that integrates RGB images, high-precision remote sensing feature data, and satellite images. On this new dataset, MMGLT significantly outperforms existing single-modal methods, and achieves improvement in low-texture forest regions where traditional approaches underperform.

Vision sensor-based simultaneous localization and mapping (SLAM) systems are essential for mobile robots to locate and generate spatial models of their surroundings. However, the majority of visual SLAM systems assume static settings, leading to significant accuracy degradation in dynamic scenes. We present SGDO-SLAM, a real-time RGB-D semantic-aware SLAM framework, building upon ORB-SLAM2 to address non-static environments. Firstly, a multi-constraint dynamic rejection method from coarse to fine is proposed. The method starts with coarse rejection by combining semantic and geometric information, followed by detailed rejection using depth information, where static quality weights are quantified based on depth consistency constraints. The method achieves accurate dynamic scene perceptions and improves the accuracy of the system’s positioning. Then, a position optimization method driven by static quality weights is proposed, which prioritizes high-quality static features to enhance pose estimation. Finally, a visualized dense point cloud map is established. We performed experimental evaluations on the TUM RGB-D dataset and the Bonn dataset. The experimental results demonstrate that SGDO-SLAM reduces the absolute trajectory error performance metrics by 95% compared to the ORB-SLAM2 algorithm, while maintaining real-time efficiency and achieving state-of-the-art accuracy in dynamic scenarios.

UAVs typically rely on satellite navigation for positioning, yet this method proves ineffective in instances where the signal is inadequate or communication is disrupted. Visually based positioning technology has emerged as a reliable alternative. In this paper, we propose a novel end-to-end network, OBTPN. In the initial phase of the model, we optimized the distribution of attention within the primary network, aiming to achieve a balance between self-attention and cross-attention. Subsequently, we devised a feature fusion head, which enhanced the model’s capacity to process multi-scale information. OBTPN was successfully deployed on an NVIDIA Jetson TX2 onboard computer. This paper also proposes a high-altitude complex environment dataset, Crossview9, which addresses a research gap in the field of high-altitude visual navigation. The performance of the model on this dataset is also evaluated. Additionally, the dataset was processed to simulate a low-quality image environment to assess the model’s resilience in challenging weather conditions. The experimental results demonstrate that OBTPN_256 attains an accuracy of 84.55% on the RDS metric, thereby reaching the state-of-the-art (SOTA) level of the UL14 dataset. On the Crossview9 dataset, OBTPN_256 achieves an RDS score of 79.76%, also reaching the SOTA level. Most notably, the model’s high accuracy in low-quality image environments further substantiates its robustness in complex environments.

In high-voltage transmission lines, the tension tower needs to withstand the tension load of the overhead power line for a long time, which is prone to damage, and it is an important part of the inspection in the circuit inspection. In the modern circuit inspection process, operation and maintenance personnel mostly use unmanned aerial vehicles (UAVs) to photograph various parts of the tension tower, obtain inspection images, and manually classify and name the massive inspection images, which is low in accuracy and efficiency. Based on the above problems, this paper collects a large number of real-life UAV inspection images of various parts of a tension tower, and proposes the Stower-13 inspection image dataset, which is used to train the classification model to achieve automatic classification and naming of inspection images. Based on this dataset, this paper also proposes an improved MobileViT model, in which the Scale-Correlated Pyramid Convolution Block Attention Block (SCPCbam) module is introduced, which adds the Convolution Block Attention Module (CBAM) to the four branches of the original Scale-Correlated Pyramid Convolution (SCPC) module, so as to strengthen the ability of image multi-scale information extraction and improve the classification accuracy. This paper discusses a number of experiments on the model, and the experimental results show that the dataset proposed in this paper helps the model to understand the feature information. At the same time, the improved MobileViT model has a strong ability to extract image spatial feature information, the classification accuracy is higher than that of other models of the same type, so it is able to cope with a wide range of problems that arise in the course of practice, and it meets the practical needs of automatically naming transmission line inspection images.

Wire clamps and vibration-proof hammers are key components of high-voltage transmission lines. The wire clips and vibration-proof hammers detected in Unmanned Aerial Vehicle (UAV) power inspections suffer from small size, scarce edge information, and low recognition accuracy. To address these problems, this paper proposes a small object detection (SOD) model based on the YOLOv8n, called SOD-YOLO. Firstly, an extra small target detection layer was added to YOLOv8, which significantly improves the small target detection accuracy. In addition, in order to enhance the detection speed of the model, the RepVGG/RepConv ShuffleNet (RCS) and a OneShot Aggregation of the RCS (RCSOSA) module were introduced to replace the C2f module in the model backbone and neck shallow networks. Finally, to address the excessive focus on low-quality sample bounding boxes during model training, we introduced Wise-CIoU loss instead of CIoU loss, which improved the detection accuracy of the model. The experimental results indicate that SOD-YOLO achieved a mean average precision of 90.1%, surpassing the YOLOv8n baseline model by 7.5% while maintaining a model parameter count of 3.4 M; the inference speed reached 88.7 frames/s, which meets the requirement of real-time recognition.