Beicheng Li

In complex environments such as cities and canyons, cross-view matching of UAV and satellite imagery serves as a complementary approach to self-localization in the absence of GNSS signals. To address the limitations of existing datasets—namely their limited variety of scenes and low annotation accuracy—we have constructed the Cross-View, Multi-Altitude, Multi-Scene UAV Localization Dataset (CMHUAV). This dataset covers nine adjacent, non-overlapping subregions in Hangzhou and comprises 8625 precisely matched pairs of drone and satellite images, providing high-precision supervision for cross-view feature learning. Furthermore, we designed a Transformer-based global feature enhancement localization model (SLUG) equipped with a lightweight Adaptive Part Weighting (APW) module, which learns part-level scalar importance weights to dynamically recalibrate the contribution of each annular region. By integrating adaptive part-level weighting with a joint supervision strategy, we significantly enhance SLUG’s spatial discrimination capability. We also propose the Height-aware Normalized Distance Matching (HNDM@K) metric, which unifies retrieval ranking quality and geospatial bias within a single framework, enabling joint evaluation of elevation-adaptive localization accuracy and retrieval performance. Experiments demonstrate that, on the CMHUAV dataset, SLUG achieves a recall of 93.53% and a mean precision of 93.9%; under the HNDM@1 metric, the localization scores at all three altitude levels exceed 94.8%. These results demonstrate that the proposed method, evaluated via offline visual retrieval, exhibits strong potential as a vision-based localization module for UAVs under GNSS disruption. Nevertheless, closed-loop flight tests and system-level localization performance still require further validation.

Traditional visual SLAM systems often suffer from localization drift in dynamic environments due to interference from moving objects. Although semantic segmentation and depth-based masking methods have improved performance, they may still suffer from boundary under-segmentation and missed detections due to truncation of dynamic objects. To address these challenges, we propose a cascaded framework, DMSG-SLAM, a cascaded visual SLAM system that fuses Depth-Mask, Semantic information and Geometry constraints for dynamic environments. A lightweight object detection network, combined with depth consistency, is first employed to generate instance-like masks for preliminary dynamic feature removal. Then, a rotation-aware local epipolar geometric filtering mechanism is introduced to suppress residual features near object boundaries and mitigate perceptual blind spots caused by occlusion or truncation. Within potential dynamic regions, the epipolar threshold is adaptively switched according to the estimated inter-frame rotation to provide a more conservative filtering effect under challenging motion conditions. In addition, a TSDF-based dense volumetric map is incorporated to reconstruct more consistent surfaces. Experiments on highly dynamic sequences from the TUM RGB-D dataset indicate that DMSG-SLAM achieves competitive accuracy in dynamic environments, with localization performance improving by up to 90% compared to ORB-SLAM2.

Simultaneous Localization and Mapping (SLAM) is a fundamental methodology that underpins autonomous navigation in robotic systems. Conventional approaches perform well in static environments but rely on the assumption of environmental rigidity, which leads to significant accuracy degradation in dynamic environments. To address this challenge, this study presents RTS-SLAM, a real-time semantic visual SLAM system designed for dynamic environments. Based on the ORB-SLAM2 framework, a multi-layer, constraint-driven dynamic feature-rejection strategy is introduced. The proposed approach first removes dynamic features by combining semantic information with geometric constraints. Subsequently, residual dynamic points are eliminated via trajectory-consistency constraint analysis, thereby effectively improving localization accuracy. Furthermore, a dense mapping strategy featuring global sparsification and critical region refinement is proposed. By reducing redundancy in the dense point cloud, the method decreases memory usage while preserving important object geometries. Experimental evaluations on the TUM RGB-D and Bonn datasets indicate that RTS-SLAM reduces the average absolute trajectory error by more than 95% compared with ORB-SLAM2 in dynamic environments. Meanwhile, the system maintains real-time performance and achieves high localization accuracy in dynamic environments.