The under-canopy environment, which is inherently inaccessible to humans, necessitates the use of unmanned aerial vehicles (UAVs) for data collection. The implementation of UAV autonomous navigation in such environments faces challenges, including dense obstacles, GNSS signal interference, and varying lighting conditions. This paper introduces a UAV autonomous navigation method specifically designed for under-canopy environments. Initially, image enhancement techniques are integrated with neural network-based visual feature extraction. Subsequently, employs a high-dimensional error-state optimizer coupled with a low-dimensional height filter to achieve high-precision localization of the UAV in under-canopy environments. Furthermore, proposes a boundary sampling autonomous exploration algorithm and an advanced Rapidly exploring Random Tree (RRT) path planning algorithm. The objective is to enhance the reliability and safety of UAV operations beneath the forest canopy, thereby establishing a technical foundation for surveying vertically stratified natural resources.

Real-time analysis and mining of vehicle GPS data is essential for effective traffic regulation. Currently, most vehicle analysis algorithms based on GPS data are designed for static datasets, with fewer algorithms addressing dynamic streaming GPS data. Moreover, the limited number of real-time algorithms primarily focus on public transportation vehicles such as taxis and buses. There remains a significant gap in the analysis of specialized vehicles like dump trucks, which fails to meet the regulatory needs for monitoring these vehicles. To address this, this paper proposes a method for identifying driving states and detecting collision risks for dump trucks based on real-time GPS stream data. First, by partitioning the data, the method enables the separate calculation and identification of various operational states of different vehicles, such as their location, speed, and direction. Second, we partition the data based on vehicle positions to detect potential collision risks among vehicles in nearby areas. Experimental results show that the data throughput reaches 25,000 and 66,000 records per second for each method, with a data skew rate controlled below 0.1, demonstrating the method’s efficiency in real-time driving state recognition and collision risk detection for dump trucks.

Achieving real-time indoor 3D reconstruction using low-cost devices in GNSS-denied environments is highly valuable for applications such as indoor navigation, indoor spatial measurement, and augmented reality guidance. However, vision–IMU-based real-time indoor 3D reconstruction inevitably involves repeated scanning of the same scene regions. When pose inconsistency exists among repeated observations, the same structures may be incorrectly fused into different locations, leading to artifacts such as surface layering and ghosting, which are further amplified by accumulated estimation errors. These artifacts are not caused by errors in a single module, but are closely related to pose estimation, topological relationships, and reconstruction error propagation, which severely degrade the geometric reliability, measurability, and task usability of indoor maps. To address these issues, this paper proposes an end-to-end real-time 3D reconstruction framework that reduces artifacts from three aspects: encoder enhancement, structural relationship constraints, and adaptive reconstruction. Specifically, IMU measurements are encoded into tokens and injected into a Vision Transformer to improve pose estimation accuracy. Then, a Fiedler-vector-based topological region partitioning strategy is introduced to incorporate global scene connectivity into the end-to-end pipeline, thereby limiting artifact propagation across different regions. Finally, each segmented region is reconstructed through adaptive neural point-cloud reconstruction and global optimization to further improve local geometric consistency. Extensive quantitative, qualitative, and ablation experiments are conducted on multiple self-collected and public datasets, together with a comprehensive comparison against existing state-of-the-art vision-based end-to-end and 3DGS-based indoor reconstruction methods. The experimental results demonstrate that the proposed method achieves clear advantages in artifact suppression, geometric accuracy, reconstruction quality, and real-time performance. This method provides a new solution for improving the reliability and practical usability of real-time indoor 3D reconstruction in GNSS-denied environments. The source code and supplementary data are available at: https://github.com/starlingonearth/end-to-end-artifact-suppression-real-time-3d-reconstruction.