Visual geo-localization can achieve UAVs (Unmanned Aerial Vehicles) position during GNSS (Global Navigation Satellite System) denial or restriction. However, The performance of visual geo-localization is seriously impaired by illumination variation, different scales, viewpoint difference, spare texture, and computer power of UAVs, etc. In this paper, a fast detector-free two-stage matching method is proposed to improve the visual geo-localization of low-altitude UAVs. A detector-free matching method and perspective transformation module are incorporated into the coarse and fine matching stages to improve the robustness of the weak texture and viewpoint data. The minimum Euclidean distance is used to accelerate the coarse matching, and the coordinate regression based on DSNT (Differentiable Spatial to Numerical) transform is used to improve the fine matching accuracy respectively. The experimental results show that the average localization precision of the proposed method is 2.24 m, which is 0.33 m higher than that of the current typical matching methods. In addition, this method has obvious advantages in localization robustness and inference efficiency on Jetson Xavier NX, which completed to match and localize all images in the dataset while the localization frequency reached the best.

Unmanned Aerial Vehicles (UAVs) require the ability to robustly perceive surrounding scenes for autonomous navigation. The semantic reconstruction of the scene is a truly functional understanding of the environment. However, high-performance computing is generally not available on most UAVs, so a lightweight real-time semantic reconstruction method is necessary. Existing methods rely on GPU, and it is difficult to achieve real-time semantic reconstruction on CPU. To solve the problem, an indoor dense semantic Simultaneous Localization and Mapping (SLAM) method using CPU computing is proposed in this paper, named CDSFusion. The CDSFusion is the first system integrating RGBD-based Visual-Inertial Odometry (VIO), semantic segmentation and 3D reconstruction in real-time on a CPU. In our VIO method, the depth information is introduced to improve the accuracy of pose estimation, and FAST features are used for faster tracking. In our semantic reconstruction method, the PSPNet (Pyramid Scene Parsing Network) pre-trained model is optimized to provide the semantic information in real-time on the CPU, and the semantic point clouds are fused using Voxblox. The experimental results demonstrate that camera tracking is accelerated without loss of accuracy in our VIO, and a 3D semantic map is reconstructed in real-time, which is comparable to one generated by the GPU-dependent method.

Real-time 3D reconstruction combined with MAVs has garnered significant attention in a variety of fields, including building maintenance, geological exploration, emergency rescue, and cultural heritage protection. While MAVs possess the advantages of speed and lightness, they also exhibit strong image blur and limited computational resources. To address these limitations, this paper presents a novel approach for onboard, depth-only, real-time 3D reconstruction capable of accommodating fast-moving MAVs. Our primary contribution is a dense SLAM system that combines surface hierarchical sparse representation and particle swarm pose optimization. Our system enables the robust tracking of high-speed camera motion and facilitates scaling to large scenes without being constrained by GPU memory resources. Our robust camera tracking framework is capable of accommodating fast camera motions and varying environments solely by relying on depth images. Furthermore, by integrating path planning methods, we explore the capabilities of MAV autonomous mapping in unknown environments with restricted lighting. Our efficient reconstruction system is capable of generating highly dense point clouds with resolutions ranging from 2 mm to 8 mm on surfaces of different complexities at rates approaching 30 Hz, fully onboard a MAV. We evaluate the performance of our method on both datasets and real-world platforms and demonstrate its superior accuracy and efficiency compared to existing methods.