Conference Agenda

Accurate large-scale dense 3D reconstruction is fundamental for geospatial mapping, robotics, and autonomous navigation. Conventional photogrammetric workflows can reconstruct environments from ground-level imagery but often suffer from cumulative drift over kilometer-scale trajectories and require extensive calibration. Recent advances in feed-forward 3D reconstruction, notably the Visual Geometry Grounded Transformer (VGGT), have demonstrated the ability to generate dense point clouds directly from RGB images without explicit optimization. VGGT jointly estimates camera poses, depth, and dense geometry from multiple uncalibrated frames in a single forward pass. However, its scalability is limited by two factors: (1) the lack of absolute metric scale and (2) high GPU memory demands.

Many national mapping agencies (e.g., USGS, IGN, Ordnance Survey) have released Airborne Laser Scanning (ALS) datasets covering vast urban and rural areas. These high-quality aerial point clouds provide globally consistent, metrically referenced data that can serve as external constraints for ground-level reconstructions. Building upon this opportunity, we propose VGGT-ALS, a framework that leverages open ALS point clouds to enable VGGT-based systems to produce large-scale, metrically accurate dense maps from mobile mapping imagery.

High-precision vehicle localization in GNSS-denied urban areas requires alternatives to costly HD maps. In this paper, we present a novel framework for feature extraction and benchmark generation to enable high-precision localization using abstract LoD2/DTM maps as a replacement for HD maps. Our first contribution, a semantic-geometric pipeline, processes dense LiDAR and camera data to extract map primitives. This is accomplished by a RANSAC-fitted ground plane extraction step, followed by a semantic filter that discards dynamic objects. Finally, geometric clustering (HDBSCAN) and RANSAC plane fitting isolate large-scale vertical facades. Our second contribution, a multi-stage GT generation framework, resolves annotation ambiguity using a Human-In-The-Loop (HITL) system. A robust 2D pose is computed by finding the geometric median of bootstrapped transformation samples on the SE(2) manifold, which is then refined to a 6-Degree-of-Freedom pose via point-to-plane ICP, before being validated by a human for a final check. We evaluated our feature extraction pipeline against the generated benchmark, achieving 95.04% precision and 83.74% recall. An analysis of this performance shows the pipeline correctly rejects small, ambiguous features while achieving high recall on all large, stable features, proving its suitability for a robust localization filter.

As civil infrastructure in South Korea ages, the demand for systematic monitoring has grown. While Unmanned Aerial Vehicles (UAVs) provide a safer alternative to manual inspections, Global Navigation Satellite System (GNSS) signal degradation beneath bridge structures remains a critical barrier to autonomous flight. Unlike existing hardware-centric or SLAM-based solutions that require high costs and computational overhead, this study proposes a robust, algorithm-based path planning methodology using a 3D spatial grid framework.

Two strategies were evaluated through field tests at the Bukhangang Bridge: the Photography Point Method (PPA) and the GNSS Non-Shadowing Area Method (WPS-GNSA). Results demonstrated that WPS-GNSA significantly enhances signal reliability by strategically positioning waypoints outside electromagnetic shadow zones. At an optimal 19-meter separation distance, WPS-GNSA maintained GNSS Level 5 connectivity—the threshold for recommended autonomous flight—for 9.4% of the duration and Level 4 or higher for 79.7%, whereas the PPA peaked at Level 4.

These findings indicate that WPS-GNSA enables reliable autonomous inspections using standard commercial drones without specialized hardware modifications. While the current model relies on pre-existing digital blueprints, future research will integrate AI-based navigation and real-time environment perception to enhance scalability and adapt to dynamic obstacles in complex infrastructure environments.

Accurate and reliable positioning is fundamental for the development of a wide number of applications. Despite in most of the regular working conditions the use of a GNSS receiver is sufficient for properly solving the problem, determining a reliable solution in challenging conditions can be difficult. In such conditions, exploiting the information shared by different sensors and platforms can be useful for reliably determining the platforms' positions. In this work, both ground and aerial platforms are considered: each platform is assumed to be provided with communication capabilities, which can be exploited to share its knowledge. Since GNSS positioning is usually less effective at ground level than on a flying platform, the aerial platforms are assumed to be provided with good GNSS-based positioning information. Instead, GNSS is assumed to be unavailable to the ground vehicles, which, instead, can use LiDAR/visual odometry for dead-reckoning positioning, UWB inter-platform ranging for relative positioning, and camera-based positions, provided by aerial platforms, for assessing their georeferenced positions. This work focuses on assessing the positioning performance when exploiting vision-based information about the georeferenced ground vehicle positions from a camera mounted on a UAV. The camera acquired oblique views of the scene while moving over the case study area during the test. YOLO was used to detect cars from the image frames and the vehicle coordinates have been extracted from 3D reconstructions obtained from the MoGe-2 network. Average errors at meter level on the determined georeferenced coordinates were obtained when combining UWB vehicle-to-vehicle ranges with MoGe-2 reconstructions.

Reliable positioning and mapping in dense urban environments remain challenging due to signal blockage, multipath, and dynamic scenes. Progress on multi-sensor integrated positioning and visual/lidar SLAM has been driven by open datasets, yet most existing resources are either perception-centric with limited raw navigation data, focused on controlled environments, or built around outdated software platforms and/or data formats. In this paper, we present the Melbourne Multi-Sensor Urban Positioning and Mapping Dataset, a new resource targeting urban vehicle navigation and mapping tasks. The dataset was collected using a custom mobile mapping platform equipped with a tactical-grade INS, a survey-grade Leica GNSS receiver, a low-cost UBLOX GNSS receiver, a high-resolution Ouster OS1 128 lidar, and four industrial FLIR cameras providing 360° coverage. Seven data collection trips were recorded on dynamic streets in several inner suburbs of Melbourne, including multiple closed loops and a repeated route with day–night variation. For better compatibility and future-proofing, all raw data are provided as standard ROS2 message streams in MCAP format, complemented by commonly used individual formats and GNSS products for multi-sensor integrations. We benchmark three GNSS--based positioning packages (RTKLib, Net_Diff and Ginan) and four state-of-the-art lidar(-inertial) odometry/SLAM methods (FAST-LIO2, KISS-ICP, KISS-SLAM and PIN-SLAM), demonstrating the applicability and compatibility of our dataset for modern positioning and mapping software pipelines. The dataset is designed as a robust, ROS2-native testbed for research on GNSS/IMU/lidar/camera fusion for the testing and validation of vehicle positioning and mapping in urban environments, which is available open-source at https://github.com/zjjdes/melbourne_dataset.

Accurate and robust localization is essential for autonomous vehicles and mobile robots operating in complex, dynamic environments. However, existing learning-based LiDAR-inertial odometry (LIO) methods typically rely on simple weighted fusion or purely global attention, which may not fully exploit cross-modal complementarity. In this paper, we propose CMLGF-LIO, a cross-modal local-global fusion framework that improves LIO accuracy and robustness. At the local level, we design a Local Split-Attention (LSA) module that injects IMU-derived motion priors into local LiDAR feature groups and adaptively allocates attention weights, suppressing redundant information while preserving discriminative local geometry for fine-grained fusion. At the global level, we introduce a Global MLP-Mixer (GMM) module that aligns LiDAR and IMU token sequences and models global cross-modal interactions using an MLP-Mixer backbone. Experiments demonstrate that CMLGF-LIO is more robust than learning-based baselines under challenging conditions, and ablation studies validate the effectiveness of the proposed local-global fusion strategy.

This paper presents a low-cost mobile mapping system built on a consumer vehicle (2026 Tesla Model Y) equipped with a roof-mounted Velodyne VLP-16 LiDAR and three post-processed kinematic (PPK) GNSS receivers. A processing pipeline was developed that fuses scan-to-scan LiDAR registration via KISS-ICP with multi-receiver PPK trajectories to produce georeferenced 3D point clouds. The system was tested on a 2 km loop on the Ohio State University campus. KISS-ICP achieved a registration fitness of 1.0 across all 3,400+ frames, producing locally crisp point clouds. A yaw-only Procrustes alignment followed by interactive 7-parameter refinement maps the SLAM trajectory into an East-North-Up geodetic frame. We document the complete pipeline architecture, including automated GPS time synchronization, multi-receiver vehicle pose estimation, and a streaming LAS export capable of handling 70+ million points. We systematically evaluate ten post-hoc trajectory correction strategies and identify a fundamental trade-off between inter-frame consistency (point cloud crispness) and absolute geodetic accuracy. The primary unresolved challenge is a staircase artifact caused by ~40 m of accumulated SLAM drift over the loop, which cannot be corrected without degrading local registration quality. We conclude that loop closure detection and pose graph optimization within the SLAM pipeline are necessary to resolve this tension and outline a path toward survey-grade mobile mapping from consumer vehicle platforms.