Inland-LOAM: Voxel-Based Structural Semantic LiDAR Odometry and Mapping for Inland Waterway Navigation

TL;DR

Inland-LOAM tackles the challenge of GNSS-denied navigation on inland waterways by integrating a LiDAR-only SLAM framework with a water-surface planar constraint and a voxel-based semantic map pipeline. The two main modules—robust LOAM for pose estimation and a semantic-conversion stage that yields a 2D structural semantic map and IENC-compatible outputs—enable real-time estimation of bridge clearances and shoreline boundaries. Key contributions include the variance-based feature extraction tailored to waterway geometry, a joint optimization that incorporates a global water plane, and an end-to-end 3D-to-2D map transformation with IENC export capabilities. Experimental results on real-world datasets show improved localization accuracy and semantically rich maps compared to several state-of-the-art baselines, supporting safer and more efficient autonomous inland navigation.

Abstract

Accurate geospatial information is crucial for safe, autonomous Inland Waterway Transport (IWT), as existing charts (IENC) lack real-time detail and conventional LiDAR SLAM fails in waterway environments. These challenges lead to vertical drift and non-semantic maps, hindering autonomous navigation. This paper introduces Inland-LOAM, a LiDAR SLAM framework for waterways. It uses an improved feature extraction and a water surface planar constraint to mitigate vertical drift. A novel pipeline transforms 3D point clouds into structured 2D semantic maps using voxel-based geometric analysis, enabling real-time computation of navigational parameters like bridge clearances. An automated module extracts shorelines and exports them into a lightweight, IENC-compatible format. Evaluations on a real-world dataset show Inland-LOAM achieves superior localization accuracy over state-of-the-art methods. The generated semantic maps and shorelines align with real-world conditions, providing reliable data for enhanced situational awareness. The code and dataset will be publicly available

Figures (15)

• Figure 1: Our experimental platform, a catamaran named the Maverick, was equipped with a Sensor Box for situational awareness and data collection. Image source: Wouters2023.
• Figure 2: Overview of the Inland-LOAM framework, which processes LiDAR data to generate IENC-compatible structural semantic maps for inland waterways. The LOAM pipeline (upper section) performs robust and low-drift LiDAR odometry. The semantic interpretation and conversion module (lower section) then uses this map to build a Voxel Map, analyze geometric properties for structural semantic understanding, and produce a 2D map converted into an IENC-compatible format.
• Figure 3: The layout of sensors on our self-developed Sensor Box.
• Figure 4: Comparative performance of feature extraction. Purple points indicate planar features, while green points denote rough features. The upper figure illustrates the feature distribution by our method, where semantic information of the targets is annotated, including building facades, trees, shore, and water surface. Notably, the shore is largely covered by vegetation, with only a small portion appearing as a flat surface. The lower figure shows the feature distribution of LOAM-based methods.
• Figure 5: Inland-LOAM factor graph structure, showing prior, water surface, LiDAR odometry, loop closure factors.