Semantic-Guided Geometric Feature Extraction from Dense LiDAR for Vehicle Localization - with Abstract Maps

Oct 28, 2026

Upload by Mia

High-precision vehicle localization in GNSS-denied dense urban environments depends on robust alternatives to costly HD maps. This paper proposes a semantic-geometric pipeline that extracts map primitives from dense LiDAR and camera data, using RANSAC ground-plane extraction, semantic filtering to remove dynamic objects, and clustering with HDBSCAN plus RANSAC to isolate large-scale vertical facades. A multi-stage HITL ground-truth generation resolves annotation ambiguity and refines pose on SE(2) then to 6-DoF via point-to-plane ICP, achieving 95.04% precision and 83.74% recall.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages9

Reading Time23 min

AccessFree to download

1. Introduction & Background

Abstract Maps and Dense-to-Abstract Localization Gap

Semantic-Geometric Feature Extraction Pipeline

RANSAC ground plane extraction and semantic filtering

HDBSCAN clustering and RANSAC facade fitting

Multi-Stage Ground-Truth Generation (HITL)

SE(2) pose estimation via geometric median

6-DoF refinement and validation

Evaluation and Benchmark Results

Semantic-Guided Geometric Feature Extraction from Dense LiDAR for Vehicle Localization with Abstract Maps Mohamed Wahbah 1 , Rozhin Moftizadeh1 , Christopher Klugmann2 , Daniel Kondermann2 , Ingo Neumann1 , Hamza Alkhatib1 1 Geodetic Institute, Leibniz University Hanovver, Germany- (wahbah,moftizadeh,neumann,alkhatib)@gih.uni-hannover.de 2 Quality Match GmbH, Heidelberg Germany- (christopher.klugmann,dk)@[quality-match.com](quality-match.com) Keywords: Feature Extraction, Geo-referencing, LoD2 models, Digital Terrain Models (DTM) . Abstract 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-TheLoop (HITL) system. A robust 2D pose is computed by finding the geometric median of bootstrapped transformation sampleson 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. 1. Introduction & Background Accurate and robust global localization is a fundamental prerequisite for higher levels of driving automation, enabling core functions such as path planning and mission execution. The challenge is particularly critical in dense urban environments, where positional awareness requires accuracy on the order of several centimeters (Reid et al., 2019) . While Global Navigation Satellite Systems (GNSS) can often meet this requirement, they are rendered unreliable in dense urban environments due to signal occlusion and severe multi-path effects. However, these same environments offer a high density of stable, structured features—such as building facades, road surfaces, road signs—that can serve as geospatial landmarks for localization, providing a viable alternative to GNSS. The dominant approach for achieving the required centimeterlevel accuracy, without GNSS, relies on matching vehicle sensor data to a pre-built digital map, with High-Definition (HD) maps emerging as the state-of-the-art solution (Ghallabi et al., 2019, Chen et al., 2021) . These multi-layered representations typically contain a rich set of geometric and semantic features, including dense 3D point clouds, road networks, poles, and curbs. HD maps enable a variety of localization methods, including Light Detection and Ranging (LiDAR)-based point-voxel (Shi et al., 2020) or segment matching (Dub et al., 2017), as well as vision-based techniques that utilize features like road markings (Zhao et al., 2024) or traffic lights (Zhu et al., 2025) . Multimodal approaches fusing camera and LiDAR data also see common use with HD maps (Zuo et al., 2020) . Despite their proven efficacy, the creation and continuous maintenance of HD maps at scale incur prohibitive costs, as the map must be frequently updated to reflect changes in the real world (Lee and Ryu, 2024) . This high overhead has inspired increasing research into localization using low-detail, abstract maps that are widely available and simpler to maintain. Examples include OpenStreetMap

FAQs

Why are abstract LoD2/DTM maps used instead of HD maps for vehicle localization?

HD maps require frequent costly updates to stay consistent with real-world changes. Abstract maps offer a lower-detail representation that is simpler to maintain while still providing structured landmarks for localization.

How does the proposed pipeline extract features from dense LiDAR and camera data?

It performs RANSAC-fitted ground plane extraction, applies a semantic filter to discard dynamic objects, then uses geometric clustering (HDBSCAN) and RANSAC plane fitting to isolate large-scale vertical facades.

How is ground truth (GT) generated and pose refined for evaluation?

A multi-stage HITL system resolves annotation ambiguity, computes a robust 2D pose using the geometric median of bootstrapped SE(2) transformation samples, refines it to a 6-DoF pose using point-to-plane ICP, and validates it with a human. The resulting benchmark evaluation reports 95.04% precision and 83.74% recall.

Oct 28, 2026

Upload by Mia