Online Mapping for Autonomous Driving: Addressing Sensor Generalization and Dynamic Map Updates in Campus Environments

TL;DR

HD maps are critical for accurate localization but costly to create and maintain, especially in dynamic campus environments. The authors deploy a sensor-generalizable online mapping system based on SemVecNet/SemVecMap on a campus golf cart with dual cameras and LiDAR to generate and continually update HD maps. They integrate a 3D ground-truth labeling pipeline, environment-specific fine-tuning, and a semi-automatic map-update process that fuses multi-frame predictions to detect changes and revise the map. Experiments demonstrate substantial cross-domain generalization improvements after campus fine-tuning and show real-world map generation in unseen areas with adaptive updates, highlighting the practical potential of online HD-map maintenance.

Abstract

High-definition (HD) maps are essential for autonomous driving, providing precise information such as road boundaries, lane dividers, and crosswalks to enable safe and accurate navigation. However, traditional HD map generation is labor-intensive, expensive, and difficult to maintain in dynamic environments. To overcome these challenges, we present a real-world deployment of an online mapping system on a campus golf cart platform equipped with dual front cameras and a LiDAR sensor. Our work tackles three core challenges: (1) labeling a 3D HD map for campus environment; (2) integrating and generalizing the SemVecMap model onboard; and (3) incrementally generating and updating the predicted HD map to capture environmental changes. By fine-tuning with campus-specific data, our pipeline produces accurate map predictions and supports continual updates, demonstrating its practical value in real-world autonomous driving scenarios.

Figures (5)

• Figure 1: Campus construction zone; regions marked with red crosses indicate areas where data collection was not possible.
• Figure 2: Overview: The blue section represents the 3D ground truth map labeling pipeline, the orange section denotes the golf cart platform, and the green section corresponds to the online mapping SemVecNetSemVecNet pipeline. The orange arrows illustrate the interaction steps between the map labeling and online mapping pipelines for model fine-tuning and map updates.
• Figure 3: Four main annotated regions in the campus map. Green polylines denote road boundaries, red polylines represent lane dividers, and blue polygons indicate pedestrian crosswalks. Key regions of interest are highlighted—yellow represents the roundabout, blue indicates the intersection, green indicates the loop, and pink denotes the multi-lane two-way road.
• Figure 4: Qualitative visualization of the fine-tuned online mapping model across various driving scenarios in the campus environment. Each triplet shows the semantic map input (left), the predicted vectorized map (middle), and the ground truth annotation (right). The examples include straight roads (a)(e), intersections(b)(f), loop(c), and roundaboud(d).
• Figure 5: New map generation in unseen areas using our fine-tuned online mapping model. Top: heatmaps of boundaries, dividers, and crosswalks accumulated from multiple frames, with the final vectorized map extracted from high-confidence regions. Bottom: zoom-in views of two intersections, comparing the prediction map with the old satellite map and the current camera views.