SIO-Mapper: A Framework for Lane-Level HD Map Construction Using Satellite Images and OpenStreetMap with No On-Site Visits

TL;DR

Lane-level HD maps enable precise vehicle localization but traditional pipelines require on-site MMS data, limiting geographic coverage. SIO-Mapper constructs city-scale lane-level HD maps from publicly available satellite images and OpenStreetMap, using $L_i=\phi(s_i,c_i)$ with $\phi=D(E_t(\cdot),E_c(\cdot))$ to generate lane images, and a novel Mapper that merges segments via a hybrid clustering-graph approach. It introduces standardized vector-based evaluation metrics and demonstrates robust performance on NAVER Labs Open Dataset and NuScenes, achieving broad coverage and competitive accuracy across diverse urban environments. These contributions offer scalable, automated lane-level map generation suitable for localization and planning in real-world autonomous driving contexts.

Abstract

High-definition (HD) maps, particularly those containing lane-level information regarded as ground truth, are crucial for vehicle localization research. Traditionally, constructing HD maps requires highly accurate sensor measurements collection from the target area, followed by manual annotation to assign semantic information. Consequently, HD maps are limited in terms of geographic coverage. To tackle this problem, in this paper, we propose SIO-Mapper, a novel lane-level HD map construction framework that constructs city-scale maps without physical site visits by utilizing satellite images and OpenStreetmap data. One of the key contributions of SIO-Mapper is its ability to extract lane information more accurately by introducing SIO-Net, a novel deep learning network that integrates features from satellite image and OpenStreetmap using both Transformer-based and convolution-based encoders. Furthermore, to overcome challenges in merging lanes over large areas, we introduce a novel lane integration methodology that combines cluster-based and graph-based approaches. This algorithm ensures the seamless aggregation of lane segments with high accuracy and coverage, even in complex road environments. We validated SIO-Mapper on the Naver Labs Open Dataset and NuScenes dataset, demonstrating better performance in various environments including Korea, the United States, and Singapore compared to the state-of-the-art lane-level HD mapconstruction methods.

Figures (7)

• Figure 1: Conventional lane-level HD maps only cover very limited portion of the entire road. For example, in Seoul which contains 5470.67 km of roads, only 63.21 km which is 1.16% of the road is covered as shown in this figure.
• Figure 2: Overall pipeline of the proposed SIO-Mapper. First, SIO-Net generate lane images. Transformer encoder and convolutional encoder each extract features in satellite images. Then, decoder concatenate both features and generate lane images. Then, mapper integrates lane images using cluster-based mapper and graph-based mapper and merge them using map merger to construct lane-level HD map. Using SIO-Mapper, we can construct a lane-level HD map in city-scale without any physical visit to the target area.
• Figure 3: This figure illustrates the pipeline of the suggested network, SIO-Net. In preprocess step, we mask the satellite image using public map mask and calculate road shape embedding by approximating the shape of road into quadratic function. Then, we extract image features using two different encoders, Transformer encoder and convolutional encoder, and concatenate both features. Finally, the decoder generate a lane image by decoding the concatenated feature. Lane image is consists of three labels which are respectively color coded in red, green, and blue: white broken lanes, white lanes, and yellow lanes.
• Figure 4: Result at three challenging cases: under overpass, under shadow, and complex lane structure. Even a query image is partially blocked by diverse obstacle, OST still can generate lane image well. The last case, complex lane structure, does not perform perfectly, but they still perform properly except road boundary, where the vehicle actually drives.
• Figure 5: Map accuracy, $A_m$, based on different thresholds for each sequence. Since the NAVER LABS Open Dataset does not provide LiDAR and camera data for the Yeouido sequence, there are no results using HDmapNet for the Yeouido sequence.