Source-linked AI summary
Building Lane-Level Maps from Aerial Images
Jiawei Yao, Xiaochao Pan, Tong Wu, Xiaofeng Zhang
TL;DR
Aerial-image lane-level mapping is less developed than vehicle-view lane detection, although it targets offline maps for autonomous driving. The paper introduces an aerial lane dataset and a two-stage vertex-matching method, and reports consistent outperformance of state-of-the-art methods.
Problem
Prior lane-detection datasets focus on vehicle-perspective imagery, while aerial-image lane-level polylines for offline maps are a comparatively underdeveloped research target.
Source: §1 · p. 1, §2.1 · pp. 1–2Method
The paper introduces the AEL aerial lane benchmark and AerialLaneNet, a two-stage approach that predicts lane segmentation and vertices, then matches each vertex to its successor.
Source: §1 · p. 1, §2.1 · p. 2, §3 · p. 3Results
AerialLaneNet consistently outperforms state-of-the-art methods in experiments, and excels in precision, recall, and F1-score across thresholds.
Source: §5 · p. 4, §4.2 · p. 4Takeaways & Limitations
The authors conclude their study can be used in online automatic mapping for autonomous-driving and intelligent-transportation applications.
Source: §5 · p. 4
Abstract
from arXiv · showhide
Detecting lane lines from sensors is becoming an increasingly significant part of autonomous driving systems. However, less development has been made on high-definition lane-level mapping based on aerial images, which could automatically build and update offline maps for auto-driving systems. To this end, our work focuses on extracting fine-level detailed lane lines together with their topological structures. This task is challenging since it requires large amounts of data covering different lane types, terrain and regions. In this paper, we introduce for the first time a large-scale aerial image dataset built for lane detection, with high-quality polyline lane annotations on high-resolution images of around 80 kilometers of road. Moreover, we developed a baseline deep learning lane detection method from aerial images, called AerialLaneNet, consisting of two stages. The first stage is to produce coarse-grained results at point level, and the second stage exploits the coarse-grained results and feature to perform the vertex-matching task, producing fine-grained lanes with topology. The experiments show our approach achieves significant improvement compared with the state-of-the-art methods on our new dataset. Our code and new dataset are available at https://github.com/Jiawei-Yao0812/AerialLaneNet.
1. INTRODUCTION
This introduction identifies a gap in offline lane mapping from aerial images and presents a new benchmark and the AerialLaneNet baseline to address it.
- Lane detection is described as essential to autonomous-driving perception, with results that may affect driving performance and personal safety.Source: §1 · p. 1
- Prior work focuses on online lane detection from vehicle sensors or coarse road-network and curb maps, while offline aerial-image lane mapping lacks a benchmark and baseline.Source: §1 · p. 1
- The benchmark provides 7,763 high-resolution aerial images and over 150,000 labeled lanes spanning lane standards, terrain, regions, and lane types.Source: §1 · p. 1
- AerialLaneNet is reported to outperform state-of-the-art methods in comprehensive experiments.Source: §1 · p. 1
- AerialLaneNet uses two cascaded stages: point-level coarse results and features inform vertex matching to produce fine-grained, instance-level lanes.Source: §1 · p. 1
2. AERIAL LANE (AEL) DATASET
AEL is introduced as the first aerial-image dataset with lane-level annotations, designed to support offline autonomous-driving maps with accurate lane polylines. It provides a benchmark and baseline for detecting vectorized lanes in high-resolution aerial imagery.
2.1. Overview
- AEL is presented as the first aerial-image dataset with lane-level annotations, producing accurate lane polylines for offline maps used in autonomous-driving planning and navigation.The images come from Google Map service, and QGIS was used to label lane lines.Source: §2.1 · p. 2
2.2. Annotation Specification
- Lane lines are annotated as directed polylines whose vertices use latitude and longitude coordinates, with adjacent vertices chosen to approximate curved lanes.Annotations also record whether a lane is single or double, white or yellow, and solid or dashed, along with road and lane IDs; exports include 1280 × 1280 image patches, JSON annotations, and QGIS lane masks.Source: §2.2 · p. 2
2.3. Dataset Statistics
- The dataset covers 11 regions, with roads 3–27 kilometers long, and includes 7,763 images and over 150,000 lane lines split 7:2:1 into training, validation, and test sets.Regions span varied backgrounds and terrain, including Dubai coastline desert, Valencia urban areas, and forest in Perak State.Source: §2.3 · pp. 2–3
3. THE PROPOSED BASELINE: AERIALLANENET
AerialLaneNet uses a two-stage pipeline: it predicts lane segmentation and vertex maps, then matches vertices to produce instance-level lane lines. Its matching stage combines nearby-vertex information with high-resolution semantic features to classify lane connections and correct vertex locations.
3. THE PROPOSED BASELINE: AERIALLANENET
- AerialLaneNet first uses HRNet to produce lane segmentation and high-resolution features, then predicts vertex maps before matching each vertex to its next vertex.The two stages use segmentation, features, and vertex maps to produce instance-level lane lines with topology.Source: §3 · p. 3
3.1. Vertex Prediction
- Vertices are encoded as Gaussian heatmaps after downsampling coordinates by stride 4, and the vertex prediction objective is similar to focal loss.When Gaussian heatmaps overlap, the method takes their maximum rather than their average to preserve peak-point relationships.Source: §3.1 · p. 3
3.2. Vertex Matching
- For each vertex, the matching stage selects its top-K Euclidean-nearest candidates and uses cropped, aggregated segmentation, vertex-map, and feature inputs to classify the next vertex and correct its location.The classification branch uses cross-entropy loss, while the location regression branch uses smooth L1 loss.Source: §3.2 · p. 3
4. EXPERIMENTS
AerialLaneNet outperforms the compared methods across precision, recall, and F1-score at different thresholds. Increasing the number of vertex candidates improves prediction metrics but increases runtime.
4.1. Implementation Details
- Training uses randomly cropped 1280 × 1280 patches with random flips, HRNet for segmentation and feature extraction, and ResNet-50 for next-vertex prediction.The network is trained with Adam for 100 epochs at an initial learning rate of 1e −3, momentum 0.9, batch size 4, and cosine annealing; training took around 16 hours.Source: §4.1 · p. 4
4.2. Results and Comparisons
- AerialLaneNet outperforms all compared methods in precision, recall, and F1-score across different thresholds.The paper describes qualitative results as showing that the method generally restores lane topology, with some predicted lanes warped by vertex-location shifts.Source: §4.2 · p. 4
- The method connects most predicted vertices correctly and restores lane topology, though vertex-location shifts can warp predicted lanes.The qualitative evaluation compares predictions with the image and ground-truth lanes and vertices.Source: §4.2 · p. 4
- Increasing K reduces next-vertex regression MSE and raises next-vertex classification F1-Score, but runtime increases linearly.The runtime increase is attributed to the additional classification burden.Source: §4.2 · p. 4
5. CONCLUSION
The paper presents a public aerial-image lane dataset and AerialLaneNet, a baseline for detecting lane lines from aerial images. Experiments report consistent outperformance of state-of-the-art methods, and the study can support online automatic mapping for autonomous driving and intelligent transportation.
- The proposed method consistently outperforms state-of-the-art methods in extensive experiments.The paper says the dataset and study can support online automatic mapping for autonomous driving and intelligent transportation applications.Source: §5 · p. 4
- The paper proposes an aerial-image dataset with varied backgrounds and fine-grained lane types, to be made publicly available upon publication.Source: §5 · p. 4
- AerialLaneNet is introduced as a baseline method for detecting lane lines from aerial images.Source: §5 · p. 4