Existing orthophoto maps based on satellite or aerial photography often lack the precision required for detailed road marking. This paper presents a robust pipeline for generating ultra-high-resolution road maps using vehicle-mounted sensors. By combining a novel offline trajectory interpolation method with wavelet-based image fusion, the researchers achieved a mapping resolution of less than 2 centimeters per pixel—sharp enough to automate the identification of road signs and lane markings.
The Challenge of High-Resolution Road Mapping
Standard orthophoto maps rely on satellite or aerial photography. While useful for general navigation, these images lack the resolution required for precise "road marking"—the process of identifying lane lines, arrows, and street-level signs.
Current high-resolution road data often requires manual labor, which is slow and expensive to refresh. While vehicle-based mapping offers higher resolution, it faces two major technical hurdles:
- Positioning Jumps: GPS and Extended Kalman Filter (EKF) data can be "jumpy," leading to misalignments between image frames.
- Fusion Blurring: When multiple overlapping images are merged, simple averaging creates a blurry "ghosting" effect due to slight perspective or positioning errors.
Proposed Methodology
The authors propose a multi-stage pipeline to solve these issues, moving from raw data collection to a final fused map.
1. Data Collection and Localization
The system utilizes a vehicle equipped with two oblique cameras, GPS, and an Inertial Measurement Unit (IMU). To fix "jumpy" positioning, the authors developed an Offline Interpolation Algorithm. This algorithm takes accurate but discontinuous EKF points and fills the gaps with Dead Reckoning (DR) data, creating a smooth, continuous trajectory for the vehicle.
2. Inverse Perspective Mapping (IPM)
This technique transforms the perspective view from the car's cameras into a top-down "bird's-eye" view, creating orthophoto fragments. This step is crucial for aligning ground-level imagery with a global coordinate system.
3. Wavelet-Based Image Fusion
The core novelty lies in the fusion step. Instead of merging images directly, they use a Wavelet Transform to decompose the images into low-frequency (general structure) and high-frequency (fine detail) components.
For the fusion step, they employ a Weighted Median Algorithm rather than a simple mean. The weight is determined by the pixel's distance to the camera; closer pixels have higher resolution and less rotation error. Using a median ensures that the system selects the "best" representative pixel rather than an average of several misaligned ones, which preserves sharp edges.
Experimental Results
The approach was evaluated using real-world data collected on the campus of Shanghai Jiao Tong University (SJTU).
- Ultra-High Resolution: The system achieved a real-world resolution of 0.0187 meters per pixel (approx. 1.8 cm).
- Visual Clarity: The resulting maps clearly show road signs that are blurred in standard satellite imagery.
- Robustness: The fusion method successfully handled "double-collected" roads and varying illumination angles without losing detail.
Conclusion and Future Work
The proposed method offers a significant improvement over traditional aerial mapping for road-level detail. However, limitations remain, such as the presence of dynamic obstacles like pedestrians in the final map. Future work will focus on integrating LiDAR to detect and automatically remove these obstacles and adding moving shadow detection to further improve image consistency.
Summary updated