Introduction
For autonomous vehicles, the road surface is far more than a flat plane; it is a complex geometric landscape of bumps, ruts, and potholes. Accurate reconstruction of this geometry is critical for ensuring passenger comfort through active suspension control and maintaining safety during high-speed maneuvers. Traditionally, this task has been approached through perspective-view depth estimation. However, these methods often struggle with the fine-grained precision required for centimeter-level surface analysis.
This paper introduces RoadBEV, a novel framework that shifts road surface reconstruction from the traditional perspective view to a Bird’s-Eye-View (BEV) representation. By focusing on vertical elevation rather than skewed longitudinal depth, RoadBEV achieves unprecedented accuracy, enabling vehicles to "see" road irregularities with sub-centimeter precision.
The Limitations of Perspective-View Depth
The authors identify a fundamental "angle bias" inherent in perspective-view depth estimation that hinders its application to road surfaces:
- Inconsistent Search Directions: In a perspective view, the network searches for depth along rays that are nearly parallel to the road surface. This geometry makes it mathematically difficult to capture subtle vertical elevation changes.
- Information Sparsity: As distance increases, road features become extremely sparse in the image plane. Computational resources are often wasted on global scene hierarchy rather than the local texture necessary for surface reconstruction.
- Coarse Resolution: Global depth searching typically lacks the resolution needed to detect centimeter-level variations, which are essential for advanced chassis control.
The RoadBEV Framework
RoadBEV addresses these challenges by defining a 3D "Region of Interest" (ROI) specifically around the road surface. This ROI is discretized into a voxel grid. Unlike traditional methods that "lift" 2D features into 3D space—often introducing noise from depth priors—RoadBEV utilizes 3D-to-2D projection to query pixel features directly from the input images.
The framework consists of two primary architectures:
RoadBEV-mono
Designed for real-time applications, the monocular model extracts multi-scale features from a single image and projects them into the voxel grid. These features are then reshaped into a BEV feature map. To determine elevation, the model treats the problem as a classification task over pre-defined vertical "bins" (e.g., 1 cm intervals). A soft argmin operation is then applied to convert these discrete classifications back into a continuous elevation value, providing a balance between categorical stability and numerical precision.
RoadBEV-stereo
The stereo variant builds a 4D cost volume within the BEV space by correlating features from both left and right camera images. Because the search range is restricted to a narrow vertical window (typically +/- 20 cm), the network can dedicate its full computational capacity to high-resolution similarities. This effectively transforms the model into a high-precision 3D occupancy network specialized for the road surface.
Experimental Validation
The models were evaluated using the Road Surface Reconstruction Dataset (RSRD), a high-fidelity benchmark containing 2,800 pairs of high-resolution stereo images. The ground truth for this dataset consists of dense point clouds and precise motion pose information, which the authors used to generate a digital elevation map (DEM) with a 3x3 cm grid resolution.
Key Results
RoadBEV significantly outperformed existing state-of-the-art (SOTA) depth and stereo matching methods:
- Monocular Performance: RoadBEV-mono achieved an absolute elevation error of 1.83 cm, representing a 29.3% reduction in error compared to leading models like AdaBins. It maintains a high inference speed of 26.8 FPS, making it viable for onboard deployment.
- Stereo Performance: RoadBEV-stereo reached a remarkable 0.50 cm error rate, surpassing specialized stereo models such as ACVNet and DVANet. While more computationally intensive (8.0 FPS), its precision offers a reliable "preview" of the road ahead.
Insights and Future Directions
The success of RoadBEV stems from two primary insights. First, vertical alignment is essential; by aligning the search direction with the vertical axis, the model captures denser and more recognizable profile features. Second, narrowing the search space to a 40 cm vertical window simplifies the learning task and minimizes the impact of background noise.
Despite these gains, some limitations remain. Accuracy still tends to degrade as the longitudinal distance from the camera increases due to the inherent resolution limits of the original perspective images. Future work will focus on temporal integration to stabilize reconstructions over time and handle occlusions. Additionally, the authors suggest exploring the joint reconstruction of geometry and texture using emerging techniques like 3D Gaussian Splatting to further enhance the vehicle's perception of its environment.
Summary updated