RoGs: Large Scale Road Surface Reconstruction with Meshgrid Gaussian (2405.14342v3)

Abstract: Road surface reconstruction plays a crucial role in autonomous driving, which can be used for road lane perception and autolabeling. Recently, mesh-based road surface reconstruction algorithms have shown promising reconstruction results. However, these mesh-based methods suffer from slow speed and poor reconstruction quality. To address these limitations, we propose a novel large-scale road surface reconstruction approach with meshgrid Gaussian, named RoGs. Specifically, we model the road surface by placing Gaussian surfels in the vertices of a uniformly distributed square mesh, where each surfel stores color, semantic, and geometric information. This square mesh-based layout covers the entire road with fewer Gaussian surfels and reduces the overlap between Gaussian surfels during training. In addition, because the road surface has no thickness, 2D Gaussian surfel is more consistent with the physical reality of the road surface than 3D Gaussian sphere. Then, unlike previous initialization methods that rely on point clouds, we introduce a vehicle pose-based initialization method to initialize the height and rotation of the Gaussian surfel. Thanks to this meshgrid Gaussian modeling and pose-based initialization, our method achieves significant speedups while improving reconstruction quality. We obtain excellent results in reconstruction of road surfaces in a variety of challenging real-world scenes.

• The paper introduces RoGS, which leverages 2D Gaussian surfels for faster, more accurate road surface reconstruction compared to traditional methods.
• The method employs trajectory-based initialization, reducing elevation errors by 17.62% and achieving a 15.84× speedup on KITTI and NuScenes datasets.
• These improvements enhance autonomous driving data quality and open opportunities for urban planning and augmented reality applications.

Efficient Road Surface Reconstruction with RoGS

Overview

Road surface reconstruction is crucial for autonomous driving, enabling better road lane perception and auto-labeling tasks. Traditional methods like mesh-based algorithms show promise but fall short in speed and rendering quality. Enter RoGS, an innovative approach using 2D Gaussian Splatting (2DGS) to achieve faster and more accurate road surface reconstructions.

Why Road Surface Reconstruction Matters

Reconstructing road surfaces from video data aids in creating high-definition maps and offers valuable annotations for training autonomous driving systems. It plays a significant role in understanding road regions, lane lines, and road markings. Traditional methods face challenges in sparse-feature areas like roads, leading to incomplete reconstructions with holes and noise.

Traditional Approaches and Their Limitations

Conventional methods fall into two categories:

1. Classical 3D Reconstruction: Techniques like Structure-from-Motion (SfM) and Multi-View Stereo (MVS) generate sparse or semi-dense point clouds. These approaches struggle with road surfaces due to sparse texture features, often producing incomplete reconstructions.
2. Implicit 3D Reconstruction: Methods such as Neural Radiance Fields (NeRF) offer impressive rendering effects but struggle with geometric accuracy, especially in large-scale scenes. Additionally, these methods can be time-consuming and computationally intensive.

The RoGS Approach

RoGS introduces a new perspective by using 2D Gaussian surfels. Each surfel stores not just geometric data but also color and semantic information. This explicit representation aligns more closely with the physical reality of roads and offers several advantages:

• Speed: RoGS achieves significant speedups compared to traditional mesh-based methods.
• Quality: The 2D Gaussian surfels provide better accuracy in representing road geometry.

Data Initialization

RoGS leverages vehicle trajectory data for initialization. By aligning surfel coordinates with the vehicle trajectory, the method ensures a more accurate starting point for optimization. This trajectory-based initialization is more effective than random or point-cloud based approaches.

Results and Performance

The authors tested RoGS on the KITTI and NuScenes datasets, achieving impressive results:

• Speed: RoGS exhibits a 15.84× speedup compared to traditional methods like RoMe.
• Accuracy: With LiDAR data for supervision, RoGS shows a 17.62% reduction in elevation error, surpassing RoMe in both PSNR and mIoU metrics.

Implications and Future Prospects

The implications of this research are twofold:

1. Practical: Faster and more accurate road reconstructions mean better training data for autonomous vehicles and more reliable maps.
2. Theoretical: The combination of explicit geometric representations with trajectory-based initialization offers a new avenue for improving 3D scene reconstruction techniques.

Looking forward, this approach could benefit various applications beyond autonomous driving, such as urban planning and augmented reality. Future developments might focus on refining the initialization process and further improving the robustness of the method against less accurate trajectory data.

Conclusion

RoGS represents a significant step forward in road surface reconstruction. By leveraging 2D Gaussian surfels and trajectory-based initialization, this method not only speeds up the process but also enhances the accuracy of the reconstructions. These advancements pave the way for more efficient and reliable autonomous driving technologies.