RoGs: Large Scale Road Surface Reconstruction with Meshgrid Gaussian

TL;DR

This work proposes a novel large-scale road surface reconstruction approach with meshgrid Gaussian, named RoGs, that achieves significant speedups while improving reconstruction quality and obtains excellent results in reconstruction of road surfaces in a variety of challenging real-world scenes.

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.

Figures (22)

• Figure 1: Road surface reconstruction results (KITTI odometry sequence-00) using our proposed RoGs, covering a trajectory length of roughly 3724m. We reconstruct the complete scene in just 4 minutes, while RoMe mei2024rome takes 2.5 hours.
• Figure 2: We propose RoGs, a large-scale road surface reconstruction method based on Gaussian Splatting. Inputting the surround view video, semantic segmentation results, and vehicle poses, RoGs is able to reconstruct RGB maps, semantic maps, and elevation maps.
• Figure 3: Overview of RoGs. The left side shows the road representation with meshgrid Gaussian. The blue curve indicates the vehicle trajectory. Mesh-distributed Gaussian surfels are used to represent the road surface. Each surfel stores position, scale, rotation, color, opacity, and semantic information. The learnable parameters are indicated in blue. The middle demonstrates the pose-based initialization. For each Gaussian surfel, its elevation (z) and rotation are initialized using the nearest vehicle pose. Finally, the rendering is supervised by camera images and semantic segmentation results. Additionally, to improve the reconstruction quality, LiDAR point clouds can be introduced to supervise the elevation. $L^{*}$ indicates an optional loss.
• Figure 4: Comparison of the two layouts. Layout-1 covers the entire road with fewer Gaussian surfels and reduces the overlap between Gaussian surfels during training. Therefore, Layout-1 is used to place Gaussian surfels.
• Figure 5: Pose-based initialization. For each Gaussian surfel, its elevation and rotation are initialized using the nearest vehicle pose. The blue curve represents the vehicle trajectory, and the red arrows represent the x-axis.