EMIE-MAP: Large-Scale Road Surface Reconstruction Based on Explicit Mesh and Implicit Encoding

TL;DR

This work proposes EMIE-MAP, a novel method for large-scale road surface reconstruction based on explicit mesh and implicit encoding, which achieves remarkable road surface reconstruction performance in a variety of real-world challenging scenarios.

Abstract

Road surface reconstruction plays a vital role in autonomous driving systems, enabling road lane perception and high-precision mapping. Recently, neural implicit encoding has achieved remarkable results in scene representation, particularly in the realistic rendering of scene textures. However, it faces challenges in directly representing geometric information for large-scale scenes. To address this, we propose EMIE-MAP, a novel method for large-scale road surface reconstruction based on explicit mesh and implicit encoding. The road geometry is represented using explicit mesh, where each vertex stores implicit encoding representing the color and semantic information. To overcome the difficulty in optimizing road elevation, we introduce a trajectory-based elevation initialization and an elevation residual learning method based on Multi-Layer Perceptron (MLP). Additionally, by employing implicit encoding and multi-camera color MLPs decoding, we achieve separate modeling of scene physical properties and camera characteristics, allowing surround-view reconstruction compatible with different camera models. Our method achieves remarkable road surface reconstruction performance in a variety of real-world challenging scenarios.

Figures (18)

• Figure 1: We propose EMIE-MAP, a novel large-scale road surface reconstruction method based on Explicit Mesh and Implicit Encoding. By taking input surround-view video sequences and localization information, EMIE-MAP is capable of reconstructing RGB maps, semantic maps, and elevation maps. The RGB maps are corresponding to different cameras. EMIE-MAP integrates the advantages of explicit and implicit representations, enabling accurate road surface reconstruction and rendering.
• Figure 2: Overview of EMIE-MAP. The left side presents the proposed road surface representation based on explicit mesh and implicit encoding. We utilize a mesh composed of equilateral triangular faces to represent the road structure. Each vertex stores its initial coordinates $(x,y,z_0)$, semantic information, and implicit color encoding. An elevation residual network predicts the elevation residual at each vertex, while multiple RGB MLPs decode the implicit color features into observed colors for the corresponding camera. This yields a road surface with explicit information in the framework's middle section. The right side shows optimization losses. The generated RGB and semantic maps through direct rendering are supervised by observed images. Additionally, Lidar point clouds are utilized to supervise the road surface elevation.
• Figure 3: Road surface reconstruction results in city street scene. From top to bottom are the input images, reconstructed road surface RGB images from different camera perspectives, semantic map, and elevation map. Despite the presence of dense traffic, the road surface in occluded areas is effectively filled in by images taken at different times.
• Figure 4: Road surface reconstruction results in night scene. The only light source is the car headlights. In the input image with poor lighting quality, the RGB results are also dark. However, our method is still able to reconstruct the desired semantic map.
• Figure 5: Road surface reconstruction results in ramp scenes. Our method remains remarkable performance in ramp scenes, capturing road surface color, semantics, and elevation.