Xin, Chaoyin (Shanghai Jiao Tong University), Xie, Wende (didi), Wang, Yafei (Shanghai Jiao Tong University), Li, Zexing (Shanghai JiaoTong University), Wang, Bowen (Shanghai Jiao Tong University), Shi, Hongyang (DiDi Chuxing)

GeRAD: Geometry-Aware Neural Implicit Surface Rendering for Autonomous Driving

Scheduled for presentation during the Regular Session "Sensing, Vision, and Perception I" (WeAT2), Wednesday, September 25, 2024, 11:30−11:50, Salon 5

2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC), September 24- 27, 2024, Edmonton, Canada

This information is tentative and subject to change. Compiled on October 7, 2024

Abstract

Neural Radiance Field (NeRF) has emerged as a groundbreaking method for synthesizing novel views, gaining great popularity in the autonomous driving (AD) community for rendering photorealistic driving scenarios. However, the application of NeRF in AD is hindered by its limitations in processing sensor data from vehicles with free movements, particularly due to the lack of surface geometry details and sparse viewpoints. Thus, this paper proposes a grid-based NeRF method that integrates geometry-enhanced priors to reconstruct realistic unbounded AD scenarios from real-world data. Perspective-warped hash grids are introduced to represent unbounded scenes captured from free trajectories. For accurate surface representation, we introduce the Signed Distance Function (SDF) to extract unbiased density expression. To achieve better representation under sparse viewpoints, we additionally employ monocular priors to constrain sample depths and the gradients of SDF field. We conduct experiments on KITTI, Free-dataset, and a self-collected urban road dataset. Results indicate that our method outperforms state-of-the-art approaches in AD scenarios in terms of reconstruction quality and robustness.