Autonomous driving scenario generation based on neural radiance fields or 3D Gaussian splatting: State-of-the-art investigations, reviews, and perspectives

Reconstruction of autonomous driving scenarios plays a pivotal role in vehicle testing. In recent years, approaches based on Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3D GS) have opened new avenues for enhancing the performance of vehicle testing systems. This paper provides a systematic review of the latest research progress on NeRF and 3D GS in the context of autonomous driving scene reconstruction and explores their potential for future applications in the development of autonomous driving test systems. First, the paper briefly reviews the development trajectories of NeRF and 3D GS, presenting a concise timeline of representative works based on major databases such as Web of Science, IEEE Xplore, and arXiv, thereby offering essential background for the subsequent discussion. Next, the current state of research is comprehensively analyzed, with a particular focus on advancements in model architecture, scene editing, and the reconstruction of challenging autonomous driving scenarios. Finally, the paper outlines a research outlook on an integrated test system centered on NeRF or 3D GS, encompassing modules for large-scale dynamic scene reconstruction, autonomous scene editing, and perception-decision algorithms. This integrated framework aims to provide a valuable reference for building efficient and scalable testing platforms for autonomous driving.