Research and Implementation of a Neural Radiance Fields-Based 3D Reconstruction Method for Autonomous Driving Scenes

This paper addresses the scarcity of training and testing data in autonomous driving scenarios. We propose a 3D reconstruction framework for autonomous driving scenes based on Neural Radiance Fields (NeRF). Compared to traditional multi-view geometry methods, NeRF offers superior scene representation and novel view synthesis capabilities but suffers from low training efficiency and limited generalization. To overcome these limitations, we integrate existing NeRF optimization techniques and introduce a scene-specific data reuse strategy tailored for autonomous driving, enabling continuous 3D reconstruction directly from 2D images without requiring elaborate calibration. This approach significantly improves reconstruction efficiency, achieving reliable reconstruction and real-time visualization in complex traffic environments. Furthermore, we develop an interactive scene editing plugin in Unreal Engine 5, supporting the addition, removal, and adjustment of static objects. This extension allows the generation of customizable training and testing data, providing richer data support for autonomous driving algorithms.