A Controllable Editing Closed-Loop 3D Gaussian Autonomous Driving Sensor Simulator

With the development of Neural Radiance Fields (NeRF) and 3D Gaussian techniques, significant progress has been made in autonomous driving sensor simulation. However, current data-driven sensor simulations are limited to the re-production of autonomous driving scenes, with restricted scene editing capabilities. To address this limitation, this paper proposes a 3D Gaussian-based closed-loop controllable editing sensor simulator for the joint simulation and planning of sensors. The simulator reconstructs a road network model aligned with the 3D Gaussian model, enabling the planning of background vehicle trajectories within the scene and facilitating scene editing. This paper introduces a technique for the simultaneous construction of the road network and 3D Gaussian model, automating the creation of both. Evaluation results validate that the proposed method can automatically construct high-precision rendered camera sensors and aligned road networks, demonstrating its potential for application in large-scale scenarios.