DrivingGaussian: Composite Gaussian Splatting for Surrounding Dynamic Autonomous Driving Scenes

TL;DR

DrivingGaussian introduces a novel framework for dynamic autonomous driving scene reconstruction, combining static background modeling with composite dynamic Gaussian graphs, enhanced by LiDAR priors for detailed, panoramic, and consistent surround-view synthesis.

Contribution

The paper proposes a new composite Gaussian splatting method that effectively models both static and dynamic elements in driving scenes, improving reconstruction accuracy and visual fidelity.

Findings

Outperforms existing scene reconstruction methods in dynamic environments.

Achieves high-fidelity, panoramic surround-view synthesis.

Maintains multi-camera consistency in complex scenes.

Abstract

We present DrivingGaussian, an efficient and effective framework for surrounding dynamic autonomous driving scenes. For complex scenes with moving objects, we first sequentially and progressively model the static background of the entire scene with incremental static 3D Gaussians. We then leverage a composite dynamic Gaussian graph to handle multiple moving objects, individually reconstructing each object and restoring their accurate positions and occlusion relationships within the scene. We further use a LiDAR prior for Gaussian Splatting to reconstruct scenes with greater details and maintain panoramic consistency. DrivingGaussian outperforms existing methods in dynamic driving scene reconstruction and enables photorealistic surround-view synthesis with high-fidelity and multi-camera consistency. Our project page is at: https://github.com/VDIGPKU/DrivingGaussian.