A System for Evaluating Autonomous Driving Algorithms via Gaussian Splatting and Mesh Representation

This paper presents a system for capturing and representing spatial data extracted from an autonomous vehicle equipped with dual frontal cameras with limited overlapping field of view, LiDAR and wheel sensors. The methodology includes a sequence of steps to process data into a meaningful representation for simulation: initial image and camera pose capture, pose refinement, offline multi-view stereo reconstruction and Gaussian Splatting model extraction. By minimizing camera overlap, the system maximizes the quantity of spatial information captured. The resulting mesh and Gaussian Splatting point cloud enable realistic simulations, providing a robust framework for testing and evaluating autonomous driving systems.