3D Dense Reconstruction for Autonomous Heavy-Duty Vehicles Using Gaussian Splatting

Sammanfattning

Accurate 3D reconstruction of environments is a crucial component in the perception systems of autonomous vehicles, particularly in structured domains such as highways. Recent advances have introduced techniques like 3D Gaussian Splatting, which offer promising capabilities for photorealistic and efficient 3D reconstruction. However, the application of such methods for dense geometric scene reconstruction in highway-based autonomous heavy-duty driving remains underexplored. This thesis investigates how prior knowledge, specifically from Light Detection and Ranging (LiDAR) data, can be leveraged to improve the geometric fidelity of 3D reconstruction using a pipeline based on 3D Gaussian Splatting, applied to highway scenes. By examining how different initialization strategies affect the completeness and accuracy of the resulting reconstructions, this work aims to assess the sensitivity of Gaussian Splatting to initial conditions in structured outdoor environments. Several initialization methods are proposed and evaluated, including Structure-from-Motion (SfM) only, SfM combined with LiDAR, and LiDAR-based ground flattening. Their performance is compared both qualitatively and quantitatively using one-sided and symmetric Chamfer distance metrics. Results show that initializing with LiDAR points and flattening Gaussians aligned with the ground yields substantial improvements in geometric accuracy, particularly in key regions such as the drivable road surface. However, all strategies converge to similar levels of photometric loss, despite varying levels of geometric fidelity. This highlights a key limitation of image-based supervision and highlights the challenge of 3D-to-2D ambiguity, where photometric consistency does not necessarily imply accurate 3D geometry. These findings suggest that while informed initialization is beneficial, further gains may require more advanced approaches. By demonstrating the value of incorporating scene-specific priors, this work lays a foundation for extending Gaussian Splatting to real-world autonomous driving scenarios focused on dense geometric reconstruction.