RDC‐GS: Enhanced 3D Gaussian Splatting for Robust Dash Cam Video Reconstruction

Abstract Recent advancements in 3D Gaussian Splatting (3DGS) have made significant improvements in real‐time novel view synthesis and 3D reconstruction. 3DGS has seen significant development in driving scenarios, but existing methods are mainly designed for videos captured by autonomous vehicles. It is not suitable for the complexity and dynamic lighting challenges present in dash cam videos. Despite the progress made by previous work in dealing with reflections and occlusions, the distribution of 3D Gaussian points remains inaccurate due to the inherent complexity of dash cam scenes, causing geometric distortions in the rendered output. Additionally, uncontrolled dynamic illumination exacerbates Gaussian point density anomalies and local geometric distortions. These challenges significantly hinder the development of scene reconstruction techniques based on dash cam videos. To address these challenges, we present RDC‐GS, an innovative method featuring a point correction mechanism to eliminate distribution errors of Gaussian points during training, and a brightness‐aware illumination technique to enhance detailed representation under dynamic lighting conditions. This approach yields more robust scene reconstruction. Experiments conducted on real dash cam videos demonstrate that our method achieves a 1.5‐dB PSNR improvement over current state‐of‐the‐art techniques. Comprehensive experiments validate the efficacy of our approach across challenging scenarios.