Self-Supervised Gaussian-Based Framework for Robust Static-Dynamic Decomposition and Real-Time Urban Scene Reconstruction

High-fidelity 3D reconstruction of large-scale urban scenes is critical for autonomous driving perception and simulation. Existing neural rendering methods, including NeRF and Gaussian-based variants, often face challenges like unstable geometry, noisy motion segmentation, and poor performance under sparse viewpoints or varying illumination. This paper presents a self-supervised Gaussian-based framework to address these challenges, enabling robust static–dynamic decomposition and real-time scene reconstruction. The proposed method introduces three innovations: (1) a semantic–geometric feature fusion module that combines semantic context and geometric cues for reliable motion prior estimation; (2) a cross-sequence geometric consistency constraint that enforces depth and surface continuity across time and viewpoints; (3) an efficient Gaussian parameter optimization strategy that stabilizes geometry by jointly constraining scale and normal updates. Experiments on the Waymo Open Dataset and KITTI benchmarks show that the proposed framework improves PSNR by up to +1.3 dB, SSIM by +0.015, and reduces depth L1 error by over 25%, while achieving real-time rendering speeds exceeding 40 FPS. These results demonstrate that the proposed framework provides a robust, scalable solution for urban scene reconstruction, with practical applications in autonomous driving.