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

We propose a self-supervised Gaussian-based framework that enhances the robustness and efficiency of static-dynamic decomposition and high-fidelity reconstruction in autonomous driving scenarios. Existing approaches demonstrate promising performance but remain prone to noisy motion segmentation, overfitting on sparse LiDAR observations, and degraded accuracy in low-texture or distant regions, resulting in unstable geometry. To address these challenges, our method introduces a feature fusion module that jointly exploits image semantics and geometric cues, enabling more reliable motion prior extraction under noise and illumination variations. A cross-sequence geometric consistency constraint is further designed to regularize reconstruction across temporal and viewpoint changes, ensuring physically coherent scene representations. In addition, we formulate an efficient Gaussian parameter optimization strategy that constrains scale and normal updates, preventing unrealistic ellipsoid growth while maintaining competitive rendering performance. Experiments conducted on the Waymo Open Dataset and KITTI benchmark confirm the effectiveness of our approach. Compared with state-of-the-art self-supervised baselines, the proposed framework achieves up to +1.3 dB PSNR and 0.015 SSIM improvements, and reduces depth L1 error by over 25%, while sustaining rendering speeds exceeding 40 FPS. These results establish our framework as a robust and scalable solution for realistic urban scene modeling, with direct relevance to autonomous driving perception, simulation, and safety validation.