GraphGSOcc: Semantic-Geometric Graph Transformer With Dynamic-Static Decoupling for 3D Gaussian Splatting-Based Occupancy Prediction

Focusing on the task of 3D semantic occupancy prediction for autonomous driving, we address three key issues in existing 3D Gaussian splitting (3DGS) methods: (1) unified feature aggregation that neglects semantic correlations among similar categories and across regions, (2) boundary ambiguities caused by the lack of geometric constraints in MLP iterative optimization, and (3) bias issues in dynamic-static object coupling optimization. We propose the GraphGSOcc model, a novel framework that combines semantic and geometric graph transformers and decouples dynamic-static object optimization for 3D Gaussian splitting-based occupancy prediction. We propose a dual Gaussian graph attention approach, which dynamically constructs dual graph structures: a geometric graph that adaptively calculated KNN search radii based on Gaussian poses, enabling large-scale Gaussians to aggregate features from broader neighborhoods while compact Gaussians focus on local geometric consistency and a semantic graph that retains top-M highly correlated nodes via cosine similarity to explicitly encode semantic relationships within and across instances. Coupled with the multiscale graph attention framework, the fine-grained attention applied at lower layers optimizes boundary details, whereas the coarse-grained attention occurring at higher layers models the object-level topology. Additionally, we decouple dynamic and static objects by leveraging semantic probability distributions and design a dynamic-static decoupled Gaussian attention mechanism to optimize the prediction performance for both dynamic objects and static scenes. GraphGSOcc achieves state-of-the-art performance on the SurroundOcc-nuScenes, Occ3D-nuScenes, OpenOcc, and KITTI occupancy benchmarks. Experiments on the SurroundOcc dataset achieve an mIoU of 25.20%, reducing GPU memory to 6.0 GB, demonstrating a 1.97% mIoU improvement and a 13.7% memory reduction compared with GaussianWorld.