Geometry-Aware Spatio-Temporal Context Modeling for 4D Occupancy Forecasting
This study addresses the challenges of static geometric distortion and poor long-term temporal consistency in 4D occupancy prediction by proposing GAST. The method achieves end-to-end joint optimization of geometry-aware spatiotemporal context through progressive explicit-implicit generation and dual-path spatiotemporal modeling, integrated with pose-driven deformation and motion-aware feature modulation. Experimental evaluations on the Occ3D-nuScenes benchmark demonstrate that GAST improves mIoU by 7.67% and accelerates inference speed by 2.84× compared to existing approaches. Furthermore, the proposed framework significantly enhances long-horizon prediction performance, thereby achieving high-fidelity 4D occupancy prediction. These results validate the effectiveness of combining geometric awareness with dynamic temporal modeling for robust scene understanding in autonomous driving applications.