Efficient Dual-Branch 3D Occupancy Prediction for Autonomous Driving

Engineering notes

Experiments on the nuScenes benchmark demonstrate that the proposed method achieves a real-time inference speed of 22.0 FPS while maintaining a high IoU (39.2), significantly outperforming existing mainstream methods and validating the excellent balance between accuracy and efficiency achieved by our approach.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

With the increasing demand for high-precision environmental perception in autonomous driving systems, 3D occupancy prediction has emerged as a prominent research focus as a generic 3D scene representation method that does not rely on predefined object categories. This paper proposes a real-time 3D occupancy prediction method for autonomous driving. By designing a dual-branch voxel feature extraction mechanism, semantic information is extracted from low-resolution features while geometric details are preserved from high-resolution features, achieving complementary enhancement of semantics and geometry through a feature fusion strategy. Furthermore, a virtual camera preprocessing module is introduced to unify different camera parameters, enhancing the model's cross-dataset generalization capability. Experiments on the nuScenes benchmark demonstrate that the proposed method achieves a real-time inference speed of 22.0 FPS while maintaining a high IoU (39.2), significantly outperforming existing mainstream methods and validating the excellent balance between accuracy and efficiency achieved by our approach.

Links and sources

• Official / arXiv page

Comments