3D Occupancy Perception Network Based on Temporal Fusion of Bird's-Eye-View Features

In view of the difficulties of long-tailed obstacle perception and the high complexity of dynamic environment modeling in unmanned driving scenarios, this paper proposes a 3D occupancy perception network based on temporal fusion of bird's-eye-view (BEV) features. Firstly, image features are extracted by image backbone network and mapped into BEV features, and then BEV features are temporally fused by deformable attention mechanism. Secondly, a dual-branch prediction structure of 3D semantic occupancy and 2D velocity flow field is designed to decouple the heterogeneous tasks. It realizes fine-grained voxel semantic occupancy prediction through 3D convolution, and generates 2D velocity flow field combined with temporal cost volume matching mechanism, so as to reduce multi-task competition while maintaining real-time performance. Finally, a dynamic supervision strategy is proposed, which uses ray-extended sampling to generate key voxel supervision masks covering obstacles and surrounding empty voxels, and combines random sampling of empty voxels with differentiated supervision of dynamic and static voxels to alleviate the imbalance of category distribution and suppress the trailing effect of prediction. Experiments on the FlowOcc3D dataset and real vehicle show that the proposed network achieves good performance in both semantic occupancy and velocity prediction, which verifies its effectiveness in various driving scenarios. Its lightweight design provides reliable support for real-time environment perception and path planning of unmanned systems, and promotes the application of 3D occupancy perception technology in various scenarios.