AlignOcc: Alignment-Aware LiDAR-Camera Fusion for 3D Occupancy Prediction in Autonomous Driving

Engineering notes

It achieves tightly coupled multi-modal representation via a geometry-semantic framework, promoting fine-grained fusion of structural and semantic information. Extensive experiments on the nuScenes dataset demonstrate that our method achieves superior performance.

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

Comprehensive modeling of real-world autonomous driving scenarios is critical for intelligent transportation systems. Multi-modal fusion-based 3D occupancy prediction methods effectively address the limitations of conventional 2D object detection task in perceiving irregularly shaped obstacles and unknown object categories. However, most existing methods insufficiently leverage the rich semantic and geometric information embedded in raw data. Moreover, current multi-sensor fusion approaches often neglect the inherent misalignment between LiDAR and camera modalities, thereby compromising perception accuracy. This paper introduces AlignOcc as a novel LiDARCamera fusion-based framework for 3D occupancy prediction. It achieves tightly coupled multi-modal representation via a geometry-semantic framework, promoting fine-grained fusion of structural and semantic information. In addition, we design an Alignment-Aware Fusion Module that performs global alignment between the two modalities via bidirectional dynamic offsets. Extensive experiments on the nuScenes dataset demonstrate that our method achieves superior performance.

Links and sources

• Official / arXiv page

Comments