Environment Perception Method for Autonomous Vehicle Platoons Based on LiDAR–Vision Fusion and Multi-Scale Attention Mechanism

Engineering notes

Experimental results show that, compared with mainstream fusion models, the proposed method achieves significant improvements in obstacle detection accuracy and environmental robustness.

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

To enhance the perception accuracy and robustness of unmanned vehicle formations in complex dynamic environments, this paper proposes an environmental perception method based on LiDAR-visual fusion and a multi-scale attention mechanism (MSA). First, spatial geometric features are extracted from LiDAR point clouds using a sparse voxel Transformer module, while image data is processed through a spatiotemporal depth-aware convolutional network to extract semantic information. Subsequently, the two are fused in a cross-modal selfattention alignment module to explicitly model cross-view spatial correspondences and occlusions within the formation. To improve the ability to recognize obstacles at multiple scales, a multi-scale hierarchical attention decoder is designed, integrating pyramid scene embedding, deformable cross-scale interactions, and BEV semantic aggregation mechanisms. Additionally, a graphstructured inter-vehicle consistency perception module is introduced, which fuses formation structural constraints through a multi-agent attention propagation mechanism to enhance perception consistency and confidence. Experimental results show that, compared with mainstream fusion models, the proposed method achieves significant improvements in obstacle detection accuracy and environmental robustness.

Links and sources

• Official / arXiv page

Comments