LiDAR-Radar Dual-Stream Pointcloud Aggregation Mid-Fusion With Multi-Coordinate System for 3D Semantic Segmentation
Ayoung Lee, Sungpyo Sagong, H. Y. Lee, Kyongsu Yi · IEEE Access · 2026
This paper proposes Dual-stream Point aggregation Mid-fusion with Multi-coordinate system (DPMM), which leverages LiDAR-radar fusion with velocity information and 3D coordinate compatibility for improved semantic segmentation in autonomous driving. While LiDAR-camera fusion has been widely adopted, the space modality gap poses significant challenges in data alignment and computational efficiency. LiDAR-radar fusion mitigates these issues through shared 3D coordinate space, but existing methods rely on coarse BEV or voxel representations unsuitable for point-level semantic segmentation. To address this, we first introduce a LiDAR-Radar Dual-stream Point cloud Concat Attention Fusion Block (DPCAF) that efficiently processes two sensor modalities through separate pathways and a fusion process via multi-head self-attention. Second, we employ a multi-coordinate partition system processing point clouds in cubic, cylindrical, and spherical representations, capturing diverse geometric features. Experimental results on the nuScenes lidar semantic segmentation dataset demonstrate an overall 80.1 mIoU, with our qualitative analysis demonstrating high performance on challenging cases such as velocity-dependent object classification and distant, sparse objects. This study contributes to 3D semantic segmentation by introducing a novel LiDAR-radar fusion-based model that effectively combines the strengths of LiDAR and radar sensors.