LiDAR Point Cloud Semantic Segmentation Based on Feature Interaction Fusion

LiDAR point cloud semantic segmentation is a critical task for environmental perception in various outdoor scenarios. Point cloud semantic segmentation based on projection views, e.g., Range View (RV) and Bird’s Eye View (BEV), has been widely studied, while single-view projection methods fail to exploit complementary information between different views during feature representation learning. To address this issue, we propose a LiDAR point cloud semantic segmentation method based on feature interaction fusion. Specifically, multiple views are first obtained through 2D projections. Then, a dual-stream neural network is designed for feature interaction and fusion, which bidirectionally aligns and integrates features from different views. Additionally, a residual attention mechanism is introduced to further refine features and reduce information loss. Finally, a confidence-weighted decision-level fusion method is employed to weight and combine the results from both views. Tested on the SemanticKITTI dataset, experimental results demonstrate that the proposed fusion network outperforms single-view methods.