Adaptive Window Multi-Feature Fusion Point Cloud Semantic Segmentation Network

Abstract

With the widespread application of 3D point cloud data, point cloud semantic segmentation technology has shown tremendous potential in fields such as autonomous driving, robot navigation, and urban modeling. However, the high dimensionality, sparsity, and complex local structures of 3D point cloud data make it challenging for traditional point cloud processing methods to effectively capture fine-grained features. This challenge is particularly evident when dealing with point clouds that have different scales, densities, and structural characteristics. To address these issues, this research proposes Adaptive Window Multi-Feature Fusion Point Cloud Semantic Segmentation (AWFusionNet), aimed at simultaneously considering both global and local features of point clouds, optimizing their representation capability and segmentation accuracy. The method combines dynamic and fixed window feature extraction mechanisms, using dynamic windows to model global features and fixed windows to enhance local features, effectively improving segmentation accuracy and robustness. Specifically, the dynamic window utilizes the farthest point sampling (FPS) algorithm for division and performs global information fusion through inter-window relative attention and global cross-attention. The fixed window employs a local relative attention feature expansion module to extract fine-grained local features. Additionally, the method improves edge feature recognition during the upsampling stage through an inter-layer edge enhancement and suppression module. Experimental results demonstrate that AWFusionNet achieves high accuracy and better robustness when processing point cloud data in complex scenarios.