Point Cloud Semantic Segmentation Based on Geometric Feature Fusion

In recent years, with the development of point cloud technology, its applications in computer vision, robotics, and autonomous driving have become increasingly widespread.As a pivotal research direction in this area, point cloud semantic segmentation is highly important for fields such as urban planning, construction, and autonomous driving.However, current point cloud semantic segmentation techniques still face numerous challenges.For example, existing segmentation methods often prioritize segmentation performance over processing time, and segmentation performance tends to decline noticeably when objects undergo rotation.To address this issue, this paper proposes an encoder-decoder architecture called the geometric feature fusion network (GFF-Net) for point cloud semantic segmentation.In this network, a local polar conversion (LPC) module is introduced to address the effects of object rotation on the segmentation results.The LPC module obtains a geometric spatial representation that does not vary with rotation about the Z-axis.Building upon this feature, a geometric feature fusion (GFF) module is proposed to compensate for the information lost in the downsampling stage.Experiments conducted on the S3DIS and SemanticKITTI datasets demonstrate that the proposed GFF-Net method not only enhances segmentation accuracy but also achieves a favorable balance between computational efficiency and segmentation performance.These results highlight the potential of GFF-Net to improve the practical applicability of point cloud semantic segmentation in real-world applications.