Search papers, labs, and topics across Lattice.
The paper introduces Superpoint Transmamba (SPTM), a novel architecture for semantic segmentation of 3D power tower point clouds, addressing the challenge of segmenting sparse and discontinuous power lines in UAV LiDAR data. SPTM employs a 3D U-Net backbone with superpoint pooling for local feature aggregation, followed by a hybrid Transformer-Mamba decoder to capture both local geometric details and global dependencies. Experiments on a UAV-collected power tower dataset demonstrate that SPTM significantly outperforms existing methods like Point Transformer and 3DSwin Transformer, especially in power line classification, as measured by mIoU and mAcc.
By fusing Transformers and Mamba state-space models, SPTM achieves state-of-the-art semantic segmentation of power tower point clouds, significantly improving power line classification in complex, occluded environments.
Semantic segmentation of three-dimensional (3D) power tower point clouds is a key technology for smart grid inspection and digital modeling of transmission corridors. However, slender structures such as power lines are sparse and discontinuous, and susceptible to occlusion in uninhabited aerial vehicle (UAV) light detection and ranging (LiDAR) point clouds, resulting in insufficient segmentation precision for existing methods. To address this issue, a novel 3D power tower semantic segmentation method, superpoint transmamba (SPTM), is proposed. Firstly, a sparse 3D U-Net is used as the backbone network, and a superpoint pooling layer is used to aggregate features in locally geometrically consistent regions. Then, a hybrid query decoder that fuses the Transformer and the Mamba state-space model is introduced in the decoding stage to balance local geometric details with the modeling of global long-range dependencies. Finally, a series of experiments are conducted on a dataset of power tower point clouds collected by UAVs to validate the effectiveness of the proposed method. Experiments demonstrate that SPTM outperforms existing mainstream methods (such as Point Transformer, Stratified Transformer, and 3DSwin Transformer) in metrics such as mean intersection over union (mIoU) and mean accuracy (mAcc), achieving significant improvements in power line classification. Ablation experiments further validate the complementarity and effectiveness of superpoint pooling and the Mamba module in identifying slender structures. This method provides an efficient and accurate solution for 3D point cloud semantic segmentation in complex power transmission scenarios.