U-Agent: A Point Cloud Segmentation Network Based on Local Feature Enhancement and Agent Attention for Global Modeling

Light Detection and ranging (LiDAR) can rapidly acquire large-scale, high-precision 3-D data, providing critical support for fields such as surveying and mapping, forestry, and disaster monitoring. Significant progress has been achieved in the semantic segmentation of LiDAR point cloud using deep learning techniques. Although transformer-based methods have significantly improved segmentation accuracy due to their strong contextual modeling capabilities, their inherent quadratic complexity limits their application in real-world scenarios. Most existing studies adopt local attention or sliding window attention to alleviate computational pressure, but this compromise inevitably weakens the model's global modeling ability. To address this critical issue, this article innovatively proposes a point cloud segmentation network, named U-Agent, based on local feature enhancement and agent attention for global modeling. This network effectively represents the local and global information of point cloud and improves segmentation performance. The method first introduces a multiscale local feature enhancement module, which strengthens the details of local point cloud by performing differential operations between features from multiscale neighborhood and incorporating a distance-adaptive factor. Meanwhile, a global feature extraction module based on agent attention is designed to intelligently evaluate and select high-value agent points, construct a bridge for global information aggregation, and realize efficient expansion of the receptive field. In addition, these two modules are integrated into the classical U-Net to achieve complete fusion of local and global features. The proposed method yields excellent results on the LASDU and DFC2019 datasets, and highly competitive results on the DALES dataset. Experimental results on the LASDU dataset show that the network achieves an overall accuracy (OA) of 87.9% and a mean F1-score (mF1) of 79.0%; on the DFC2019 dataset, it attains an OA of 97.3% and an mF1 of 91.2%; on the DALES dataset, it achieves an OA of 97.3% and a mean Intersection over Union (mIoU) of 78.1%. These results fully verify the advancement and effectiveness of the proposed method.