摘要
面向复杂山区机载LiDAR点云地面点提取效果不佳的问题,文中提出地形感知增强滤波网络AttPos-Net提高山区地面点滤波精度。融合高程梯度与局部高程方差构建地形感知位置编码模块,增强对阶梯地形和陡坡的垂直几何敏感性;其次,通过设计分组向量自注意力机制强化边界特征与长程依赖捕捉,提升如梯田等特殊地貌边界细节区分能力;采用轻量化残差连接平衡浅层细节保留与深层特征融合,实现地表点云的高精度提取。实验利用两个基准数据集、四川省乐山市地区和云南省山区10%的实测数据作为训练数据,在剩余90%的数据上进行测试,从而与经典网络进行对比。结果表明,AttPos-Net在各类地形中滤波效果优异,尤其在梯田、陡坡、山谷等复杂山地地形上相较经典网络优势更显著。
Aiming at solving the problem of poor extraction effect of airborne LiDAR ground points in complex mountainous terrain,this paper proposes a terrain-aware enhanced filtering network,AttPos-Net.The network constructs a terrain-aware position encoding module by fusing elevation gradients and local elevation variance to enhance vertical geometric sensitivity to stepped terrain and steep slopes.Secondly,it designs a grouped vector self-attention mechanism to strengthen boundary feature extraction and long-range dependency capture,improving the ability to distinguish excessive details at boundaries of special geomorphologies such as terraced fields.A lightweight residual optimization connection is adopted to balance shallow detail retention and deep feature fusion,achieving high-precision extraction of surface point clouds.The experiments adopt two benchmark datasets,as well as 10%of the measured data from Leshan,Sichuan,and mountainous areas in Yunnan as training data,and test on the remaining 90%of the data,comparing with classic networks.The results show that AttPos-Net has excellent filtering performance in various terrains,especially in complex mountainous terrains such as terraced fields,steep slopes,and valleys,where it has more significant advantages over classic networks.
作者
谢玉凤
李炎秋
仲鸿儒
何欣悦
郭鑫
袁伟杰
XIE Yufeng;LI Yanqiu;ZHONG Hongru;HE Xinyue;GUO Xin;YUAN Weijie(The Surveying and Mapping Engineering Brigade of Sichuan Metallurgical Geological Exploration Bureau,Chengdu,610000,China;Southwest Jiaotong University,Chengdu,610000,China)
出处
《测绘工程》
2025年第6期27-39,共13页
Engineering of Surveying and Mapping
基金
国家自然科学基金(42471484)。
