针对输电线路杆塔点云分类中存在的结构稀疏、方向性强、尺度差异大等问题，提出了一种融合角度信息建模与关键点自适应采样机制的改进KPConv网络。在卷积算子中引入邻域点与核点之间的夹角余弦相似度权重，优先捕捉长条形导线等方向性结构，并通过基于特征最大响应的降采样策略选取分类贡献最大的关键点，有效保留微小部件信息；在升采样阶段，结合空间距离与高程差进行加权投票修正，平滑边界并纠正误分类。在真实输电线路杆塔点云数据集上的实验结果表明，该方法的mIoU为0.840，杆塔、导线与绝缘子的IoU分别为0.907、0.947和0.802，均显著优于PointNet、SPGraph和KPConv等方法，特别是在长细结构与微小部件识别方面优势明显，验证了其有效性与工程应用价值。

Aiming at the problems of sparse structure, strong directionality and large scale difference in the point cloud classification of transmission power line towers, we proposed an improved KPConv network integrating angular information modeling and key point adaptive sampling mechanism. In this method, we introduced the angle cosine similarity weights between neighboring points and kernel points in the convolution operator to preferentially capture directional structures such as long conductors, and selected the key points that contributed the most to the classification through the downsampling strategy based on the maximum response of features, so as to effectively retain the information of tiny components. In the upsampling stage, we made weighted voting corrections by combining spatial distances with the difference in elevation, to smooth out the boundaries and rectify the misclassification. The experimental results on the real transmission power line tower point cloud dataset show that the mIoU of this method reaches 0.840, and the IoUs of tower, conductor and insulator are 0.907, 0.947 and 0.802, respectively, which are significantly better than that of PointNet, SPGraph and KPConv. Especially in the recognition of long and thin structures and tiny parts, the advantages are obvious, which verifies the effectiveness and engineering application value of proposed method.