基于双路径特征提取网络的三维点云分割算法

3D point cloud segmentation based on dual path feature extraction network

三维点云分割是点云处理的一项关键任务,其作为场景理解的重要步骤,在机器人、自动驾驶等多个领域中得到了愈发广泛的应用和关注。本文针对点云具有的无序性和变密度等难点,使用核点卷积算子和点云展平卷积算子构造了相互增强的双路径特征提取网络。结合预处理模块、残差式特征融合模块及空间、通道注意力模块以编码器-解码器式的架构,提出了可以实现点云分类和分割的双任务神经网络算法。在ModelNet40分类数据集和S3DIS、SemanticKITTI分割数据集上进行实验,与当前最新算法的精度对比显示,所提算法在点云分类和分割任务上具有先进性能。另外,消融实验的结果证明了本文所提出的双路径特征提取网络与注意力模块结合的有效性和可行性。

3D point cloud segmentation is a key task of point cloud processing. As an important step of scene understanding, it has been more and more widely used and concerned in many fields such as robot, automatic driving and so on. Aiming at the difficulties of disorder and variable density of point cloud, KPConv and FPC point convolution operator are used to construct a mutually enhanced dual path feature extraction network. Combined with preprocessing module, residual feature fusion module, spatial and channel attention module, a dual task neural network algorithm that can realize point cloud classification and segmentation is proposed in the encoder-decoder architecture. Experiments on ModelNet40 classification dataset and S3DIS, SemanticKITTI segmentation datasets show that the proposed algorithm has advanced performance in point cloud classification and segmentation tasks. In addition, the results of ablation experiment prove the effectiveness and feasibility of the combination of dual path feature extraction network and attention module proposed in this paper.