基于深度卷积神经网络的点云三维目标识别方法研究

Research on 3D Object Recognition Method for Laser Point Cloud Based on Deep Convolution Neural Network

摘要:

为了提高对三维点云目标的识别精确度，提出一种基于深度卷积神经网络(Convolutional Neural Network, CNN)的点云目标识别模型。针对已有的深度卷积点云目标识别网络无法有效提取点云局部拓扑特征的问题，采用迭代最远点采样(Iterative Farthest Point Sampling, FPS)结合方向卷积编码方式来捕获局部形状特征。并引入空间变换网络(Spatial Transform Network, STN)使点云数据能够自适应进行空间变换和对齐，以解决点云数据旋转性会造成目标识别结果不稳定的问题。实验结果表明：文中提出的点云目标识别方法有效提高了识别精度度，相较于PointNet在ModelNet40和ShapeNetCore两个数据集上分别提高1.2%和1.4%。

Abstract:

In order to improve the accuracy of point cloud target recognition, a target recognition model for point cloud based on Convolutional Neural Network (CNN) is proposed. Aiming at the problem that the existing deep convolutional point cloud target recognition network can not effectively extract the local topological features of point cloud, Iterative Farthest Point Sampling (FPS) combined with directional convolutional coding is used to capture the local shape features. In view of the instability of target recognition results caused by the rotation of point cloud data, the introduction of Spatial Transform Network (STN) enables point cloud data to self-adaptively perform spatial transformation and alignment. The experimental results show that the point cloud target recognition method proposed in this paper effectively improves the recognition accuracy which increases by 1.2% and 1.4% respectively compared with PointNet on ModelNet40 and ShapeNetCore datasets.