Abstract: We propose a 3D point cloud classification method that adaptively selects the optimal neighborhood size of a single point and learns deep-level features with stronger generalization ability. Firstly, we obtain the optimal local neighborhood information of each point based on the adaptive optimal neighborhood size selection, and then extract the low-level features of the point cloud based on the local neighborhood information; then we design a convolutional neural network model with the low-level features of the points to be classified as the input, learn the deep-level features that can reflect the inherent properties of the target features and realize the classification. Finally the experiment is conducted using Topcon's 3D point cloud dataset, which is obtained by a mobile platform equipped with a TOPCON GLS-2200 3D laser scanner. The results show that the overall accuracy of the classification results of this paper reaches 90.48%, which is better than other point cloud classification methods.

Key words: point cloud classification, adaptive optimal neighborhood size selection, deep level features, neural networks