3D Point Cloud Semantic Segmentation In Large Scale Outdoor Scenes

In the field of computer vision research,effective understanding of 3D scenes is a prerequisite for many applications.This paper mainly discusses the 3D point cloud semantic segmentation problem for large-scale outdoor scenes,and uses two algorithms to get effective semantic segmentation results for it.In order to overcome the challenge of huge data and representation,this paper first proposes a method for generating point cloud image representations.This method sets a cylindrical model at the scene viewpoint and projects the entire 3D scene onto the side of the cylinder.Then expand the side of the cylinder and divide the pixel grid,and finally calculate the pixel value to generate a grayscale image.The point cloud image model can transform the disordered 3D point cloud data into an ordered 2D image,which creates conditions for the subsequent use of the mature feature extraction method in the image processing field.In the traditional point cloud semantic segmentation algorithm based on handcrafted features,it is more effective to extract the multi-scale geometric features on neighborhood,but the method relies on the intervention of human experience and is time consuming.In this paper,we first borrow the point cloud image model to quickly extract multi-scale image texture features.Then,in order to enhance the distinguishability of the feature set,we extract the geometric features of the 3D point under a single optimal neighborhood size for achieving the balance between the feature representation ability and the computational time.Finally,the random forest classifier is used for two combinations of the above features,and the semantic segmentation result of the 3D point cloud scene is achieved.When applying convolutional neural networks in the field of deep learning to 3D data,how to design a regular and efficient input representation is a common challenge faced by the 3D visual community.This paper adopts a novel point cloud representation method proposed by Landrieu et al.,namely the super point graph,and adapts the point cloud semantic segmentation algorithm based on the super point graph.The adaptation focuses on improving the representation ability of the feature set in the part of super point graph construction of the original algorithm.The experimental results show that the adaptation can effectively improve the semantic segmentation result without increasing the time-consuming of the algorithm.This paper verifies the above two semantic segmentation algorithms using public datasets from the ETH Zurich and MINES ParisTech,and compares the performance of the two algorithms.The results show that both algorithms can realize the semantic segmentation task for large-scale outdoor scenes.The improved version of the point cloud semantic segmentation algorithm based on the super point graph has advantages in overall accuracy,while the semantic segmentation algorithm based on image model features takes advantage of time-consuming aspects.