Point Cloud Semantic Segmentation Algorithm Based on Multi-information Markov Random Field

There is a strong spatial relationship between point clouds obtained from the same object, nevertheless it usually takes a great quantity of time to directly establish the relationship model between individual point clouds. In this paper, we draw on this idea and propose a new semantic segmentation method based on 3D mesh model, which is applied to do the semantic segmentation of point cloud data in traffic scene. Firstly, the Markov random field is used to model according to the attribute information of the triangular patches of the 3D mesh model and the spatial dependence between the bins. Secondly, the attribute information of the triangle patch is represented by the intensity information of the point cloud data included in each triangle patch, and it is clustered by Gaussian Mixture model which describes the matching degree of each attribute with each class. The algorithm combines topology information of grid model and intensity information of point cloud data, eliminates the over segmentation effectively and makes the boundary of the partition smooth. Finally, ISDF (improved shape and diameter function) is proposed to determine the final class of two side triangular patches on the boundary. The proposed method is evaluated on two public point cloud datasets, and shows the competitive performance.