Semantic Segmentation And Modeling Of3D Scene Point Cloud

Scene understanding, which refers to simultaneously segmenting, classifying and modeling objects in3D point cloud, is one of the fundamental problems of computer vision. An excellent scene understanding can be of great benefit to various computer vision applications, like object classification, automatic driver and3D scene modeling. With the increasing abilities of3D sensors such as lidar sensors, the3D point cloud of residential scene is easier to capture than before. It is important to propose an approach to semantically segment the point cloud.With the increasing availability of3D sensors such as lidar sensors, stereo and SFM(Structure From Motion) systems, the3D point clouds of urban scenes are easier than ever to collect. The lidar sensor is faster and more accurate than other3D sensors. It is important to provide a suite that can automatically process the huge data and semantic segment them into comprehensible meanings.This thesis studies semantic segmentation on3D point cloud, including segmen-tation, feature extraction, classification and optimization. We also study3D modeling of trees and houses based on semantic segmentation, and analysis the scene rendering in Maya and Unity. We propose a cluster-group based framework for the semantic segmentation of3D residential scenes point cloud. We introduce a dual-scale analy-sis, clustering and grouping, which is different from most recent proposals and help to separate different scale objects in scene. Another contribution of this work is that we only use the simple geometrical features of clusters and groups to train the classifier and recognize objects. After classification, we use GraphCuts to optimize. Scenes are effectively semantic segmented by this framework.