Geometry Identification from Point Cloud Data

Abstract

Abstract Abstract: The aim of this thesis is to conduct a comparative study of different algorithms which can learn from 3D (x,y,z) point cloud data and are able to conduct per-point classification. The point clouds of interest for this thesis are the one’s which represent a 360-degree view of an object. For example, the input is a point cloud representing a knife with two parts: the blade and the handle. For such an input, the algorithms must be able to classify which points of this point cloud belong to blade and to the handle. To solve this task different types of deep-learning based models like graph-based neural networks and non-grid point convolution networks were implemented and analyzed. These models perform with over 80% accuracy which is quite remarkable given the input is only raw 3D coordinates with no need of voxelization. The models have also been tested on synthetic datasets as well as an object scanned using a LiDAR camera. These algorithms can be applied in autonomous driving, augmented/virtual reality applications, medical data processing and 3D animation industry. Keywords: