Classification and Segmentation of 3D Cloud Points

Abstract

In recent years, the advent of advanced 3D sensing technologies has facilitated the acquisition of detailed spatial data in the form of point clouds. These 3D point clouds, composed of discrete data points in a spatial coordinate system, offer a comprehensive representation of object surfaces and environments, makingthemindispensable in various applications, ranging fromautonomousdriving and robotics to architecture and healthcare. This thesis explores the methodologies and advancements in the classification and segmentation of 3D point clouds, focusingonboth traditionalmachinelearning approachesandcontemporary deep learning techniques. Central to this thesis isanin-depth analysis of state-of-the-art deep learning frameworks tailored for 3D data, including PointNet, PointNet++. These models, by leveraging the spatial structure of point clouds, have demonstrated remarkable performance in both classification and segmentation tasks. The research further examines advanced segmentation techniques, differentiating between semantic and instance segmentation, and evaluates their effectiveness in partitioning complex scenes into meaningful segments. In conclusion, this thesis contributes to the growing body of knowledge in 3D point cloud analysis by providing a comprehensive review of existing techniques, introducingnovelenhancements, andidentifying future research directionsaimed at further improving the accuracy and applicability of 3D point cloud processing technologies. i