Duboki modeli za registraciju oblaka točaka

master's thesis

• Title (english): Deep models for point cloud registration
• Author: Manuel Matošević
• Mentor: Tomislav Petković (mentor)
• Committee member: Tomislav Petković (predsjednik povjerenstva)
• Committee member: Marko Subašić (član povjerenstva)
• Committee member: Stjepan Begušić (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2025-02-17, Croatia

Abstract

Registracija oblaka točaka jedan je od ključnih problema u obradi 3D podataka, a sastoji se od pronalaska rotacije i translacije koje omogućuju preklapanje dva oblaka točaka u jedinstvenu i povezanu cjelinu. Ovaj diplomski rad pruža pregled osnovnih koraka klasičnog postupka registracije 3D oblaka točaka, s posebnim naglaskom na Iterativna najbliža točka (Iterative Closest Point, ICP) algoritam, koji se često koristi za rješavanje ovog problema. Tijekom rada provedeno je istraživanje literature o primjeni dubokog učenja u registraciji 3D podataka. Identificirani su najvažniji modeli dubokog učenja te su podijeljeni u skupine prema njihovim glavnim karakteristikama. Iz svake skupine odabran je po jedan reprezentativni model, a za daljnje istraživanje odabrani su modeli DCP (Deep Closest Point) i PointNetLK. Ovi modeli uspoređeni su međusobno i s klasičnim ICP algoritmom koristeći skup podataka ModelNet40. Eksperimentalni dio rada obuhvaća evaluaciju učinkovitosti odabranih metoda prema kriterijima točnosti registracije. Rezultati istraživanja ističu prednosti i ograničenja klasičnih i dubokih pristupa registraciji oblaka točaka te pružaju smjernice za daljnji razvoj ovog područja. Zaključci rada doprinose boljem razumijevanju problema registracije oblaka točaka i ukazuju na potencijal primjene dubokog učenja za unaprjeđenje ovog procesa.

Abstract (english)

Point cloud registration is one of the key challenges in 3D data processing. It involves determining the rotation and translation that align two point clouds into a single, cohesive representation. This thesis provides an overview of the fundamental steps in the classical 3D point cloud registration process, with a particular focus on the Iterative Closest Point (ICP) algorithm, which is widely used to address this problem. A literature review was conducted to explore the application of deep learning in 3D point cloud registration. The most significant deep learning models were identified and categorized based on their main characteristics. From each category, one representative model was selected, with DCP (Deep Closest Point) and PointNetLK chosen for further investigation. These models were compared against each other and the classical ICP algorithm using the ModelNet40 dataset. The experimental part of the thesis includes an evaluation of the selected methods based on registration accuracy. The results highlight the strengths and limitations of both classical and deep learning-based approaches to point cloud registration, providing insights for further development in this area. The findings of this thesis contribute to a better understanding of the point cloud registration problem and emphasize the potential of deep learning to enhance this process.

• Keywords: registracija oblaka točaka; iterativna najbliža točka (ICP); Deep Closest Point (DCP); PointNetLK
• Keywords (english): point cloud registration; Iterative Closest Point (ICP); Deep Closest Point (DCP); PointNetLK
• Language: croatian
• URN:NBN: urn:nbn:hr:168:446480
• Study programme: Title: Computing; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra inženjer/inženjerka računarstva (magistar/magistra inženjer/inženjerka računarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Open access
• Created on: 2025-03-12 11:13:43