Razvoj sistema za prepoznavanje objektov na osnovi oblakov točk

Štepec, Dejan (author), Skočaj, Danijel (mentor)

V diplomskem delu se posvečamo prepoznavanju objektov na osnovi oblakov točk. Sistem najprej implementiramo z različnimi 3D opisniki ter naredimo temeljito analizo s preučevanjem rezultatov dobljenih na uveljavljeni zbirki oblakov točk. Ovrednotimo tudi, kako na klasifikacijsko točnost vpliva redukcija dimenzionalnosti opisnika ter predstavimo različne načine selekcije ključnih točk. Rezultate primerjamo z rezultati sistema, ki prepoznava kategorijo objektov iz slik. Sistem ovrednotimo na različno velikih množicah objektov ter na primeru zelo podobnih objektov, kjer pokažemo slabosti opisnikov, ki temeljijo zgolj na 3D informaciji. Na podlagi rezultatov, doseženih na zbirki, implementiramo sistem za prepoznavanje objektov na osnovi oblakov točk v realnem času. Sistem implementiramo tako, da omogoča gradnjo baze objektov, ki jih želimo prepoznavati. Sistem ovrednotimo na bazi 18 objektov, ki jo zgradimo v ta namen. Rezultati pokažejo, da lahko zgolj z informacijo o obliki objekta dosežemo dobre rezultate.

• Language: Slovenian
• Keywords: prepoznavanje objektov, 3D, vreča besed, PFH, FPFH, VFH, spin slike, SVM
• Work type: Bachelor thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2015
• PID: 20.500.12556/RUL-72446
• Publication date in RUL: 17.09.2015

Secondary language

• Language: English
• Title: Development of a system for point cloud based object recognition

Abstract

In this thesis we address the problem of point cloud based 3D object recognition. First we develop the system with different 3D shape descriptors and then make a thorough analysis of results gathered from testing the system on a publicly available benchmark database. We also evaluate how dimensionality reduction and different methods for feature point selection influence on classification accuracy. The system is compared with the system for object recognition in 2D. We evaluate the system on different number of objects, including on objects that are very similar by shape. Based on results from the dataset, we develop a real-time system for object recognition. The system allows us to build our own database of objects that we want to recognize. We constructed a database of 18 objects on which we evaluated the system. Results presented in this thesis have shown that we can recognize objects sufficiently good only with 3D shape information.

• Keywords: object recognition, 3D, bag of words, PFH, FPFH, VFH, Spin Images, SVM