1. UKM Learning and Research Repository
2. PHD Thesis / Tesis PHD
3. Faculty of Information Science and Technology / Fakulti Teknologi dan Sains Maklumat

Title:	Multiple visual descriptor combination for loop closure detection and visual odometer trajectory estimation
Authors:	Mohammed Omar Moh'd Salameh (P68462)
Supervisor:	Azizi Abdullah, Prof. Madya Dr.
Keywords:	Robots -- Control systems SLAM (Computer program language) Universiti Kebangsaan Malaysia -- Dissertations Dissertations, Academic -- Malaysia
Issue Date:	2-Feb-2018
Description:	Memory management is one of the crucial elements in Visual Simultaneous Localization and Mapping (VSLAM) for long-time autonomous navigation in a real environment. The real environment involves landmark variations which create challenges to VSLAM for a robot to recognise the visited landmark locations and estimate its trajectory. In VSLAM, many algorithms utilised a single descriptor for describing landmarks for loop closure detection (LCD). However, LCD using single descriptors may hinder the recognition of landmarks and may worsen LCD performance as a database of image maps grows. Thus, this work proposes an Ensemble Bayesian Filter for Active Locations (EBF-AL) for LCD. The EBF-AL is based on the Real-Time Appearance-Based Mapping (RTAB-Map) model to deal with the growing of image maps. In RTAB-Map, it manages the active locations in Working Memory (WM) for LCD, and stores the passive locations in Long Term Memory (LTM). However, some relevant of passive locations that are transferred to LTM may be undetected by LCD. Thus, a proposed algorithm, namely Ensemble Bayesian Filter for Active and Passive Locations (EBF-APL) is used to manipulate information from LTM and WM for LCD. Besides LCD, one important element that utilises visual descriptor for VSALM is visual odometry trajectory estimation (VOTE). Similar to LCD, the most widely used solution for VOTE is arguably using a single keypoints descriptor. However, VOTE using a single keypoints detector seems to be unreliable due to image variation problems in finding the best corresponding features in other target images. Therefore, this research proposes a method that used a random sampling scheme. The scheme extracts the best keypoints from a different type of keypoints detector to reduce the trajectory estimation errors. The proposed algorithms then evaluated on several benchmark datasets namely Lip6 Indoor, Lip6 Outdoor and City Centre for LCD and KITTI for VOTE. The results show that EBF-AL and EBF-APL outperformed the standard RTAB-Map that uses WM for LCD. EBF-AL achieves a recall of 80%, 97% and 86% and EBF-AL score a recall of 91%, 98% and 88% respectively on the Lip6 Indoor, Lip6 Outdoor and City Centre data sets respectively. In trajectory estimation experiment, the proposed algorithm can eliminate the trajectory error of 44%, 8% and 13% on KITTI dataset for the sequence 00, 02 and 05 respectively.,Ph.D.
Pages:	192
Call Number:	TJ211.35.S245 2018 3 tesis
Publisher:	UKM, Bangi
Appears in Collections:	Faculty of Information Science and Technology / Fakulti Teknologi dan Sains Maklumat

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.