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

Title:	Soft Assignment Visual Descriptors For Visual Place Recognition
Authors:	Abbas M. Ali (P47845)
Supervisor:	Md. Jan Nordin, Assoc. Prof. Dr.
Keywords:	Soft Assignment Visual Descriptors Visual Place Recognition Soft Assignment For Visual Place Recognition Visual Place Recognition Using Soft Assignment Visual Descriptors Pattern recognition systems
Issue Date:	27-Jun-2013
Description:	Upon increasing the popularity of using the Hard Bag of Features (HBOF) for accurate object and place categorization problem, there are some issues which are still being scrutinized. In fact, most of the previous researches in place recognition area are based on using Histogram descriptors. Based on the literature, these methods have several issues such as the inability to include spatial relation among the local appearance features for representing the scene image in a more informative way. Therefore, the main objective of this research is to improve the performance of the HBOF in visual place recognition by developing spatial relations for Soft assignment features. These features extracted by measuring the distances of patches from the centroids of codebook constructed by clustering SIFT features by K-means. The covariance of minimum distance (CMD) with whitening filters and some normalization parameters are used to increase the accuracy performance. The visual place confusion has been decreased by implementing Entropy of covariance feature vectors (ECV) which is investigated alone and combined with the edge histogram descriptors (EHD) using the conceptual semantic representation. To demonstrate the effectiveness of the proposed approaches in visual place recognition, several experiments have been setup such as CMD, ECV, and Semantic in order to evaluate the accuracy rate. Practically, several comparative studies were conducted with other related approaches namely a conventional BoW or HBOF, Minimum distance table, and Covariance of Distance table. The proposed methods have been evaluated based on different datasets such as IDOL and on real images from a handheld camera taken for some places in FTSM-UKM. Based on the obtained results, the combined features of EHD and ECV bring a significant improvement in the classification accuracy rates up to 98.6% and 93.423% for IDOL and FTSM-UKM dataset respectively.,PhD
Pages:	222
Call Number:	TA1634 .A454 2013 3
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.