1. UKM Learning and Research Repository
2. PHD Thesis / Tesis PHD
3. Faculty of Engineering and Built Environment / Fakulti Kejuruteraan dan Alam Bina

Title:	Attribute-based structured ranking for person re-identification
Authors:	Mohammad Ali Saghafi (P59727)
Supervisor:	Aini Hussain, Prof. Dr.
Keywords:	Universiti Kebangsaan Malaysia -- Dissertations Dissertations, Academic -- Malaysia Person re-identification Classification Image
Issue Date:	6-Nov-2020
Description:	Person re-identification (PRI) is defined as finding the same person across different surveillance cameras with disjoint fields of view. Although several methods have been investigated previously for PRI, several issues were still inefficiently resolved. The issues mostly were due to the use of low-level features and descriptors that are not robust enough against noise, occlusion, varying light and pose of images. Furthermore, the accuracy of the classification stage results is still low. So far, different learned metrics techniques have been deployed for classification purposes. In the majority of the metric learning methods for PRI, point-wise or pair-wise loss functions have been utilized to optimize the metric. However, re-identification is a ranking problem inherently. The main objective of this thesis is to improve the ability of the feature extraction process and then to solve PRI as a ranking optimization problem. This approach is believed to have significant capacity in improving PRI outcome. Human body silhouette is first extracted from the video streams images and are later preprocessed. Next asymmetry axes and pictorial structures body partitioning are applied on segmented silhouettes to detect head, torso and legs. Spatial Chromatic Component Histogram (SCCH) descriptor was then applied to extract colour and spatial information of partitioned silhouettes. Afterwards, visual attributes using HSV (Hue, Saturation and Value), Texton and SIFT (Scale-Invariant Feature Transform) descriptors are extracted from the partitioned labelled silhouette parts and trained with SVM (Support Vector Machine) classifiers. Opposite attribute detection scheme and balanced data training were applied to improve detection accuracy. Finally, structured SVM is used to analyse PRI as a ranking problem. In this framework called Structural Attribute Ranking (SAR), Normalized Discounted Cumulative Gain (NDCG) is used as the loss function of the ranking problem. The use of visual attributes enables the images to be ranked based on their relevance to the query. In this novel view to PRI, a total order ranking based on relevancy had been successfully developed. SCCH deployment resulted on 25% of correct matches in CMC (Cumulative Matching Characteristic) first rank on VIPeR dataset and 30% 1st rank correct match on local dataset. Experiments on VIPeR and PRID datasets using proposed structured ranking scheme resulted 47.7% and 39% of correct matches in CMC 1st rank respectively that outperformed classifications based on current learning metric. Opposite training schemes helped significant improvement of attribute detection with 69.6% on PRID and 75.5% accuracies on VIPeR datasets. In conclusion, this research proved that the use of visual attributes as input features and rank-wise loss functions in the classification process has promising potentials in solving PRI problems compared to existing methods.,Ph.D.
Pages:	121
Publisher:	UKM, Bangi
Appears in Collections:	Faculty of Engineering and Built Environment / Fakulti Kejuruteraan dan Alam Bina

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