Integrated Framework For Single Visual Object Tracking

Visual object tracking is a process that imitates the visual perception of a human eye to observe the dynamic configuration of an object or target in visual surveillance. However, under unconstrained and changing environments, the target of the visual surveillance also undergoes appearance variatio...

Full description

Saved in:
Bibliographic Details
Main Author: Ong, Lee Yeng
Format: Thesis
Published: 2019
Subjects:
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Visual object tracking is a process that imitates the visual perception of a human eye to observe the dynamic configuration of an object or target in visual surveillance. However, under unconstrained and changing environments, the target of the visual surveillance also undergoes appearance variation as well as vulnerable to distortion effect. The ultimate aim of a visual object tracker not only has to track a single object accurately but also has to maintain the tracking process until the end of the video sequence. This work aims to devise a single-object tracker based on an integrated framework that specifically incorporates four interrelated modules (appearance modelling, motion modelling, object localisation and multi-level scrutiny) to handle appearance variation and distortion effect. Both primary modules, namely appearance and motion modelling establish the global and local features of visual content with the movement information of a target across the consecutive frames. The outcomes of the primary modules are integrated to verify the tracked location of the target in the object localisation module. Lastly, the distortion effect is closely monitored with the multilevel scrutiny approach before selectively update the appearance variation. The approaches that are proposed for each module are elaborated in the following contributions. In the appearance modelling module, two new region descriptors that exploit moment descriptor using the normalisation (GRD) and weighted (LRD) approaches are proposed.