Document Type

Thesis

Degree Name

Master of Applied Computing

Department

Physics and Computer Science

Program Name/Specialization

Applied Computing

Faculty/School

Faculty of Science

First Advisor

Safaa Bedawi

Advisor Role

providing support and guidance

Abstract

Object Tracking is one of the essential tasks in computer vision domain as it has numerous applications in various fields, such as human-computer interaction, video surveillance, augmented reality, and robotics. Object Tracking refers to the process of detecting and locating the target object in a series of frames in a video. The state-of-the-art for tracking-by-detection framework is typically made up of two steps to track the target object. The first step is drawing multiple samples near the target region of the previous frame. The second step is classifying each sample as either the target object or the background. Visual object tracking remains one of the most challenging task due to variations in visual data such as target occlusion, background clutter, illumination changes, scale changes, as well as challenges stem from the tracking problem including fast motion, out of view, motion blur, deformation, and in and out planar rotation. These challenges continue to be tackled by researchers as they investigate more effective algorithms that are able to track any object under various changing conditions. To keep the research community motivated, there are several annual tracker benchmarking competitions organized to consolidate performance measures and evaluation protocols in different tracking subfields such as Visual Object Tracking VOT challenges and The Multiple Object Tracking MOT Challenges [1, 2]. Despite the excellent performance achieved with deep learning, modern deep tracking methods are still limited in several aspects. The variety of appearance changes over time remains a problem for deep trackers, owing to spatial overlap between positive samples. Furthermore, existing methods require high computational load and suffer from slow running speed.

Recently, Generative Adversarial Networks (GANs) have shown excellent results in solving a variety of computer vision problems, making them attractive in investigating their potential use in achieving better results in other computer vision applications, namely, visual object tracking. In this thesis, we explore the impact of using Residual Network ResNet as an alternative feature extractor to Visual Geometry Group VGG which is commonly used in literature. Furthermore, we attempt to address the limitations of object tracking by exploiting the ongoing advancement in Generative Adversarial Networks. We describe a generative adversarial network intended to improve the tracker’s classifier during the online training phase. The network generates adaptive masks to augment the positive samples detected by the convolutional layer of the tracker’s model in order to improve the model’s classifier by making the samples more difficult. Then we integrate this network with Multi-Domain Convolutional Neural Network (MDNet) tracker and present the results. Furthermore, we introduce a novel tracker, MDResNet, by substituting the convolutional layers of MDNet that were originally taken from Visual Geometry Group (VGG-M) network with layers taken from Residual Deep Network (ResNet-50) and the results are compared. We also introduce a new tracker, Region of Interest with Adversarial Learning (ROIAL), by integrating the generative adversarial network with the Real-Time Multi-Domain Convolutional Network (RT-MDNet) tracker. We also integrate the GAN network with MDResNet and MDNet and compare the results with ROIAL.

Convocation Year

2019

Convocation Season

Fall

Download

Share

COinS