Scale-invariant visual tracking using P-N leaening & particle filters / Ahmed Naglah

Supervisor: Mohamed A. El-Helw

Thesis (M.A.)—Nile University, Egypt, 2016 .

"Includes bibliographical references"

Contents:
Chapter 1: Introduction ....................................................................................................... 1
1.1 Motivation & Research Goals ............................................................................. 1
1.2 Contributions ...................................................................................................... 2
1.3 Organization of Thesis ........................................................................................ 3
Chapter 2: Technical background and research work ......................................................... 5
2.1 Overview ............................................................................................................. 5
2.2 Common Classifiers use in Visual Tracking ...................................................... 8
2.3 Image processing as a preparation stage in Visual Tracking ............................ 14
2.4 Some addressed visual tracking approaches ..................................................... 19
2.5 Conclusion:................................................................................................... 28
Chapter 3: Preliminaries ................................................................................................... 29
3.1 Compressive Tracking ....................................................................................... 29
3.2 Tracking Learning and Detection ...................................................................... 36
3.3 Statistical Estimation ......................................................................................... 40
3.4 Conclusion:................................................................................................... 44
Chapter 4: Proposed scale-adaptive visual tracking framework ....................................... 46
4.1 Scale-Adaptive Compressive Tracking ............................................................. 46
4.2 Compressive Sensing with PN Learning ........................................................... 50
xi
4.3 Particle Filters with PN Learning ...................................................................... 53
4.4 Conclusion:................................................................................................... 57
Chapter 5: Results and discussions ................................................................................... 58
5.1 Experiment Setup .............................................................................................. 58
5.1.1 Video Sequences......................................................................................... 58
5.1.2 Benchmark with other Trackers ................................................................. 60
5.2 Evaluation Criteria ............................................................................................ 62
5.3 Qualitative Analysis .......................................................................................... 62
5.4 Quantitative Analysis ........................................................................................ 69
5.5 Conclusion:................................................................................................... 72
Chapter 6: Conclusion & future work ............................................................................... 73
6.1 Review of Contributions ................................................................................... 73
6.2 Conclusion & Future work ................................................................................ 74
REFERENCES .................................................................................................................

Abstract:
ABSTRACT
In this thesis, we will introduce a novel scale-invariant visual tracking framework
using PN learning, particle filters and compressive sensing. Visual tracking can be
considered as a dual-class classification problem. The first class is representing the
object and the second class is representing the background. This classification
problem can be divided into two steps: testing phase and training phase. In order to
get the best out of any classifier, PN learning was commonly used to perform
accuracy boosting for the labels of training samples during the training phase. In our
proposed framework, we enriched the use of PN learning by combining it with
particle filters. Particle filters is a widely-used statistical estimator that been used in
many applications. The particle filters is introduced to our framework by defining
motion model and measurement model. The motion model is defined by the moving
object and the tracker. Meanwhile, the measurement model is defined by the detector.
In our proposed framework, particle filters & PN learning will simultaneously
estimate the hidden state of the visual tracking system and boost the accuracy of the
training phase for the classifier. Building on top of this idea, a motion estimation
mechanism will be introduced in our proposed framework by the use of compressive
sensing. The compressive sensing will be done by using Haar-like features. The idea
here is that instead of using scale-invariant-features to estimate the tracked object
size, we will use the compressed features at different scales then will perform a
second stage of classification to detect the best-fit scale. According to our
experiments, the obtained results demonstrated enhanced tracking accuracy and
reliability compared to the other recent algorithms. As a future work, we advise to use
some statistical models to perform directional-biased particle resampling. We also
advise to use a circular spatial representation to reduce computational cost.

Text in English, abstracts in English.