A GPU-Accelerated Semi-Supervised Learning-Based Monitoring Systems / Mohamed Mahmoud Mohamed El-Zahhar

Supervisor: Mohamed A. El-Helw

Thesis (M.A.)—Nile University, Egypt, 2012 .

"Includes bibliographical references"

Contents:
Chapter 1: Introduction ........................................................................................................ 1
1.1 Introduction ..................................................................................... 1
1.2 Motivation ....................................................................................... 2
1.3 Objectives ....................................................................................... 2
1.4 Dissertation Overview .................................................................... 2
1.5 Publications and Awards ................................................................. 3
Chapter 2: Background ........................................................................................................ 4
2.1 Introduction ..................................................................................... 4
2.2 Shadow Perception .......................................................................... 5
2.3 Shadow Detection Process .............................................................. 7
2.4 Categories of Shadow Detection Approaches ................................. 7
2.4.1 Geometrical Information .................................................. 9
2.4.2 Spatial Information ........................................................ 10
2.4.3 Temporal Information .................................................... 11
2.4.4 Illumination Change Information (Model-based) .......... 11
2.4.5 Illumination Change Information (non-model-based) ... 12
2.4.5 Edge Information ........................................................... 13
2.5 Semi-Supervised Learning ............................................................ 14
2.5.1 Introduction .................................................................... 14
2.5.2 Popular Semi-Supervised Learning Methods ................ 15
2.6 Ensemble Learning ....................................................................... 17
2.6.1 Introduction .................................................................... 17
2.6.2 Diversity ......................................................................... 17
2.6.3 Design of Multiple Classifier Systems .......................... 17
2.7 Semi-Supervised Multiple Classifier Systems .............................. 18
2.8 Review on Soft Labels: ................................................................. 19
Chapter 3: Framework for Improved Shadow Detection and Removal ............................ 21
3.1 Introduction ................................................................................... 21
3.2 Proposed Semi-Supervised Multiple Classifier Systems .............. 22
3.2.1 Ensemble Driven Self-Training ..................................... 22
3.2.2 Co-Training of Multiple Classifiers ...............................
3.3 Framework Description ................................................................ 24
3.3.1 Background Subtraction................................................. 25
3.3.2 Features Extraction ........................................................ 25
3.3.3 Ensemble-Driven Self-Training for Adaptive Shadow
Detection ................................................................................. 27
3.3.3.1 Initial Training Phase .................................................. 28
3.3.3.2 Shadow Detection Phase ............................................. 29
Chapter 4: Experimental Results ....................................................................................... 30
4.1 Experiments and Analysis of Semi-Supervised Multiple Classifiers
using Soft Labeled Data ...................................................................... 30
4.2 Experiments and Analysis of Shadow Detection and Removal
Framework .......................................................................................... 33
4.2.1 Quantitative Evaluation ................................................. 34
4.2.2 Qualitative Evaluation ................................................... 37
4.3 Conclusions ................................................................................... 39
Chapter 5: GPU-Accelerated Shadow Features Extraction ............................................... 40
5.1 Introduction ................................................................................... 40
5.2 Heterogeneous Compute: .............................................................. 41
5.3 GPU Suitability ............................................................................. 41
5.4 OpenCL Programming Framework .............................................. 42
5.4.1 What is OpenCL? ........................................................... 42
5.4.2 Platform Model .............................................................. 42
5.4.3 Execution Model ............................................................ 44
5.4.4 Memory Model .............................................................. 44
5.5 Features Extraction on GPU ......................................................... 45
5.6 Experimental Results .................................................................... 45
Appendix A: List of OpenCl Kernels Used in Features Extraction ................................... 47
References ..........................................................................................................................

Abstract:
roblems that must be alleviated in order to achieve robust segmentation of moving objects. The computer treats cast shadow as a part of the object during object detection causing errors in segmentation and tracking. There are large number of algorithms that have been proposed for resolving the shadow detection problem.
However, many of these methods tackled the problem in static settings with significant human input and with known scene conditions. In practice, these assumptions are violated as scene conditions change overtime.
This thesis proposes a novel approach for adaptive shadow detection by using semi-supervised learning which is a technique that has been widely utilized in various pattern recognition applications and exploits both labeled and unlabeled data to improve classification. The approach can be summarized as follows: First, we extract color, texture, and gradient features that are useful for differentiating between moving objects and their shadows. Second, we use a semi-supervised learning approach for adaptive shadow detection.
This thesis presents two key contributions. First, the adaptation of semisupervised multiple classifiers learning algorithm to use soft labeled data, as the soft labeled data is more useful and informative than crisp labels in many machine learning applications. Based on this, we derived the second contribution which is a novel GPU-accelerated semi-supervised learning-based approach for cast shadow detection in video sequences. This approach provides robust, flexible and accurate results in outdoor traffic surveillance video sequences under different or changing illumination conditions. The practical values of the proposed technique are the minimized human intervention by using few labeled pixels for initial training, the adaptive learning of shadows by using unlabeled pixels overtime, and the use of GPU for shadow features extraction to accelerate computations. Experiments on benchmark video sequences demonstrate that the proposed framework achieves improved shadow detection (classify shadow points as shadows) and discrimination (not to classify object points as shadows) rates under different scene conditions.

Text in English, abstracts in English.