Hybrid Background Modeling based on Local Edge Shape and Color Descriptors

다국어 초록 (Multilingual Abstract)

In recent years, the use of video-based surveillance systems and its applications has increased owing to its indefeasible necessity in ensuring people safety. However, the management of the increasing volume of video data is quite challenging since existing surveillance systems are still being run by manual human-monitoring. To minimize such human involvement, numerous intelligent monitoring agents have been introduced by the researchers in recent past. Moving object (foreground) segmentation is the prime and initial task in most of such systems. The task of segmenting the moving object is challenging in practise because of the inclusion of many un-stationary situations in real-world environments, including gradual or sudden illumination changes, dynamic backgrounds, shadows, noise pixels and many more.
Several methods have been studied to tackle the above-mentioned problems, which can be classified into three major groups: pixel-based methods, edge-based methods, and hybrid methods. The major limitations of these methods can be summarized as, i) pixel-based methods are prone to segmentation-errors due to illumination changes, ii) edge-based methods are sensitive to background motions, and iii) hybrid methods using different background models do not ensure optimal utilization of multiple features simultaneously.
In this thesis, we propose a hybrid background model that relies on both edge and non-edge features of the image in the construction of the model. We also introduce a new coding scheme, Local Hybrid Pattern (LHP) that selectively encodes edge or non-edge feature for each pixel. Furthermore, to represent per-pixel background dynamics, we use an adaptive code dictionary that is updated by adding the stable code while discarding the unstable one. We also use weight with each LHP code of the dictionary to ensure appropriate description of each code. Moreover, to avoid the ambiguous region having irrelevant texture-information, we propose a statistical approach that considers edge and inner information simultaneously with an adaptive region segmentation for the LHP representation. We create three support regions (edge, inner, and ambiguous) through a hysteresis operation on the LHP codes of the incoming frame to select foreground candidates more accurately which helps us to overcome the edge-variation problem. In segmentation step, a pixel is detected as foreground if the LHP code deviates significantly from its corresponding background model (dictionary). Along with the segmentation of foreground or background, our proposed approach also classifies each pixel as edge or inner region simultaneously. We separate foreground regions from noisy background using a morphological approach for each candidate region, and classify them as edge or inner simultaneously.
We validate our proposed algorithms against the illumination changes by performing experiments on several publicly available datasets. Our comparative evaluation also shows that our proposed method performs better than other methods showing less sensitivity to the background motions. Overall, we show the potential of the proposed edge-feature demonstrating better detection of foreground objects.

목차 (Table of Contents)

• Abstract i
• Acknowledgment iii
• Table of Contents v
• List of Figures viii
• List of Tables xvi
• Chapter 1 Introduction 1
• 1.1 Background 1
• 1.2 Motivation 1
• 1.3 Contributions 6
• 1.4 Organization 8
• Chapter 2 Background Study 10
• 2.1 Overview 10
• 2.2 Pixel-based Methods 11
• 2.3 Edge-based Methods 11
• 2.4 Hybrid Methods 12
• Chapter 3 Local Hybrid Pattern 14
• 3.1 Local Hybrid Pattern Generation 15
• 3.1.1 Region Classification 15
• 3.1.2 Local Edge Shape Pattern 16
• 3.1.3 Inner Patterns 17
• 3.2 Adaptive Local Hybrid Pattern 18
• 3.2.1 Adaptive Region Classification 19
• 3.3 Local Adaptive Hybrid Pattern 21
• 3.3.1 Code Definition 23
• 3.3.2 Code matching 25
• Chapter 4 Hybrid Background Modeling 28
• 4.1 Overview 28
• 4.2 Adaptive Dictionary Model 29
• 4.2.1 Code-based Background Modeling 30
• 4.2.2 Background Model Updating 30
• 4.3 Statistical Hybrid Modeling 31
• 4.3.1 Code-based Non-parametric Background Modeling 32
• 4.3.2 Background Model Update 33
• 4.4 Extended Adaptive Dictionary Model 34
• Chapter 5 Foreground Detection 37
• 5.1 Overview 37
• 5.2 Probability Map Generation 37
• 5.2.1 Probability Map based on Adaptive Dictionary Model 38
• 5.2.2 Probability Map based on Statistical Hybrid Modeling 39
• 5.2.3 Probability Map based on Extended ADM 41
• 5.3 Foreground Detection 42
• 5.3.1 Foreground Candidates Extraction 43
• 5.4 Foreground Classification 44
• Chapter 6 Experimental Results 46
• 6.1 Performance Measurement 46
• 6.2 Adaptive Dictionary Model based Method 47
• 6.2.1 LHP Descriptor Selection 48
• 6.2.1.1 Edge Descriptor for LHP 48
• 6.2.1.2 Inner Descriptor for LHP 50
• 6.2.2 Parameter Setting 53
• 6.2.3 Quantitative Results 56
• 6.2.4 Qualitative Results 57
• 6.3 Statistical Hybrid Modeling based Method 65
• 6.3.1 Parameter Setting 66
• 6.3.2 Quantitative Results 66
• 6.3.3 Qualitative Results 70
• 6.4 Extended ADM for Local Adaptive Hybrid Pattern 81
• 6.4.1 Parameter Setting 81
• 6.4.2 Quantitative Results 83
• 6.4.3 Qualitative Results 87
• Chapter 7 Conclusions and Future work 93
• 7.1 Conclusions 93
• 7.2 Future work 94
• Bibliography 96
• Appendix A List of Publications 103