|
Hokkaido University Collection of Scholarly and Academic Papers >
Theses >
博士 (情報科学) >
Co-occurrence Pixel-Block Background Model and its Application to Robust Event Detection
Files in This Item:
|
|
Related Items in HUSCAP:
|
Title: | Co-occurrence Pixel-Block Background Model and its Application to Robust Event Detection |
Other Titles: | 共起ピクセルブロック背景モデルとそのロバストイベント検出への応用 |
Authors: | 周, 文俊 Browse this author |
Keywords: | Co-occurrence pixel-block pairs | CPB | Hypothesis on degradation modication | HoD | Foreground detection | Illumination changes | Background motion |
Issue Date: | 25-Sep-2019 |
Publisher: | Hokkaido University |
Abstract: | As a basic approach utilized in many computer vision applications, foreground detection
plays an important role in various tasks like video surveillance, traffic monitoring, scene
background initialization and object tracking. One simple way to do background model
is to acquire a background image without any moving objects. However, foreground
detection is faced with many practical challenges, especially the background changes,
not least of which is related to illumination changes, e.g. variable sunlight or lights being
switched on and off indoors, and background motion, e.g. the swaying motion of the
trees, eeting cloud and moving waves on the water. To handle such challenges, previous
statical methods have been proposed, in which the intensity of each pixel is independently
analyzed in the temporal domain and then the current frame is subtracted, such as the
Gaussian Mixture Model (GMM) to build a pixel-wise model for each pixel, however such
kind of methods is difficult to solve illumination changes with the intensity varies rapidly
and signi cantly. Recent many local feature based methods have been put forward for
background modeling such as Barnich et al. proposed ViBe, a method that involves
comparing each pixel with a set of previous values located the same or neighboring
positions to evaluate whether a pixel belongs to the background. However, such local
feature based background models are susceptible to be affected by the dynamic motion
of the background, thus losing the robustness.
To overcome above problems, this thesis presents a novel background subtraction
method called Co-occurrence Pixel-Block pairs (CPB) for detecting objects in dynamic
scenes. CPB is a \pixel to block" structural model, which is evolved from the Co-
occurrence Probability based Pixel Pairs (CP3) and it uses the correlation of multiple
co-occurrence pixel block pairs to detect objects in dynamic scenes. It offers robust
background subtraction against a dynamically changing background. We rstly propose
a correlation measure for co-occurrence pixel-block pairs to realize a robust background
model. We then introduce a novel evaluation strategy named correlation depended de-
cision function for accurate object detection with the correlation of co-occurrence pixel-
block pairs. Finally, CPB can estimate the foreground from a dynamic background with
a sensitive criterion. Furthermore, a Hypothesis on Degradation Modi cation (HoD)
based on CPB is proposed to further resist background changes for foreground detec-
tion, such as illumination changes and background motion. HoD provides CPB with
a model update strategy that can be used for a long time. HoD further improves the
robustness of CPB, and stabilizes the efficiency of CPB over time.
Through the experimental comparisons with other existing foreground detection tech-
niques based on challenging datasets, we demonstrated the good performance of our
algorithms. In summary, CPB is sufficiently sensitive to detect foreground objects in
dynamic scenes and CPB performs robust detection in outdoor or indoor environments
with relatively low complexity. Furthermore, HoD provides a new and natural thought:
the structure of background model can be updated by the designed correlation weigh,
which is a new strategy can be utilized in the pixel-correlation based algorithms for the
background model update.
This thesis is organized into the following chapters:
Chapter 1 introduces the related works in foreground detection. Some general prob-
lems are involved and discussed. Furthermore, the motivations and contributions of this
study are described.
Chapter 2 introduces the Co-occurrence Pixel-Block Background Model (CPB) in
detail, including the basic concept and essential mechanism of CPB. As an extension
from the“pixel to pixel”
structure that our previous work CP3, CPB proposes a“pixel-block”
structure for the background model. In this chapter, we describe how to construct
the“pixel-block”
structure for background model and explain the process of modelbuilding in theory.
Chapter 3 discusses the application of CPB in the eld of the foreground (event)
detection. We also introduce a novel evaluation strategy named correlation depended
decision function for accurate foreground detection and explain the theoretical meaning
of the evaluation strategy. Moreover, we do a comparison to present the performance of
CPB for foreground detection.
Chapter 4 focuses on the Hypothesis on Degradation Modi cation (HoD), which
is proposed based on CPB to further improve the robustness of CPB and stabilize the
efficiency of CPB over time. In this chapter, the basic knowledge and mechanism of HoD
are discussed in detail. Finally, we verify the ability of HoD with adequate experiments.
Chapter 5 introduces the experimental setup in detail. In this chapter, the compara-
tive experiments for CPB and CPB+HoD using several challenging datasets are designed
and through these experiments we measure the robustness and efficiency of our methods,
CPB and CPB+HoD in various indoor and outdoor challenges.
The nal Chapter summarizes the main points of the study and discusses our algo-
rithms. Finally, the plan and concept of future work are presented. |
Conffering University: | 北海道大学 |
Degree Report Number: | 甲第13732号 |
Degree Level: | 博士 |
Degree Discipline: | 情報科学 |
Examination Committee Members: | (主査) 教授 金子 俊一, 教授 金井 理, 准教授 田中 孝之 |
Degree Affiliation: | 情報科学研究科(システム情報科学専攻) |
Type: | theses (doctoral) |
URI: | http://hdl.handle.net/2115/75868 |
Appears in Collections: | 課程博士 (Doctorate by way of Advanced Course) > 情報科学院(Graduate School of Information Science and Technology) 学位論文 (Theses) > 博士 (情報科学)
|
|