A Study on Real-Time Object Detection and Re-Identification
发布时间：2025-06-19 02:48
监视系统安装了大量室内和室外监视摄像机,是最常见的大数据源之一。数据源生成大量的原始可视化数据,使得分析过程成为繁重的人工任务。监视视频分析中最重要的功能之一就是目标检测,以此作为重识别任务的先决条件。为此,大量算法和技术被开发并安装在智能监视系统上,特别是在基于深度学习的方法取得了显著进步之后。但是,在开发可实时运行的算法时,检测算法的计算成本仍然是一项重大挑战。此外,在目标重识别的有效性方面,现有文献算法与实际工业应用上的需求仍然存在很大差距。与人脸识别不同,行人和车辆重识别算法在监视系统中仍不能显示出可靠的性能。本文的主要目的是为户外监控录像中的目标检测和重新识别开发更有效的学习方案。为此,本文介绍了三种主要的基于深度学习的模型,其中一种模型用于合理解决室外监控视频中小型行人的实时检测问题,另外两种用于对象重识别的新的深度学习方案。本论文的核心贡献和创新概括如下:(1)目前监视摄像机已被广泛使用,然而在某些情况下,为了扩大覆盖范围,它们的安装距离很远,这使得视频中的行人以不同的大小出现。为了更好地检测由子弹监视摄像机采集的视频帧中的行人,设计有效的检测算法显得十分必要。此外,由于速...
【文章页数】：141 页
【学位级别】：博士
【文章目录】：
摘要
ABSTRACT
LIST OF ABBREVIATIONS
CHAPTER 1: INTRODUCTION
1.1 Overview
1.2 Motivation of the Study
1.3 Aim and Objectives
1.3.1 First Phase: Object Detection
1.3.2 Second Phase: Object Re-Identification
1.4 Thesis Contributions
1.5 Thesis Organization
CHAPTER 2: BACKGROUND AND LITERATURE REVIEW
2.1 Computer Vision
2.1.1 Traditional Computer Vision Methods
2.1.2 Deep-Learning-Based Computer Vision Methods
2.2 Surveillance Video Analysis
2.3 Real-Time Processing of Deep Learning Algorithms
2.4 First Phase: Object Detection
2.4.1 General Object Detectors
2.4.2 Custom Detectors (Pedestrian/Vehicle)
2.4.3 Limitations of Object Detectors in Literature
2.5 Second Phase: Object Re-Identification
2.5.1 Limitations of Object Re-identification Methods in Literature
2.6 Performance Criteria and Metrics
2.6.1 Object Detection
2.6.2 Pedestrian Detection
2.6.3 Person/Vehicle Re-Identification
2.6.4 Face Verification and Identification
2.7 Chapter Summary
CHAPTER 3: AUTO-ZOOMING CNN-BASED FRAMEWORK FOR REAL-TIMEPEDESTRIAN DETECTION IN OUTDOOR SURVEILLANCE VIDEOS
3.1 Introduction
3.2 Problem Statement
3.3 The Major Contributions
3.4 The Proposed Detection Framework
3.4.1 Auto-Zooming-Based Detection
3.4.2 Tile Selection
3.4.3 Lightweight CNN-Based Pedestrian Detector
3.4.4 Lightweight Detectors Parameters
3.4.5 Tile Stitching and Bounding Box Post-Processing
3.5 Pedestrian Detector Settings
3.5.1 Outdoor Surveillance Pedestrian Dataset
3.5.2 Training Lightweight Pedestrian Detector
3.6 Experimental Results and Discussions
3.6.1 Impact of Input Aspect Ratio on Pedestrian Detection Performance
3.6.2 Lightweight Detector Performance against Customized One-Stage Detectors
3.6.3 Zooming Technique Evaluation
3.6.4 Zooming Lightweight Detector against Two-Stage Pedestrian Detector
3.6.5 Proposed Auto-Zooming Framework in Real-World Scenario
3.7 Chapter Summary
CHAPTER 4: SUPERVISED VARIATIONAL REPRESENTATION LEARNING FORRE-IDENTIFICATION AND VERIFICATION
4.1 Introduction
4.2 Problem Statement
4.3 The Major Contributions
4.4 Supervised Variational Representation Learning
4.4.1 KL Divergence for VAE and VFL
4.4.2 VFL as Normalization and Regularization Technique
4.5 VFL Evaluation for Person Re-Identification
4.5.1 Person Re-Identification Datasets
4.5.2 Default Training Settings for Person Re-Identification
4.5.3 Evaluation of VFL with Different CNN Baselines for Person Re-Identification
4.5.4 Impact of Pre-trained CNN Baselines on VFL
4.5.5 Evaluation of VFL Model against Recent Leading Methods for Person Re-Identification
4.6 VFL Evaluation for Face Verification
4.6.1 Face Recognition Datasets
4.6.2 Default Training Settings for Face Verification
4.6.3 Evaluation of VFL with Different CNN Baselines for Face Verification
4.6.4 Impact of Feature Embedding Size of the VFL in Face Verification
4.6.5 Performance of VFL against Recent Method for Face Verification
4.7 VFL Evaluation for Vehicle Re-identification
4.7.1 Vehicle Re-Identification Datasets
4.7.2 Default training settings of Vehicle re-identification
4.7.3 Evaluation of VFL with Different CNN Baselines for Vehicle Re-Identification
4.7.4 Evaluation of VFL Model against Recent Leading Methods for Vehicle Re-Identification
4.7.5 Impact of the Feature Embedding Size of the VFL in Person/Vehicle Re-Identification
4.8 Variational Representation Learning from Multi-Vehicle Viewpoints
4.8.1 Input Pre-processing and Augmentation
4.8.2 CNN Feature Extraction
4.8.3 Variational Feature Learning
4.8.4 Viewpoint Learning and Encoding with LSTM Network
4.8.5 Evaluation of the Extended Framework for Vehicle Re-Identification in VeRi-776
4.9 Chapter Summary
CHAPTER 5: MULTI-LABEL-BASED SIMILARITY LEARNING FOR VEHICLERE-IDENTIFICATION
5.1 Introduction
5.2 Problem Statement
5.3 The Major Contributions
5.4 Multi-Label-Based Similarity Learning Framework
5.4.1 Baseline Network
5.4.2 Feature Learning
5.4.3 Siamese Network
5.4.4 Absolute Similarity Distance
5.4.5 Sigmoidal Mapping Units
5.4.6 Multi-Label Similarity Loss
5.5 Training and Settings
5.5.1 Datasets
5.5.2 Model Training
5.6 Experimental Results
5.6.1 Modules' Analysis of the Proposed Model
5.6.2 Single-Label against Multi-Label Similarity Learning
5.6.3 Feature Learning, Similarity Learning, and Joint Learning
5.6.4 Similarity Mapping
5.6.5 Impact of Distance Calculation Scheme
5.6.6 Impact of Dropout
5.6.7 Performance of MLSL Model against Recent Related Methods
5.7 Chapter Summary
CHAPTER 6: CONCLUSION AND FUTURE WORK
6.1 Contributions Revised
6.2 Future Work
REFERENCES
PUBLICATION LIST
ACKNOWLEDGMENT
附件
本文编号：4050737
【文章页数】：141 页
【学位级别】：博士
【文章目录】：
摘要
ABSTRACT
LIST OF ABBREVIATIONS
CHAPTER 1: INTRODUCTION
1.1 Overview
1.2 Motivation of the Study
1.3 Aim and Objectives
1.3.1 First Phase: Object Detection
1.3.2 Second Phase: Object Re-Identification
1.4 Thesis Contributions
1.5 Thesis Organization
CHAPTER 2: BACKGROUND AND LITERATURE REVIEW
2.1 Computer Vision
2.1.1 Traditional Computer Vision Methods
2.1.2 Deep-Learning-Based Computer Vision Methods
2.2 Surveillance Video Analysis
2.3 Real-Time Processing of Deep Learning Algorithms
2.4 First Phase: Object Detection
2.4.1 General Object Detectors
2.4.2 Custom Detectors (Pedestrian/Vehicle)
2.4.3 Limitations of Object Detectors in Literature
2.5 Second Phase: Object Re-Identification
2.5.1 Limitations of Object Re-identification Methods in Literature
2.6 Performance Criteria and Metrics
2.6.1 Object Detection
2.6.2 Pedestrian Detection
2.6.3 Person/Vehicle Re-Identification
2.6.4 Face Verification and Identification
2.7 Chapter Summary
CHAPTER 3: AUTO-ZOOMING CNN-BASED FRAMEWORK FOR REAL-TIMEPEDESTRIAN DETECTION IN OUTDOOR SURVEILLANCE VIDEOS
3.1 Introduction
3.2 Problem Statement
3.3 The Major Contributions
3.4 The Proposed Detection Framework
3.4.1 Auto-Zooming-Based Detection
3.4.2 Tile Selection
3.4.3 Lightweight CNN-Based Pedestrian Detector
3.4.4 Lightweight Detectors Parameters
3.4.5 Tile Stitching and Bounding Box Post-Processing
3.5 Pedestrian Detector Settings
3.5.1 Outdoor Surveillance Pedestrian Dataset
3.5.2 Training Lightweight Pedestrian Detector
3.6 Experimental Results and Discussions
3.6.1 Impact of Input Aspect Ratio on Pedestrian Detection Performance
3.6.2 Lightweight Detector Performance against Customized One-Stage Detectors
3.6.3 Zooming Technique Evaluation
3.6.4 Zooming Lightweight Detector against Two-Stage Pedestrian Detector
3.6.5 Proposed Auto-Zooming Framework in Real-World Scenario
3.7 Chapter Summary
CHAPTER 4: SUPERVISED VARIATIONAL REPRESENTATION LEARNING FORRE-IDENTIFICATION AND VERIFICATION
4.1 Introduction
4.2 Problem Statement
4.3 The Major Contributions
4.4 Supervised Variational Representation Learning
4.4.1 KL Divergence for VAE and VFL
4.4.2 VFL as Normalization and Regularization Technique
4.5 VFL Evaluation for Person Re-Identification
4.5.1 Person Re-Identification Datasets
4.5.2 Default Training Settings for Person Re-Identification
4.5.3 Evaluation of VFL with Different CNN Baselines for Person Re-Identification
4.5.4 Impact of Pre-trained CNN Baselines on VFL
4.5.5 Evaluation of VFL Model against Recent Leading Methods for Person Re-Identification
4.6 VFL Evaluation for Face Verification
4.6.1 Face Recognition Datasets
4.6.2 Default Training Settings for Face Verification
4.6.3 Evaluation of VFL with Different CNN Baselines for Face Verification
4.6.4 Impact of Feature Embedding Size of the VFL in Face Verification
4.6.5 Performance of VFL against Recent Method for Face Verification
4.7 VFL Evaluation for Vehicle Re-identification
4.7.1 Vehicle Re-Identification Datasets
4.7.2 Default training settings of Vehicle re-identification
4.7.3 Evaluation of VFL with Different CNN Baselines for Vehicle Re-Identification
4.7.4 Evaluation of VFL Model against Recent Leading Methods for Vehicle Re-Identification
4.7.5 Impact of the Feature Embedding Size of the VFL in Person/Vehicle Re-Identification
4.8 Variational Representation Learning from Multi-Vehicle Viewpoints
4.8.1 Input Pre-processing and Augmentation
4.8.2 CNN Feature Extraction
4.8.3 Variational Feature Learning
4.8.4 Viewpoint Learning and Encoding with LSTM Network
4.8.5 Evaluation of the Extended Framework for Vehicle Re-Identification in VeRi-776
4.9 Chapter Summary
CHAPTER 5: MULTI-LABEL-BASED SIMILARITY LEARNING FOR VEHICLERE-IDENTIFICATION
5.1 Introduction
5.2 Problem Statement
5.3 The Major Contributions
5.4 Multi-Label-Based Similarity Learning Framework
5.4.1 Baseline Network
5.4.2 Feature Learning
5.4.3 Siamese Network
5.4.4 Absolute Similarity Distance
5.4.5 Sigmoidal Mapping Units
5.4.6 Multi-Label Similarity Loss
5.5 Training and Settings
5.5.1 Datasets
5.5.2 Model Training
5.6 Experimental Results
5.6.1 Modules' Analysis of the Proposed Model
5.6.2 Single-Label against Multi-Label Similarity Learning
5.6.3 Feature Learning, Similarity Learning, and Joint Learning
5.6.4 Similarity Mapping
5.6.5 Impact of Distance Calculation Scheme
5.6.6 Impact of Dropout
5.6.7 Performance of MLSL Model against Recent Related Methods
5.7 Chapter Summary
CHAPTER 6: CONCLUSION AND FUTURE WORK
6.1 Contributions Revised
6.2 Future Work
REFERENCES
PUBLICATION LIST
ACKNOWLEDGMENT
附件
本文编号：4050737
本文链接：https://www.wllwen.com/shoufeilunwen/xxkjbs/4050737.html
最近更新
教材专著
