基于深度学习的快速车辆再识别研究

发布时间：2022-08-01 19:19

　　智能交通系统Intelligent Transportation System（ITS）是保证城市车辆安全、平稳运行的关键系统之一。车辆再识别Re-Identification（Re-Id）是一项重要的工作,其定义为识别不同监控摄像头拍摄的不重叠视野图像中的车辆。换言之,一个摄像头中捕捉到的某个车辆是否出现在多个摄像网络中。随着对自动化视频分析需求的不断增加,车辆再识别正受到越来越多的关注。它能支持许多关键应用,如智能停车、可疑车辆跟踪、车辆事件检测、跨摄像头识别、道路通行限制管理系统和自动收费等。在过去的几年中,各种强大的计算机视觉方法被用来分析车辆再识别任务中监控摄像机的视频。然而,由于需求的特殊性,研究人员在设计鲁棒高效的模型来解决相关问题时面临着很大的挑战,比如类间相似度、视点变化、部分遮挡、类内可变性、背景杂波和跨数据集的车辆再识别等问题;当前提出的模型并不能十分有效地处理上述问题。本文旨在探索解决该问题的不同方法,基于深度学习的技术以获得更好的车辆再识别性能。首先,我们改进了一个快速的交通监控模型来识别摄像网络中发现的不同类型的车辆。采用深卷积神经网络Convolution... 

【文章页数】：132 页

【学位级别】：博士

【文章目录】：
摘要
ABSTRACT
Chapter1 Introduction
    1.1 Research Background and Significance
    1.2 Problem Statement
    1.3 Motivation
    1.4 Research Aims and Objectives
    1.5 Contributions of Dissertation
    1.6 Outline of the Dissertation
Chapter2 Literature Review
    2.1 Intelligent Transportation System(ITS)
        2.1.1 Video Surveillance
    2.2 Detection
    2.3 Recognition
    2.4 Identification
    2.5 Re-Identification
        2.5.1 Person Re-Identification
        2.5.2 Vehicle Re-Identification
        2.5.3 Methods used for Vehicle Re-Identification
        2.5.4 Performance Comparison of Vision-based State-of-the-Art Vehicle Re-Identification Approaches
        2.5.5 Potential Problems and Challenges in Vision-based Vehicle Re-Identification
        2.5.6 Publically Available Vehicle Re-Identification Datasets
        2.5.7 Evaluation Measures for Vehicle Re-Identification System
        2.5.8 Practical Applications of Vehicle Re-Identification System
    2.6 Deep Learning Basics
        2.6.1 Convolutional Neural Network
        2.6.2 Siamese Neural Network
        2.6.3 Transfer Learning
    2.7 Summary
Chapter3 Fast and Deep CNN-Model for Vehicle Type Identification
    3.1 Introduction
    3.2 Proposed Scheme and Methodology
        3.2.1 Transfer Learning using Inception-v3 Model
        3.2.2 Support Vector Machine
        3.2.3 k-Nearest Neighbors
        3.2.4 Convolutional Neural Network(CNN)Model
    3.3 Experimental Procedure
    3.4 Implementation Details
    3.5 Datasets
        3.5.1VeRi-776
        3.5.2 Vehicle ID
        3.5.3 Vehicle Re Id
        3.5.4 MIO-TCD
    3.6 Evaluation Metrics
    3.7 Experimental Results
    3.8 Applications of Traffic Surveillance System
    3.9 Summary
Chapter4 Efficient and Deep Vehicle Re-Id using Multi-Level Feature Extraction
    4.1 Introduction
    4.2 Overview of the Proposed Method
        4.2.1 Appearance Attributes-based Vehicle Re-Identification
        4.2.2 License Plate-based Vehicle Re-Identification
    4.3 Experiment and Analysis
        4.3.1 Dataset
        4.3.2 Implementation Details
    4.4 Experimental Evaluations
        4.4.1 Evaluation of Appearance-based Vehicle Filtering
        4.4.2 Evaluation of License plate-based Vehicle Re-Identification
        4.4.3 Performance Comparison with State-of-the-Art Methods
        4.4.4.Ablation Studies
    4.5 Summary
Chapter5 Visual Features with Spatio-temporal Based Fusion Model for Cross-dataset Fast Vehicle Re-Id
    5.1 Introduction
    5.2 Overview of Proposed Approach
        5.2.1 Data Augmentation
        5.2.2 Siamese Neural Network-based Classifier
        5.2.3 Spatio-temporal Pattern
        5.2.4 Calculation of Composite Similarity Score
    5.3 Experiment and Analysis
        5.3.1 Vehicle Re-Identification Benchmark Datasets
        5.3.2 Implementation Details
        5.3.3 Evaluation Measures
        5.3.4 Vehicle Re-Identification Dataset Classification
        5.3.5 Single Dataset Vehicle Re-Identification
        5.3.6 Cross Dataset Vehicle Re-Identification
    5.4 Discussion
    5.5 Summary
Chapter6 Conclusion and Future Work
    6.1 Concluding Remarks
    6.2 Future Work
Acknowledgement
References
Research Results Achieved During the Study for Doctoral Degree



本文编号：3668028