基于特征學(xué)習(xí)的視頻行人檢測(cè)
發(fā)布時(shí)間:2021-04-11 10:18
在論文中實(shí)驗(yàn)結(jié)果表明了我們的貢獻(xiàn),我們的方法是基于提高訓(xùn)練集候選框的對(duì)齊質(zhì)量和完美多幀檢測(cè)器,這兩種方法通過手動(dòng)剔除Caltech數(shù)據(jù)集的訓(xùn)練標(biāo)簽,并對(duì)剩余的訓(xùn)練樣本進(jìn)行測(cè)試。同時(shí)我們研究了用于行人檢測(cè)的卷積神經(jīng)網(wǎng)絡(luò),并描述了影響它們性能的因素。為了說明背景/前景的區(qū)別,我們研究了用于行人檢測(cè)的卷積神經(jīng)網(wǎng)絡(luò),并描述了影響它們性能的因素。我們?cè)敿?xì)研究并報(bào)告了在Caltech數(shù)據(jù)集的最好的表現(xiàn)性能,并提出了一套新的經(jīng)過篩選的訓(xùn)練和測(cè)試樣本。行人檢測(cè)是一個(gè)眾所周知的計(jì)算機(jī)視覺研究領(lǐng)域的子課題,它的應(yīng)用十分廣泛,雖然已經(jīng)被提出很多年,仍然是一個(gè)研究的重點(diǎn)問題。我們對(duì)比了現(xiàn)在的最先進(jìn)的方法和“完美的多幀檢測(cè)器和候選框的對(duì)齊”的不同。受行人檢測(cè)最新發(fā)展的啟發(fā),我們創(chuàng)建了一個(gè)行人的檢測(cè)的基準(zhǔn)(在Caltech數(shù)據(jù)集上)。不僅可以實(shí)現(xiàn)定位,也可以得到前景和背景的誤差。同時(shí)為了說明定位誤差,我們研究了訓(xùn)練標(biāo)注噪聲對(duì)檢測(cè)器性能的影響,并表明我們可以在少部分經(jīng)過篩選的訓(xùn)練數(shù)據(jù)的基礎(chǔ)上進(jìn)行改進(jìn)。盡管針對(duì)行人檢測(cè)的研究已經(jīng)有了很深入的研究,但是關(guān)于行人檢測(cè)的研究還有很大的發(fā)展空間,還有很多可以繼續(xù)深入研究的方面...
【文章來源】:北京交通大學(xué)北京市 211工程院校 教育部直屬院校
【文章頁(yè)數(shù)】:63 頁(yè)
【學(xué)位級(jí)別】:碩士
【文章目錄】:
Acknowledgement
摘要
ABSTRACT
序言 Preface
Abbreviations
1 Introduction
1.1 Problem
1.2 Objectives and Problem Statement
1.3 Machine Learning and Convolutional Neural Network
1.4 Structure of Thesis
2 Related Work
2.1 Classical Features
2.2 HOG & LUV Features
2.3 Convolutional Neural Networks (CNN)
2.4 Fast Feature Pyramids
3 Methodology
3.1 Convolutional Neural Network
3.1.1 layer Typess
3.2 Training CNN
3.2.1 Dropout
3.2.2 Random Dropout
3.2.3 Rectified Linear unit Layer
3.2.4 Positive and Negative Data Ratio
3.3 Proposed Network
3.3.1 Version One
3.3.2 Version Two
3.3.3 Version Three
3.3.4 Version Four
3.3.5 Depth
3.4 Data
3.4.1 Data Collection
3.4.2 Dataset Augmentation
3.4.3 Data Usage
3.5 The First Network Proposal
3.5.1 12-Network Structures
3.5.2 16-Network Structures
4 Results
4.1 12-Network Results
4.1.1 12-Network Performance
4.1.2 12-Network Complexity
4.2 16-Network Results
4.2.1 16-Network Performance
4.2.2 16-Network Complexity
4.2.3 16-network Discussion
4.3 First Network Comparison
4.4 Comparing Against State-of-the-Art
4.5 State-of-the-Art Discussion
5 Conclusion
5.1 Conclusion
5.2 Future Work
參考文獻(xiàn) References
附錄A Appendix A
索引 INDEX
作者簡(jiǎn)歷及攻讀碩士/博士學(xué)位期間取得的研究成果 AUTHOR PROFILE AND RESEARCH ACHIEVEMENTS OBTAINEDDURING THE STUDY FOR A MASTER'S/DOCTORAL DEGREE
學(xué)位論文數(shù)據(jù)集 DATASET FOR THE MASTER'S THESIS
本文編號(hào):3131075
【文章來源】:北京交通大學(xué)北京市 211工程院校 教育部直屬院校
【文章頁(yè)數(shù)】:63 頁(yè)
【學(xué)位級(jí)別】:碩士
【文章目錄】:
Acknowledgement
摘要
ABSTRACT
序言 Preface
Abbreviations
1 Introduction
1.1 Problem
1.2 Objectives and Problem Statement
1.3 Machine Learning and Convolutional Neural Network
1.4 Structure of Thesis
2 Related Work
2.1 Classical Features
2.2 HOG & LUV Features
2.3 Convolutional Neural Networks (CNN)
2.4 Fast Feature Pyramids
3 Methodology
3.1 Convolutional Neural Network
3.1.1 layer Typess
3.2 Training CNN
3.2.1 Dropout
3.2.2 Random Dropout
3.2.3 Rectified Linear unit Layer
3.2.4 Positive and Negative Data Ratio
3.3 Proposed Network
3.3.1 Version One
3.3.2 Version Two
3.3.3 Version Three
3.3.4 Version Four
3.3.5 Depth
3.4 Data
3.4.1 Data Collection
3.4.2 Dataset Augmentation
3.4.3 Data Usage
3.5 The First Network Proposal
3.5.1 12-Network Structures
3.5.2 16-Network Structures
4 Results
4.1 12-Network Results
4.1.1 12-Network Performance
4.1.2 12-Network Complexity
4.2 16-Network Results
4.2.1 16-Network Performance
4.2.2 16-Network Complexity
4.2.3 16-network Discussion
4.3 First Network Comparison
4.4 Comparing Against State-of-the-Art
4.5 State-of-the-Art Discussion
5 Conclusion
5.1 Conclusion
5.2 Future Work
參考文獻(xiàn) References
附錄A Appendix A
索引 INDEX
作者簡(jiǎn)歷及攻讀碩士/博士學(xué)位期間取得的研究成果 AUTHOR PROFILE AND RESEARCH ACHIEVEMENTS OBTAINEDDURING THE STUDY FOR A MASTER'S/DOCTORAL DEGREE
學(xué)位論文數(shù)據(jù)集 DATASET FOR THE MASTER'S THESIS
本文編號(hào):3131075
本文鏈接:http://sikaile.net/kejilunwen/zidonghuakongzhilunwen/3131075.html
最近更新
教材專著