Language:
English
繁體中文
Help
圖資館首頁
Login
Back
Switch To:
Labeled
|
MARC Mode
|
ISBD
語意驅動式HOG行人偵測 = Semantic-Driven HOG-B...
~
國立高雄大學電機工程學系碩士班
語意驅動式HOG行人偵測 = Semantic-Driven HOG-Based Pedestrian Detection
Record Type:
Language materials, printed : monographic
Paralel Title:
Semantic-Driven HOG-Based Pedestrian Detection
Author:
毛銓毅,
Secondary Intellectual Responsibility:
國立高雄大學
Place of Publication:
[高雄市]
Published:
撰者;
Year of Publication:
2011[民100]
Description:
67面圖,表 : 30公分;
Subject:
HOG
Subject:
HOG
Online resource:
http://handle.ncl.edu.tw/11296/ndltd/54050616948629447450
Notes:
參考書目:面55-57
Summary:
本研究提出結合全域特徵與區域特徵的語意驅動式 HOG 行人偵測,為了克服複雜背景與部分遮蔽下偵測困難的情況,整合不同類型的特徵以增進偵測的可靠度。在全域特徵方面,我們採用距離轉換樣板比對的方式,首先從一組正負訓練樣本內,進行行人邊緣輪廓資訊的提取,並使用樣板比對技術,從樣板比對的分數來評估樣板與待測物體輪廓的差距,統計全部的分類結果,從正樣本與負樣本的分布中找出適當的門檻值,作為全域偵測器的判斷標準;在區域特徵方面則是採用方向梯度直方圖與支援向量機,方向梯度直方圖是一種梯度基礎的演算法,能將影像中的小區塊像素資訊轉換為36-D 的向量,每張訓練樣本中可各自產生不同的向量資訊,這些向量會對應高維度空間中的一點,我們使用支援向量機來找出能夠在這個空間中區分兩類資訊的超平面,這個超平面可視為支援向量機的判斷準則。為了結合全域與區域資訊,我們採用Adaboost 機器學習作為整合的核心架構。在訓練階段,選出該回合最佳弱分類器之後,將訓練樣本以全域分類器做分群,分為有人與無人,依據全域的分類結果調整弱分類器的門檻值,這個調整的過程會持續至訓練回合結束為止,產生全域區域整合的強分類器。 We proposed Semantic-Driven HOG-Based Pedestrian Detection. To overcome the difficulty of detecting people in complex background or partial occlusion, we integrate global and local features to enhance reliability. In terns of global feature, we adopt DT-based template matching technique. Calculating matching score to evaluate distance between template and test object. To gather statistics of classification results, finding suitable threshold, and take as global decision rule. In terns of local feature, we adopt HOG and SVM, HOG is one kind of gradient based algorithm. It can turn one block of pixels into 36-D vector, and the vector corresponded to the point in a 36-D hyperspace. We use SVM to find a hyperplane, which is the decision rule of weak classifier. In order to fuse local and global information, we use adaboost as core architecture. In training step, when the best weak classifier selected, training samples separated into two sets with human or non-human. We adjust the hyperplane of local detector according to global result. The process continued until training round ended. Finally, produce global and local integrated strong classifier.
語意驅動式HOG行人偵測 = Semantic-Driven HOG-Based Pedestrian Detection
毛, 銓毅
語意驅動式HOG行人偵測
= Semantic-Driven HOG-Based Pedestrian Detection / 毛銓毅撰 - [高雄市] : 撰者, 2011[民100]. - 67面 ; 圖,表 ; 30公分.
參考書目:面55-57.
HOGHOG
語意驅動式HOG行人偵測 = Semantic-Driven HOG-Based Pedestrian Detection
LDR
:03654nam0a2200277 450
001
322942
005
20170214100458.0
009
322942
010
0
$b
精裝
010
0
$b
平裝
100
$a
20170214y2011 k y0chiy50 e
101
1
$a
chi
$d
chi
$d
eng
102
$a
tw
105
$a
ak am 000yy
200
1
$a
語意驅動式HOG行人偵測
$d
Semantic-Driven HOG-Based Pedestrian Detection
$z
eng
$f
毛銓毅撰
210
$a
[高雄市]
$c
撰者
$d
2011[民100]
215
0
$a
67面
$c
圖,表
$d
30公分
300
$a
參考書目:面55-57
314
$a
指導教授:蕭培墉博士
328
$a
碩士論文--國立高雄大學電機工程學系碩士班
330
$a
本研究提出結合全域特徵與區域特徵的語意驅動式 HOG 行人偵測,為了克服複雜背景與部分遮蔽下偵測困難的情況,整合不同類型的特徵以增進偵測的可靠度。在全域特徵方面,我們採用距離轉換樣板比對的方式,首先從一組正負訓練樣本內,進行行人邊緣輪廓資訊的提取,並使用樣板比對技術,從樣板比對的分數來評估樣板與待測物體輪廓的差距,統計全部的分類結果,從正樣本與負樣本的分布中找出適當的門檻值,作為全域偵測器的判斷標準;在區域特徵方面則是採用方向梯度直方圖與支援向量機,方向梯度直方圖是一種梯度基礎的演算法,能將影像中的小區塊像素資訊轉換為36-D 的向量,每張訓練樣本中可各自產生不同的向量資訊,這些向量會對應高維度空間中的一點,我們使用支援向量機來找出能夠在這個空間中區分兩類資訊的超平面,這個超平面可視為支援向量機的判斷準則。為了結合全域與區域資訊,我們採用Adaboost 機器學習作為整合的核心架構。在訓練階段,選出該回合最佳弱分類器之後,將訓練樣本以全域分類器做分群,分為有人與無人,依據全域的分類結果調整弱分類器的門檻值,這個調整的過程會持續至訓練回合結束為止,產生全域區域整合的強分類器。 We proposed Semantic-Driven HOG-Based Pedestrian Detection. To overcome the difficulty of detecting people in complex background or partial occlusion, we integrate global and local features to enhance reliability. In terns of global feature, we adopt DT-based template matching technique. Calculating matching score to evaluate distance between template and test object. To gather statistics of classification results, finding suitable threshold, and take as global decision rule. In terns of local feature, we adopt HOG and SVM, HOG is one kind of gradient based algorithm. It can turn one block of pixels into 36-D vector, and the vector corresponded to the point in a 36-D hyperspace. We use SVM to find a hyperplane, which is the decision rule of weak classifier. In order to fuse local and global information, we use adaboost as core architecture. In training step, when the best weak classifier selected, training samples separated into two sets with human or non-human. We adjust the hyperplane of local detector according to global result. The process continued until training round ended. Finally, produce global and local integrated strong classifier.
510
1
$a
Semantic-Driven HOG-Based Pedestrian Detection
$z
eng
610
0
$a
HOG
$a
SVM
$a
Adaboost
$a
行人偵測
$a
全域特徵
$a
區域特徵
610
1
$a
HOG
$a
SVM
$a
Adaboost
$a
pedestrian detection
$a
local feature
$a
global feature
681
$a
008M/0019
$b
542201 2080
$v
2007年版
700
1
$a
毛
$b
銓毅
$4
撰
$3
546965
712
0 2
$a
國立高雄大學
$b
電機工程學系碩士班
$3
166118
801
0
$a
tw
$b
NUK
$c
20120426
$g
CCR
856
7
$z
電子資源
$2
http
$u
http://handle.ncl.edu.tw/11296/ndltd/54050616948629447450
based on 0 review(s)
ALL
博碩士論文區(二樓)
Items
2 records • Pages 1 •
1
Inventory Number
Location Name
Item Class
Material type
Call number
Usage Class
Loan Status
No. of reservations
Opac note
Attachments
310002199373
博碩士論文區(二樓)
不外借資料
學位論文
TH 008M/0019 542201 2080 2011
一般使用(Normal)
On shelf
0
310002199381
博碩士論文區(二樓)
不外借資料
學位論文
TH 008M/0019 542201 2080 2011 c.2
一般使用(Normal)
On shelf
0
2 records • Pages 1 •
1
Multimedia
Multimedia file
http://handle.ncl.edu.tw/11296/ndltd/54050616948629447450
Reviews
Add a review
and share your thoughts with other readers
Export
pickup library
Processing
...
Change password
Login