Publications

Detailed Information

Autonomous Navigation by Domain Adaptation with Generative Adversarial Networks : 생성적 대립 신경망과 도메인 적응 기법을 이용한 자율 주행 방법 연구

DC Field Value Language
dc.contributor.advisor윤성로-
dc.contributor.author홍용준-
dc.date.accessioned2018-12-03T01:35:35Z-
dc.date.available2018-12-03T01:35:35Z-
dc.date.issued2018-08-
dc.identifier.other000000152176-
dc.identifier.urihttps://hdl.handle.net/10371/143665-
dc.description학위논문 (석사)-- 서울대학교 대학원 : 공과대학 전기·컴퓨터공학부, 2018. 8. 윤성로.-
dc.description.abstract대립적 생성 신경망은 (GANs) 중간 지도 학습, 이미지 전이, 도메인 적응 등 과 같은 머신 러닝의 중요한 문제들에 대한 중요한 해법으로 각광받고 있다. 특히, 대립적 생성 신경망은 기존의 생성 모델에 비해 확률 분포에 대한 가정 없이 더욱 구체적이고 실제같은 데이터를 생성할 수 있다. 본 논문에서는 로보틱스, 기계 공학, 머신 러닝 등에서 중요한 문제였던 자율 주행에 대립적 생성 신경망을 적용하였다. 깊은 인공신경망을 자율 주행에 이용한 기존 접근 방법들은 데이터에 대한 라벨 정보을 많이 모은 후, 인공신경망을 지도 학습 방법으로 훈련시켰다. 하지만 라벨 정보가 있는 거대한 데이터 집합은 이러한 방법에 필수 불가결하며 라벨 정보가 있는 실제 데이터는 얻기 어렵고 부정확한 경우가 많다.

우리는 이 문제를 대립적 생성 신경망을 이용한 도메인 적응 기법으로 해결하였 다. 시뮬레이션 환경에서는 다양한 종류의 라벨 정보들을 쉽게 얻을 수 있다. 라벨 정보가 있는 인공 데이터를 시뮬레이터로부터 추출하고 도메인 적응 기법으로 실제 환경에서의 라벨 문제를 완화하였다. 더불어, 대립적 생성 신경망을 통해 인공 데이터를 실제 데이터와 비슷하게 만들어 실제 환경에서의 라벨이 없이도 자율 주행이 성공하도록 하였다.
-
dc.description.abstractGenerative adversarial networks (GANs) has received popular attention because of their great potential for important subjects in the machine learning field, such as semi-supervised learning, image translation and domain adaptation. Specifically, GANs are capable of generating more sharp and realistic synthetic data than prior generative models without any probability distribution assumptions. In this paper, we apply GANs to autonomous navigation, which has been a unique subject for various fields such as robotics, mechanical engineering and machine learning. Prior approaches with deep neural networks to autonomous navigation are mostly gathering lots of labeled real data and training neural networks by supervised-learning method. However, large labeled datasets are indispensable for these approaches, and labeled real data is hard to acquire, laborious and often erroneous. We address this problem by domain adaptation with GANs. In simulator environment, various kinds of labels can be easily taken. By exploiting labeled synthetic data from simulator environment, we alleviate labeling issue in real environment through domain adaptation. In addition to that, by GANs, we try to make synthetic data look like real data, so that autonomous navigation can be successfully done without label information of real data.-
dc.description.tableofcontentsAbstract i

Contents ii

List of Tables iv

List of Figures v

1 INTRODUCTION 1

2 BACKGROUND 4

2.1 Generative Adversarial Networks . . . . . . . . . . . . . . . . . . . . 4

2.1.1 Integral Probability Metric . . . . . . . . . . . . . . . . . . . 5

2.1.2 Image to Image Translation . . . . . . . . . . . . . . . . . . 10

2.2 Domain Adaptation . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

2.3 Autonomous Navigation . . . . . . . . . . . . . . . . . . . . . . . . 17

2.3.1 Non-learning Based Approaches for Autonomous Navigation 17

2.3.2 Learning Based Approaches for Autonomous Navigation . . . 17

2.3.3 Employing a Simulator for Autonomous Navigation . . . . . 18

3 METHOD 19

3.1 Problem Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

3.2 Domain Adaptation with Adversarial Learning . . . . . . . . . . . . 20

3.3 Network Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . 21

3.3.1 Generator . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

3.3.2 Discriminator . . . . . . . . . . . . . . . . . . . . . . . . . . 22

3.3.3 Classifier . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

3.4 Objective Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

4 EXPERIMENT 27

4.1 Experiment setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

4.1.1 Steering Command . . . . . . . . . . . . . . . . . . . . . . . 28

4.2 Implementation Details . . . . . . . . . . . . . . . . . . . . . . . . . 28

4.3 Result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

4.3.1 Off-line Test Result . . . . . . . . . . . . . . . . . . . . . . . 29

4.3.2 Outdoor Navigation On-line Test Results . . . . . . . . . . . 30

4.3.3 Image Transfer . . . . . . . . . . . . . . . . . . . . . . . . . 32

5 DISCUSSION 42

5.1 Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

5.1.1 Content Similarity Loss . . . . . . . . . . . . . . . . . . . . 42

5.1.2 Normalization . . . . . . . . . . . . . . . . . . . . . . . . . 43

5.2 Failure Cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

5.3 Future Works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

5.3.1 Regression Problem . . . . . . . . . . . . . . . . . . . . . . 45

5.3.2 Command Condition . . . . . . . . . . . . . . . . . . . . . . 45

5.3.3 Other Vehicles . . . . . . . . . . . . . . . . . . . . . . . . . 46

6 CONCLUSION 47

Abstract(In Korean) 59
-
dc.formatapplication/pdf-
dc.format.mediumapplication/pdf-
dc.language.isoen-
dc.publisher서울대학교 대학원-
dc.subject.ddc621.3-
dc.titleAutonomous Navigation by Domain Adaptation with Generative Adversarial Networks-
dc.title.alternative생성적 대립 신경망과 도메인 적응 기법을 이용한 자율 주행 방법 연구-
dc.typeThesis-
dc.description.degreeMaster-
dc.contributor.affiliation공과대학 전기·컴퓨터공학부-
dc.date.awarded2018-08-
Appears in Collections:
Files in This Item:

Altmetrics

Item View & Download Count

  • mendeley

Items in S-Space are protected by copyright, with all rights reserved, unless otherwise indicated.

Share