Publications
Detailed Information
Abnormal Object Detection by Transformed-Canonical Scene Generation : 표준장면 생성을 이용한 비정상 물체 검출
DC Field | Value | Language |
---|---|---|
dc.contributor.advisor | 이경무 | - |
dc.contributor.author | SANGDON PARK | - |
dc.date.accessioned | 2017-07-14T02:47:09Z | - |
dc.date.available | 2017-07-14T02:47:09Z | - |
dc.date.issued | 2012-08 | - |
dc.identifier.other | 000000004152 | - |
dc.identifier.uri | https://hdl.handle.net/10371/122895 | - |
dc.description | 학위논문 (석사)-- 서울대학교 대학원 : 전기·컴퓨터공학부, 2012. 8. 이경무. | - |
dc.description.abstract | 비정상 물체 검출은 미개척 분야이지만 유망한 문제이다. 왜냐하면, 전체 이미지 중에서 비정상 이미지가 차지하는 비율이 늘어나는 추세이고, 영상 감시 시스템에서 응용가능하기 때문이다. 기존의 방법은 정상성을 모델링한 다음에 비정상성을 찾는 두 단계의 알고리즘을 사용하고 있다. 하지만 두 단계의 알고리즘은 모델링을 하는 괒어에서 비정상성을 고려하지 않기 때문에 비정상 물체를 찾는 데 한계를 가지고 있다. 비록 이 두 단계의 알고리즘이 신뢰성이 있다고 가정을 해도, 기존의 방법들은 비정상 물체를 검출하는 데 필요한 조건을 모두 만족하지 않는다. 이런 기존의 방법들의 문제점을 처리하기 위해서 본 논문은 정상 물체뿐만 아니라 비정상 물체도 생성할 수 있는 경치 잠재 변수 생성 모델을 제시한다. 게다가, 우리는 비정상 물체 검출을 성공적으로 수행하기 위해 제시한 네 가지 기준을 분석하였고, 그것을 모델링 과정에서 적용하였다. 게다가 모델의 잠재 변수는 population-based Markov Chain Monte Carlo를 통한 최적화를 이용해 추론된다. 마지막으로 우리는 새로운 비정상 데이터세트를 제안한다. 그 데이터세트는 세 가지 분류(co-occurrence/position/scale)로 나누어져 있어서 제안한 모델의 정확성을 철저하게 평가할 수 있고, 그렇게 평가된 모델은 기존의 방법보다 우수하다는 것을 보였다. | - |
dc.description.abstract | Abnormal object detection is undeveloped but promising problem because the portion
of abnormal images over total images tends to be increasing and it is applicable to visual surveillance systems. Conventional approaches to abnormal object detection adopt two-step algorithm: modeling of normalities through a contextual model and applying to find abnormalities. However, This two-step algorithm hinders to find abnormal objects because the algorithm does not considering abnormalities in modeling phase. Even if the two-step algorithm is assumed to be reliable, conventional approaches do not satisfy all properties which are necessary to properly detect abnormal objects. To handle limitations of conventional approaches, this paper proposes scene generating model which generates not only normal objects but also abnormal ones. In addition, four criteria are analyzed and proposed to successfully handle abnormal object detection and apply the criteria in modeling phase. Furthermore, latent variables of the proposed model are predicted by optimizing via population-based Markov Chain Monte Carlo, which has a relatively short convergence time. New abnormal dataset is presented which is classified into three categories (co-occurrence/position/scale) to throughly measure the accuracy of the proposed model for each categories, demonstrating the superiority of results over the existing approach. | - |
dc.description.tableofcontents | Abstract i
Contents ii List of Tables v List of Figures vi 1 Introduction 1 1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1.2 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 1.3 Thesis Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 2 Related works 7 2.1 Contextual Models for Object Recognition . . . . . . . . . . . . . . . 7 2.1.1 Object-Object Interaction . . . . . . . . . . . . . . . . . . . . 7 2.1.2 Scene-Object Interaction . . . . . . . . . . . . . . . . . . . . 8 2.2 Scene Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 2.2.1 Mid-level Feature based Approaches . . . . . . . . . . . . . . 9 2.2.2 High-level Feature based Approaches . . . . . . . . . . . . . 9 2.3 Abnormal Object Detection . . . . . . . . . . . . . . . . . . . . . . . 10 3 Transformed-Canonical Scene Generating Model 11 3.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 3.2 Image Representations and Assumptions . . . . . . . . . . . . . . . . 15 3.3 Canonical Scene for Location Modeling . . . . . . . . . . . . . . . . 16 3.3.1 Canonical Scene . . . . . . . . . . . . . . . . . . . . . . . . 17 3.3.2 Matching transformation . . . . . . . . . . . . . . . . . . . . 17 3.3.3 Similarity measure . . . . . . . . . . . . . . . . . . . . . . . 17 3.3.4 Connection with the Location Model . . . . . . . . . . . . . 18 3.4 Joint Location and Appearance Model . . . . . . . . . . . . . . . . . 19 4 Parameter Learning 21 4.1 Learning Canonical Scene and transformation . . . . . . . . . . . . . 22 4.2 Learning Appearance model . . . . . . . . . . . . . . . . . . . . . . 24 5 Inference 25 5.1 Population-based MCMC for MAP Inference . . . . . . . . . . . . . 26 5.2 Details on Pop-MCMC for MAP Inference . . . . . . . . . . . . . . . 26 5.2.1 Efficient Evaluation of Target Distribution . . . . . . . . . . . 27 5.2.2 Pop-MCMC Algorithm . . . . . . . . . . . . . . . . . . . . . 27 5.3 Marginals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28 6 Evaluation 31 6.1 Abnormal Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 6.2 Experimental Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 6.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34 6.4 Additional Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 6.4.1 Abnormal Dataset: Co-occurrence . . . . . . . . . . . . . . . 39 6.4.2 Abnormal Dataset: Position . . . . . . . . . . . . . . . . . . 44 6.4.3 Abnormal Dataset: Scale . . . . . . . . . . . . . . . . . . . . 48 7 Conclusion 52 Abstract (In Korean) 58 감사의 글 59 | - |
dc.format | application/pdf | - |
dc.format.extent | 34914974 bytes | - |
dc.format.medium | application/pdf | - |
dc.language.iso | en | - |
dc.publisher | 서울대학교 대학원 | - |
dc.subject | Object Detection | - |
dc.subject | Abnormal Object Detection | - |
dc.subject | Context | - |
dc.subject | Generative Model | - |
dc.subject | Generative Learning | - |
dc.subject | Expectation Maximization (EM) Algorithm | - |
dc.subject.ddc | 621 | - |
dc.title | Abnormal Object Detection by Transformed-Canonical Scene Generation | - |
dc.title.alternative | 표준장면 생성을 이용한 비정상 물체 검출 | - |
dc.type | Thesis | - |
dc.contributor.AlternativeAuthor | 박상돈 | - |
dc.description.degree | Master | - |
dc.citation.pages | LXI, 61 | - |
dc.contributor.affiliation | 공과대학 전기·컴퓨터공학부 | - |
dc.date.awarded | 2012-08 | - |
- Appears in Collections:
- Files in This Item:
Item View & Download Count
Items in S-Space are protected by copyright, with all rights reserved, unless otherwise indicated.