在视频中使用机器学习分析动作行为和异常检测（Machine Learning Methods for Behaviour Ana....pdf

发布时间：2022-05-31 发布人：admin 分类：说明书资料大小：4.30M 资料格式：pdf 举报版权申诉

u011669700-10266363-4744302543393514291.pdf-第1页.png

第1页 / 共144页

u011669700-10266363-4744302543393514291.pdf-第2页.png

第2页 / 共144页

u011669700-10266363-4744302543393514291.pdf-第3页.png

第3页 / 共144页

u011669700-10266363-4744302543393514291.pdf-第4页.png

第4页 / 共144页

u011669700-10266363-4744302543393514291.pdf-第5页.png

第5页 / 共144页

u011669700-10266363-4744302543393514291.pdf-第6页.png

第6页 / 共144页

u011669700-10266363-4744302543393514291.pdf-第7页.png

第7页 / 共144页

u011669700-10266363-4744302543393514291.pdf-第8页.png

第8页 / 共144页

Supervisor’s Foreword

Abstract

Acknowledgements

Contents

Acronyms

Notation

List of Figures

List of Tables

1 Introduction

1.1 Abnormal Behaviour Detection

1.1.1 Topic Modeling

1.1.2 Change Point Detection

1.2 Key Contributions and Outline

1.3 Disseminated Results

References

2 Background

2.1 Outline of Video Processing Methods

2.1.1 Object Detection

2.1.2 Object Tracking

2.2 Anomaly Detection

2.2.1 Video Representation

2.2.2 Behaviour Model

2.2.3 Normality Measure

2.3 Topic Modeling

2.3.1 Problem Formulation

2.3.2 Inference

2.3.3 Extensions of Conventional Models

2.3.4 Dynamic Topic Models

2.3.5 Topic Modeling Applied to Video Analytics

2.4 Change Point Detection

2.4.1 Change Point Detection in Time Series Data

2.4.2 Anomaly as Change Point Detection

2.5 Summary

References

3 Proposed Learning Algorithms for Markov Clustering Topic Model

3.1 Video Representation

3.2 Model

3.2.1 Motivation

3.2.2 Model Formulation

3.3 Parameter Learning

3.3.1 Expectation-Maximisation Learning

3.3.2 Variational Inference

3.3.3 Gibbs Sampling

3.3.4 Similarities and Differences of the Learning Algorithms

3.4 Anomaly Detection

3.4.1 Abnormal Documents Detection

3.4.2 Localisation of Anomalies

3.5 Performance Validation

3.5.1 Performance Measure

3.5.2 Parameter Learning

3.5.3 Anomaly Detection

3.6 Summary

References

4 Dynamic Hierarchical Dirichlet Process

4.1 Hierarchical Dirichlet Process Topic Model

4.1.1 Chinese Restaurant Franchise

4.2 Proposed Dynamic Hierarchical Dirichlet Process Topic Model

4.3 Inference

4.3.1 Batch Collapsed Gibbs Sampling

4.3.2 Online Inference

4.4 Anomaly Detection

4.5 Experiments

4.5.1 Synthetic Data

4.5.2 Real Video Data

4.6 Summary

References

5 Change Point Detection with Gaussian Processes

5.1 Problem Formulation

5.1.1 Data Model

5.1.2 Change Point Detection Problem Formulation

5.2 Gaussian Process Change Point Detection Approach Based on Likelihood Ratio Tests

5.2.1 Likelihood Ratio Test

5.2.2 Generalised Likelihood Ratio Test

5.2.3 Discussion

5.3 Gaussian Process Online Change Point Detection Approach Based on Likelihood Estimation

5.3.1 Test Formulation

5.3.2 Theoretical Evaluation of the Test

5.3.3 Test with Estimated Hyperparameters

5.3.4 Discussion

5.4 Performance Validation on Synthetic Data

5.4.1 Data Simulated by the Proposed Generative Model

5.4.2 Data Simulated by the GP-BOCPD Model

5.5 Numerical Experiments with Real Data

5.6 Summary

References

6 Conclusions and Future Work

6.1 Summary of Methods and Contributions

6.2 Suggestions for Future Work

6.2.1 Inference in Topic Modeling

6.2.2 Alternative Dynamics in Topic Modeling

6.2.3 Gaussian Process Change Point Detection

6.2.4 Potential Applications of the Proposed Statistical Methods

References

A EM for MCTM Derivation

Appendix B VB for MCTM Derivation

Appendix C Distributions of Quadratic Forms

C.1 Quadratic form of the ``Own'' Covariance Matrix

C.2 Quadratic form of an Arbitrary Symmetric Matrix

Appendix D Proofs of the Theorems for the Proposed Test Statistic

D.1 Proof of Theorem 5.1

D.2 Proof of Theorem 5.2

Appendix E Optimisation of Gaussian Process Covariance Function Hyperparameters

References

Springer Theses Recognizing Outstanding Ph.D. Research Olga Isupova Machine Learning Methods for Behaviour Analysis and Anomaly Detection in Video

Springer Theses Recognizing Outstanding Ph.D. Research

Aims and Scope The series “Springer Theses” brings together a selection of the very best Ph.D. theses from around the world and across the physical sciences. Nominated and endorsed by two recognized specialists, each published volume has been selected for its scientiﬁc excellence and the high impact of its contents for the pertinent ﬁeld of research. For greater accessibility to non-specialists, the published versions include an extended introduction, as well as a foreword by the student’s supervisor explaining the special relevance of the work for the ﬁeld. As a whole, the series will provide a valuable resource both for newcomers to the research ﬁelds described, and for other scientists seeking detailed background information on special questions. Finally, the valuable contributions made by today’s younger generation of scientists. it provides an accredited documentation of Theses are accepted into the series by invited nomination only and must fulﬁll all of the following criteria They must be written in good English. The topic should fall within the conﬁnes of Chemistry, Physics, Earth Sciences, Engineering and related interdisciplinary ﬁelds such as Materials, Nanoscience, Chemical Engineering, Complex Systems and Biophysics. The work reported in the thesis must represent a signiﬁcant scientiﬁc advance. If the thesis includes previously published material, permission to reproduce this must be gained from the respective copyright holder. They must have been examined and passed during the 12 months prior to nomination. Each thesis should include a foreword by the supervisor outlining the signiﬁ- cance of its content. The theses should have a clearly deﬁned structure including an introduction accessible to scientists not expert in that particular ﬁeld. More information about this series at http://www.springer.com/series/8790

Olga Isupova Machine Learning Methods for Behaviour Analysis and Anomaly Detection in Video Doctoral Thesis accepted by the University of Shefﬁeld, Shefﬁeld, UK 123

Author Dr. Olga Isupova Department of Engineering Science University of Oxford Oxford UK Supervisor Prof. Lyudmila Mihaylova University of Shefﬁeld Shefﬁeld UK ISSN 2190-5053 Springer Theses ISBN 978-3-319-75507-6 https://doi.org/10.1007/978-3-319-75508-3 ISBN 978-3-319-75508-3 (eBook) ISSN 2190-5061 (electronic) Library of Congress Control Number: 2018931499 © Springer International Publishing AG, part of Springer Nature 2018 This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the material is concerned, speciﬁcally the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microﬁlms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed. The use of general descriptive names, registered names, in this publication does not imply, even in the absence of a speciﬁc statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. The publisher, the authors and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, express or implied, with respect to the material contained herein or for any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps and institutional afﬁliations. trademarks, service marks, etc. Printed on acid-free paper This Springer imprint is published by the registered company Springer International Publishing AG part of Springer Nature The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland

Supervisor’s Foreword Autonomy implies safety, robustness and quick decision-making. With increasing amounts of data, it becomes almost impossible for a human to analyse and process these vast data streams. Volume, velocity, veracity and variety are some of the challenges that autonomy brings. Despite the signiﬁcant progress made in this area, there is still a big gap between the required level of autonomy and what the technology currently provides. Autonomous change detection and anomaly detection are needed in many areas, from smartphones to drones, autonomous cars to cybersecurity, and video analytical systems. In this context, the Ph.D. thesis of Olga Isupova makes important theo- retical and methodological contributions. It presents machine learning approaches able to detect changes in video streams and to replace the human involvement. The thesis develops Bayesian nonparametric inference methods for anomaly detection and behaviour analysis and represents a step change compared with existing approaches from this actively investigated area in several aspects. The methods involve elements of learning which consist of normal behaviour and are able to detect anything contradicting this normal behaviour. The thesis makes signiﬁcant progress in processing high-dimensional video data and provides a topic modelling approach for detecting any kind of anomalies, without requiring these to be speciﬁed in advance. The thesis proposes hierarchical Dirichlet process state models which assume that data consists of an unknown number, with potentially inﬁnite series of switching activities and behaviours. When newly upcoming data contradicts learnt series of behaviours, an anomaly is detected. The second signiﬁcant contribution of the thesis is in the developed Gaussian process method for change detection and in corroborating the theoretical under- pinnings. Both methods have been applied to anomaly detection and localisation in video streams, and their efﬁciency for autonomous online processing has been demonstrated. The proposed methods are general and can be applied to various applications where data does not have a clear structure nor easily predictable representations. v

vi Supervisor’s Foreword In addition, the thesis is accompanied with open-source toolboxes which can be used by researchers and engineers. I am extremely grateful to Springer Theses Series for providing this opportunity so that the research results of Olga Isupova can reach a wide audience. Shefﬁeld, UK November 2017 Prof. Lyudmila Mihaylova

Abstract Behaviour analysis and anomaly detection are key components of intelligent vision systems. Anomaly detection can be considered from two perspectives: abnormal events can be deﬁned as those that violate typical activities or as a sudden change in behaviour. Topic modelling and change point detection methodologies, respec- tively, are employed to achieve these objectives. The thesis starts with development of novel learning algorithms for a dynamic topic model. Topics extracted by the learning algorithms represent typical activities happening within an observed scene. These typical activities are used for likelihood computation. The likelihood serves as a normality measure in anomaly detection decision-making. A novel anomaly localisation procedure is proposed. In the considered dynamic topic model, a number of topics, i.e. typical activities, should be speciﬁed in advance. A novel dynamic nonparametric hierarchical Dirichlet process topic model is then developed where the number of topics is determined from data. Conventional posterior inference algorithms require pro- cessing of the whole data through several passes. It is computationally intractable for massive or sequential data. Therefore, batch and online inference algorithms for the proposed model are developed. A novel normality measure is derived for decision-making in anomaly detection. The latter part of the thesis considers behaviour analysis and anomaly detection within the change point detection methodology. A novel general framework for change point detection is introduced. Gaussian process time series data is consid- ered, and a change is deﬁned as an alteration in hyperparameters of the Gaussian process prior. The problem is formulated in the context of statistical hypothesis testing, and several tests suitable for both ofﬂine and online data processing and multiple change point detection are proposed. Theoretical properties of the pro- posed tests are derived based on the distribution of the test statistics. vii

分享到：

赞收藏

资料库

在视频中使用机器学习分析动作行为和异常检测（Machine Learning Methods for Behaviour Ana....pdf

相关推荐

人工智能

热门标签

最新资料