Determination of Pedestrian Attributes from Motion Imagery Monitoring Using a Soft Computing Approach (動作映像からのソフトコンピューティングアプローチによる歩行者属性の判定)

氏名 Handri Santoso
学位の種類 博士(工学)
学位記番号 博甲第478号
学位授与の日付 平成20年8月31日
学位論文題目 Determination of Pedestrian Attributes from Motion Imagery Monitoring Using a Soft Computing Approach (動作映像からのソフトコンピューティングアプローチによる歩行者属性の判定)
 主査 教授 中村 和男
 副査 教授 山田 耕一
 副査 教授 大里 有生
 副査 准教授 岩橋 政宏
 副査 准教授 マーラシンハ チャンドラジット アーシュボーダ

平成20(2008)年度博士論文題名一覧] [博士論文題名一覧]に戻る.

Abstract p.iii
Table of Contents p.vi
List of Figures p.ix
List of Tables p.xii
CHAPTER 1 Introduction
 1.1 General Description p.1
 1.2 Aims of Study p.12
 1.3 Dissertation Outline p.12
CHAPTER 2 Discrimination of Pedestrian from Other Moving Objects
 2.1 Introduction p.14
 2.2 Pedestrian Image Processing p.16
 2.2.1 Pedestrian Image Segmentation p.16
 2.2.2 Pedestrian Shape Analysis p.19
 2.2.3 Extraction of Pedestrian Shape Features p.21
 2.3 Pedestrian Shape Classification by Neural Netwaorks p.24
 2.4 The Experiment and Results p.25
 2.5 Conclusions p.27
CHAPTER 3 Discrimination of Sidewalk Surface Condition Based on Image Texture and Meteorological Information
 3.1 Introduction p.31
 3.2 Background Image Processing p.33
 3.3 Image Texture Extraction p.34
 3.3.1 First Type texture Extraction p.35
 3.3.2 Second Type texture Extraction p.36
 3.3.3 Meteorological Features p.37
 3.4 Discrimination Method p.38
 3.4.1 Factor Analysis p.38
 3.4.2 Artificial Neural Network p.40
 3.5 The Experiment and Results p.42
 3.6 Summary p.44
 3.7 Conclusion p.46
CHAPTER 4 Classification and Choquet Integral Agent Network (CHIAN) Concepts
 4.1 Introduction p.51
 4.2 Formalization of CHIAN p.54
 4.2.1 Hierarchical Structure by CHIAN p.54
 4.2.2 Expression of Each Agent in CHIAN p.54
 4.3 Identification of Fuzzy Measures for a Choquet Integral Agent p.56
 4.4 CHIAN-Based Back-Propagation Algorithms p.58
 4.5 The Experimet p.61
 4.5.1 Multi-Layered CHIAN Structure p.62
 4.5.2 The "At-Least Two" Problem p.63
 4.6 Conclusions p.65
CHAPTER 5 Improving Performance of Choqet Integral Agent Network (CHIAN)
 5.1 Adaptation of Learning Parameters of CHIAN by Using Genetic Algorithms p.67
 5.1.1 The Problem Background p.67
 5.1.2 Optimization Learning Parameters for CHIAN Identification by Genetic Algorithms p.68
 5.1.3 Experiment and Results p.72
 5.1.4 Conclusinos p.75
 5.2 Improving CHIAN Performance by Using Competitive Learning Algorithms p.76
 5.2.1 Problem Background p.76
 5.2.2 Generation Hidden Units of CHIAN by Using Competitive Learning Algorithms p.76
 5.2.3 Architecture of CHIAN with Competitive Learning Algorithms(CHIAN-CL) p.79
 5.2.4 The Experiments and Results p.83 Iris Problem p.83 Wine Problem p.86
 5.2.5 Conclusinos p.89
CHAPTER 6 Classification of Human Age and Gender Based on Motion Imagery by Using CHIAN with Competitive Learning Algorithms
 6.1 Introduction p.90
 6.2 Preprocessing of Human Motion Using Image Processing p.91
 6.3 Feature Extraction and Selection of Human Silhouette p.92
 6.3.1 Extraction of Human Shape-Motion Features p.92
 6.3.2 Selection of Human Motion Features p.94
 6.3.3 The Verification Methods p.95
 6.4 The Experiment and Results p.96
 6.5 Conclusions p.98
CHAPTER 7 Conclusions and Future Research
 7.1 The Summary and Conclusion p.100
 7.2 Recommendations for Future Research p.103
 List of Author's Publications p.106
 Appendix p.113

It is believed that every person has a unique way of walking which corresponds to his/her habit. There have been some attempts to make use of gait as a feature to identify and recognize a person in several applications including biometric, medical rehabilitation, sports, etc. However, only few works have done for pursuing pedestrian safety based on their motion. To cope with that problem, in this study, human mimetic knowledge and thinking are tried to be embedded into computer system based on a soft-computing approach to detect pedestrian attributes from motion imagery. At the front end, image and video processing was performed to separate the foreground from the background images, and then discrimination of pedestrian from other moving objects in the street was developed by using shape features and neural networks. Concurrently, the background image representing sidewalk surface condition is also extracted. The aim of extracting background image is to provide information relating to the safety of a pedestrian. This process included extracting texture of the background image and meteorological information at that time. At the next step, pedestrian conditions, e.g. attributes and behavioral states are to be understood. It can be attained by macroscopic information fusion mechanism based on qualitative features of space-time motion pattern of walking pedestrian. In this study, Choquet Integral Agent Networks (CHIAN) were introduced as macroscopic information fusion mechanisms because of their flexible integration of multiple qualitative input data. Identification algorithms of CHIAN were developed by using back-propagation like concepts. However, the back-propagation methods have some limitations such as trapping at local minima when a given problem has many local minimum solutions. Due to genetic algorithms (GA) mechanism, it has the characteristics finding the optimum solution among multiple local minima, and thus can overcome the difficulty of trapping at local minima; consequently it might reduce network paralysis. Thus, adaptive learning parameters of CHIAN were developed to tune learning rate and momentum coefficient by genetic algorithms to improve CHIAN performance. In addition, CHIAN has also a drawback if it has to handle huge amount of input data which can cause inferior performance in recognition and increase computation cost. In view of these facts, improvement performance of CHIAN by competitive learning algorithm was hereby proposed in this study. Finally, a CHIAN approach for classifying human age and gender are developed based on the pattern of human motion from images sequences. The widths of human shape were analyzed by 2D, i.e., space-time Fourier transform to extract the human motion features. Feature sub-set selection methods were then performed to find the salient and the effective features for the classification process. The experimental results demonstrated the capability of the proposed soft computing approach to classify the age and the gender in high accuracy rate.