氏名　Aryuanto Soetedjo
学位の種類　博士（工学）
学位記番号　博甲第383号
学位授与の日付　平成18年8月31日
学位論文題目　A Study on Fast and Robust Algorithm for Traffic Sign Recognition System　（道路交通標識認識システムのための高速かつ頑健なアルゴリズムに関する研究）
論文審査委員
　主査　教授　山田　耕一
　副査　教授　大里　有生
　副査　教授　中村　和男
　副査　教授　淺井　達雄
　副査　助教授　岩橋　政宏

• 論文目次
• 論文要旨
• 論文審査の結果の要旨

［平成18（2006）年度博士論文題名一覧］ ［博士論文題名一覧］に戻る．

Table of Contents

Title page　p.i
Acknowledgments　p.ii
Abstract　p.iii
Table of contents　p.v
List of figures　p.ix
List of tables　p.xii

Chapter1 Introduction

1.1 Introduction　p.1
1.2 Purposes of this research　p.7
1.3 Organization of the dissertation　p.7

Chapter2 Overview of Traffic Sign Recognition System

2.1 Introduction　p.9
2.2 Traffic sign detection　p.11
　2.2.1 Detection by color　p.12
　2.2.1.1 Detection using HSI（HSV）color space　p.12
　2.2.1.2 Detection using RGB color space　p.17
　2.2.2 Detection by shape　p.20
　2.2.3 Detection by joint modeling of color and shape　p.24
　2.2.3.1 Color extraction　p.24
　2.2.3.2 Shape extraction　p.25
　2.2.3.3 Color-shape integration　p.26
2.3 Traffic sign classification　p.27
　2.3.1 Template matching　p.28
　2.3.2 Statistical pattern recognition　p.29
　2.3.3 Neural network　p.31
2.4 Traffic sign tracking　p.34
　2.4.1 Parameter modeling　p.34
　2.4.2 Kalman filter　p.36

Chapter3 Ellipse Detection

3.1 Hough Transform　p.38
3.2 Ellipse detection using Randomized Hough Transform　p.41
3.3 Ellipse detection using Genetic Algorithm　p.43
　3.3.1 Individual representation　p.43
　3.3.2 Genetic operators　p.44
　3.3.3 Fitness function　p.45
　3.3.4 GA execution　p.45

Chapter4 Efficient Algorithm for Traffic Sign Detection

4.1 Introduction　p.48
4.2 Traffic sign detection　p.49
　4.2.1 Overview　p.49
　4.2.2 Edge point extraction　p.50
　4.2.3 Ellipse extraction　p.51
　4.2.3.1 Geometric fragmentation　p.51
　4.2.3.2 Search for fragments　p.53
　4.2.4 Objective function　p.55
　4.2.5 Combination　p.57
4.3 Experimental results　p.58
　4.3.1 Artificial images　p.59
　4.3.2 Real scene images　p.62
　4.3.3 Effect of red threshold　p.65
4.4 Conclusions　p.67

Chapter5 Traffic Sign Classification using Ring Partitioned Method

5.1 Introduction　p.68
5.2 Pre-processing　p.69
5.3 Ring-partitioned matching　p.71
　5.3.1 Ring-partitioned　p.71
　5.3.2 Fuzzy histogram　p.79
　5.3.3 Matching strategies　p.79
5.4 Experimental results　p.80
5.5 Conclusions　p.86

Chapter6 A New Approach for Traffic Sign Tracking from Image Sequences

6.1 Introduction　p.87
6.2 Overview of traffic sign tracking　p.89
6.3 Blog tracking　p.92
　6.3.1 Blob　p.92
　6.3.2 Blob matching　p.93
　6.3.3 Tracking algorithm　p.94
6.4 Experimental results and discussions　p.95
6.5 Conclusions　p.103

Chapter7 Efficient Red Color Thresholding for Traffic Sign Recognition

7.1 Introduction　p.105
7.2 Proposed red color thresholding　p.107
　7.2.1 CIE-RGB chromaticity diagram　p.107
　7.2.2 Red color thresholding based on CIE-RGB chromaticty diagram　p.109
　7.2.3 g-r histogram　p.111
　7.2.4 Measurement of thresholding quality　p.112
7.3 Experimental results　p.114
7.4 Conclusions　p.117

Chapter8 Conclusions and Future Research

8.1 Summary of the research　p.119
8.2 Recommendation for future research　p.121

References　p.123
List of Author's Publications　p.129

　In this research, a fast and robust algorithm for traffic sign recognition systems is newly developed. The proposed methods consist of a fast and efficient traffic sign detection technique, a traffic classification technique using ring- partitioned matching, and a traffic sign tracking technique using blob matching.
Moreover, a red color thresholding technique is newly proposed for segmenting red color from an RGB color image.
　In the detection stage, a technique called geometric fragmentation is used to detect red circular traffic signs in an image by finding and combining the left and right fragments of elliptical objects, which increases the accuracy of detection and copes with occlusion. The search for fragments resembles a genetic algorithm （GA） in the sense that it uses the terms such as individual, population, crossover and objective function used in the GA. The difference from the GA is that it conducts a concurrent random search in a small two-dimensional space devised heuristically. The objective function for evaluating individuals is devised to increase the detection accuracy and reduce the computation time. The algorithm was evaluated for detecting red circular traffic signs both from artificial sign images and real scene images.
Experimental results demonstrated that the proposed algorithm has higher detection rates, fewer false alarms and lower computation cost than GA-based ellipse detection. Compared to conventional template matching, the proposed algorithm performs better in detection and execution time and does not require a large number of carefully prepared templates.
　In the classification stage, the proposed ring-partitioned matching method uses a specified grayscale image in the pre-processing step and ring-partitioned matching in the matching step. This method does not need carefully prepared many samples of traffic sign images for the training process, alternatively only standard signs are used as the reference images. Experimental results showed the effectiveness of the method in matching of occluded, rotated, and illumination problems of traffic sign images with the fast computation time.
　To improve the performance of traffic recognition, a new approach for sign tracking from image sequences is proposed. By tracking signs, the search space is reduced and misdetection caused by temporal occlusion or poor quality of image sequences could be suppressed. The method called two-layered blob tracking is employed to track signs from a frame to the next. This tracking does not require an accurate model, which is essential in the Kalman-Filter tracking. Experimental results showed that the proposed approach could track red circular traffic signs from a moving camera effectively, without any restrictions on speed and movement of the vehicle, and camera installation.
　In the detection, classification and tracking techniques which are proposed, a red color thresholding is employed in the initial process to extract red circular sign from RGB color image. To improve the performance of this thresholding method, the new red color thresholding method based on the CIE-RGB chromaticity diagram is proposed. Compared to other red color thresholding methods, the proposed method has the best thresholding performance among the others, while the computation time is inexpensive.

　本論文は，A Study on Fast and Robust Algorithm for Traffic Sign Recognition System（道路交通標識認識システムのための高速かつ頑健なアルゴリズムに関する研究）と題し，8章より構成される．第1章「Introduction」では，自動車運転支援システムにおける道路交通標識認識の位置づけを示すとともに，道路交通標識認識における課題を整理し，本研究の目的を述べている．
　第2章「Overview of Traffic Sign Recognition System」では，道路交通標識認識のプロセスを，標識の発見と分類の2つのステージに分け，それぞれのステージにおける既存研究を概括し未解決の問題点を整理する．動画像での標識追跡を利用した標識発見研究についても整理する．
　第3章「Ellipse Detection」では，画像中から楕円形（円形）を発見する既存の各種研究を紹介し，それを道路交通標識認識に適用する場合の問題点をまとめる．
　第4章「Efficient Algorithm for Traffic Sign Detection」では，geometric fragmentationと名付けた道路交通標識発見アルゴリズムを提案する．障害物の陰にある赤い楕円形の標識の発見率を向上させるため，楕円形の右断片と左断片を組み合わせることで標識を発見する．断片の発見には遺伝的アルゴリズムに類似する，より処理の高速な発見的手法を用いる．実画像を用いて実験を行った結果，既存の各種手法より発見率が高く，処理速度も速いことが確認された．
　第5章「Traffic Sign Classification using Ring Partitioned Method」は，画像中から発見された標識を認識する新しいRing-partitioned methodを提案する．この方法は，機械学習を用いる既存研究と異なり，注意深く選択した多くの学習用標識を必要とせず，標準的な標識のみを使用する．また，認識精度および計算速度の点でも既存の各種手法に比べて優れていることが示された．
　第6章「A New Approach for Traffic Sign Tracking from Image Sequences」は動画像における標識追跡方法を提案する．従来手法は追跡対象の正確な数学モデルを必要とし，車載カメラから標識を撮影する本研究で利用するには多くの難点があった．提案手法は対象に関する数学モデルを必要とせず，移動する車の動きや速度の変化に関する制約もない．実験の結果，標識の追跡を効果的に行うことが確認された．
　第7章「Efficient Red Color Thresholding for Traffic Sign Recognition」では，CIE-RGB色度図を用いて画像中から赤色部分をより正確に抽出する方法を提案する．この方法を標識発見アルゴリズムに適用すると，より高速で正確な道路標識発見が可能になる．また，第8章「Conclusions and Future Research」では，本研究全体の成果をまとめている．
　よって、本論文は工学上及び工業上貢献するところが大きく、博士（工学）の学位論文として十分な価値を有するものと認める。