指導教授：傅立成臺灣大學：資訊工程學研究所陳逸祥Chen, Yi-HsiangYi-HsiangChen2014-11-262018-07-052014-11-262018-07-052014http://ntur.lib.ntu.edu.tw//handle/246246/261500在智慧型運輸系統領域中，道路障礙偵測一直是一個很重要的議題。行人因為其脆弱的特性，而吸引的許多專家研究此議題。但道路上仍有其他也很脆弱的障礙，如騎著自行車，機車等二輪載具的駕駛。而其中駕駛人的部分並非光源體，且機踏車上也只有少部分光源甚至沒有。由此可知建立一套夜間機踏車駕駛人偵測系統以輔助汽機車駕駛室有其必要性的。因此，本論文提出一個在移動式車輛上使用之基於近紅外線光源的夜間機踏車駕駛偵測方式。此外將機踏車納入偵測對象的一部份時，會因為機踏車的特性而增加許多的困難。例如機踏車的寬高比在不同視角的差異，機踏車駕駛與機車會在不同視角下有不同的相互遮蔽等問題。為此引入了組件式物體假設的偵測方法於本論文中，並利用機踏車明顯的外觀以及近紅外線光源的與影像上呈現的特性對偵測結果進行驗證。最後藉由實驗的結果可以證明提出的系統有著不錯的效果。Obstacle on road detection is an important issue in the field of intelligent transportation system. The research of pedestrian detection attracts attention due to the weakness of feature and descriptor. Some other obstacles, such as the cyclists, having weaker descriptive features are seldom discussed. The cyclist, of course, is not a light source and there are seldom or even no light source fixed on the two-wheel vehicles. Hence, it is notable to build a nighttime cyclist detection system to assist the vehicle driver. Thus, in this thesis, we propose a cyclist detection method for a moving vehicle equipped with a near-infrared camera. It will increase the difficulty when the detected objects including two-wheel vehicles due to some inherent properties of the two-wheel vehicles. For example, the aspect ratio varies depending on different viewpoints. Namely, the occlusion effect between cyclist and two-wheel vehicle varies with the viewpoint. To solve this problem, we employ the part-based object detection in this thesis as the main stream solution approach. Moreover, we use two kinds of additional information to verify the part-based detection result. The first one is the obvious contour appearance of two-wheel vehicle and its interior spatial relation with high stability. The second is the primary characteristic of the NIR image. Finally, experiments show that our system is verified and demonstrated in a nice performance.口試委員會審定書 i 誌謝 ii 中文摘要 iii ABSTRACT iv CONTENTS v LIST OF FIGURES vii LIST OF TABLES ix Chapter 1 Introduction 1 1.1 Motivation 2 1.2 Challenges 3 1.3 Related Work 6 1.4 Contributions 11 1.5 Thesis Organization 12 Chapter 2 Preliminaries 14 2.1 Problem Definition 14 2.2 Histogram of Oriented Gradients (HOG) Feature 16 2.3 Support Vector Machine (SVM) 17 2.3.1 Objective of SVM 18 2.3.2 Overview of SVM 18 2.4 System Overview 21 Chapter 3 Cyclist Part Generation 23 3.1 Cyclist part candidate generation 24 3.2 Cyclist part classifier construction and selection 27 3.3 Cyclist part Spatial Relationship Learning 33 Chapter 4 Near-infrared Part based Nighttime Cyclist Detection 36 4.1 Preprocessing 37 4.2 Cyclist Part Detection and Grouping 40 4.3 Cyclist Detection Result Verification 41 4.3.1 Wheel Spatial Relationship Verification 42 4.3.2 Intensity Verification 44 Chapter 5 Experimental Results 46 5.1 Environment Description 46 5.2 Dataset Description 47 5.3 Detection Result and Performance Evaluation 49 5.3.1 Evaluation Method 49 5.3.2 Cyclist part detection evaluation 50 5.3.3 Histogram specification evaluation 51 5.3.4 System Evaluation 53 Chapter 6 Conclusion 57 References 592474619 bytesapplication/pdf論文公開時間：2017/08/26論文使用權限：同意有償授權(權利金給回饋學校)物體偵測夜間近紅外線幾何資訊空間關係方向梯度直方圖thesishttp://ntur.lib.ntu.edu.tw/bitstream/246246/261500/1/ntu-103-R00922011-1.pdf