指導教授：陳柏華臺灣大學：土木工程學研究所張騎Zhang, QiQiZhang2014-11-252018-07-092014-11-252018-07-092014http://ntur.lib.ntu.edu.tw//handle/246246/260747本論文旨在運用影像視覺之方法建立一車載猶豫區間警示系統。透過對路口兩側交通號誌的辨識，獲取猶豫區間計算必要之參數，包括車輛接近路口時之速度，以及路口寬度等資訊。 本系統主要由兩個模組組成。處理與辨識模組主要負責交通號誌之偵測與辨識，猶豫區間判斷模組負責計算各項參數，以及決定是否提供警示訊息。 交通號誌在形狀與顏色上擁有明顯的特徵。在處理與辨識模組中，我們首先對影像的顏色進行分析，分隔出俱有號誌顏色的區域，再結合對形狀的判斷，來得到影像中的交通號誌的候選區域。經由機器學習訓練得到的支持向量機，可以剔除出非號誌的候選區域。我們將對被分類為號誌的候選區域進行進一步的處理。 在比較了現有的對圓形物體的偵測方法之後，我們將應性門檻值與號誌特性進行結合，以此來獲取較為準確的號誌半徑。由於在影像中，有可能會同時出現前後多個路口的號誌，我們便提出了一個判斷距離車輛最近之路口的方法。 在猶豫區間判斷模組中，我們主要藉由一個單眼鏡頭模型，將影像平面內每一個單獨像素與其在真實世界中所表示的距離進行聯繫。透過交通號誌在影像平面內的半徑，可以推估出其與路口的實際距離，並由此取得路口寬度與車速的資訊。如果猶豫區間存在，系統將向駕駛者提供兩種時間資訊。駕駛者將獲知進入猶豫區間尚需的時間，以及在進入猶豫區間後，駛離猶豫區間所需的時間。 從本研究的試驗結果來看，處理與辨識模組提供了較好的號誌偵測資訊，其平均準確率達到94.56%。號誌半徑的偵測結果顯示其在增長過程中俱有波動性，增加了判斷的挑戰性，我們透過異常值判斷以及參數調整的方法予以改善，結果仍然顯示在距離估計上存在0.96公尺到4.83公尺的誤差。預警時間的種類的準確性達到了80% 。Dilemma Zone (DZ) is a hazardous area near signalized intersections. This work develops an on-board vision-based system to detect the potential DZ for the warning to drivers. The proposed monocular vision system analyzes parameters of the vehicle and the signalized intersection by detecting and measuring traffic lights, vehicular speed, intersection width, and distance to intersection. The system contains two modules, which are the processing and recognition module and the DZ judgement module. In the processing and recognition module, a two-stage method is adopted. We perform traffic light detection in stage one through color segmentation, shape filtering and classification. In stage two, we propose a method based on adaptive threshold to obtain the radius of detected traffic lights. Distance and velocity are derived by a monocular vision model. An intersection judgement algorithm is presented for intersections with multiple traffic lights. In the DZ judgement module, DZ calculation is based on a vehicle-oriented DZ definition. The experimental results show that the proposed system achieves an average precision rate of of 94.56% traffic light detection. The error in distance estimations is from 0.96 meters to 4.83 meters due to he fluctuant characteristics in detected radius. The average Mean Absolute Deviation (MAD) of distance estimation is 2.68 meters. The DZ warning performance is tested with 2 intersections. The accuracy of warning is over 80%. In our discussion, we find the vehicle velocity may have influence on our result and our system has potential to integrate with other vision systems.口試委員會審定書 i Acknowledgements ii 摘要 iii Abstract iv Contents v List of Figures vii List of Tables ix 1 Introduction 1 1.1 Background 1 1.2 Objectives of the study 2 1.3 Scope of the study 2 1.4 Methodology of the study 2 1.5 Structure of the study 3 2 Literature Review 5 2.1 Dilemma Zone 5 2.2 Traffic Light Detection 7 2.3 Distance Measurement 9 3 Methodology 13 3.1 Preparations 15 3.1.1 Camera Calibration 15 3.1.2 ROI of Frame 15 3.1.3 Color Space Selection 16 3.1.4 Classification 17 3.1.5 Radius Detector Selection 18 3.1.6 Tracking-based Intersection Judgement Algorithm 23 3.2 Processing and Recognition Module 32 3.2.1 Stage 1: Traffic Light Detection 32 3.2.2 Stage 2: Recognition 35 3.3 DZ Judgement Module 37 3.3.1 Parameters Calculation 37 3.3.2 DZ Calculation 47 3.3.3 DZ Warning 50 4 Experimental Result 51 4.1 SVM Classifier using HOG Feature Performance 52 4.2 Processing and Recognition Module Performance 52 4.3 DZ Judgement Module Performance 55 4.3.1 Distance and Intersection Width Estimation 56 4.3.2 Speed Range Seeking 60 4.3.3 DZ Judgement and Warning 63 4.3.4 Computational Time Result 65 4.4 Discussion 66 5 Conclusions and Future Work 69 5.1 Conclusions 69 5.2 Future Work 70 Reference 7212687495 bytesapplication/pdf論文公開時間：2016/08/17論文使用權限：同意有償授權(權利金給回饋學校)猶豫區間影像處理號誌偵測距離估計速度偵測路口辨識智慧交通thesishttp://ntur.lib.ntu.edu.tw/bitstream/246246/260747/1/ntu-103-R01521533-1.pdf