傅立成臺灣大學：資訊工程學研究所黃贊宇Huang, Chan-YuChan-YuHuang2007-11-262018-07-052007-11-262018-07-052007http://ntur.lib.ntu.edu.tw//handle/246246/53760本論文提出一個偵測車輛前方多車道線以及多車輛的方法。我們假設車道線偵測以及車輛偵測會各自獨尋找影像中的車道以及車輛，透過整合車道線偵測及車輛偵測的結果資訊就可取得更為準確的偵測結果。 在車道線偵測方面，車道線特徵的分析常受到前方車輛的邊線或是車輛上的顏色所影響，導致特徵分析錯誤。同樣的在車輛偵測方面，出現在背景中與車輛相似的特徵會干擾車輛特徵之分析，導致車輛偵測不穩定的發生。 因此，本論文中使用車輛假設之位置與車道中心之間的距離，過濾不可能為車輛的物件。另一方面，使用車輛移動方向與車道線之方向的相似程度，取得最佳的車道線偵測結果。為了整合此些資訊，我們使用反覆執行之最佳化演算法將資料整合並取得最佳的結果，並利用車道線偵測以及車輛偵測之方法取得近似的結果，取得整合所需的資訊。最後在實驗結果的驗證中，我們與一般的偵測方法做比較並驗證出本論文方法之效果。This thesis presents an approach to detect multiple lane and vehicles. Instead of assuming that the processes of lane and vehicle detection should do independently, we integrate these two processes in a mutually supporting way to achieve more accurate results. In lane boundary detection, the process of identifying possible features of a lane boundary is often affected by the edges and color of the vehicles on the road. Likewise, the results of vehicle detection could be non-robust if there are some background features which can confuse the process of indentifying possible vehicle features. Thus, in the thesis, we use the distance between the central position of a lane and the position of the hypothesized vehicle to filter out the non-vehicle object. And we use the similarity between the lane boundary direction and the moving directions of the hypothesized vehicles to get the optimal lane solution. By applying iterative optimization algorithm, we can obtain the near-optimal solutions of both lane and vehicle detections. Finally the experimental results are provided to validate the effectiveness of the proposed novel approach.誌　謝 i 中文摘要 ii Abstract iii Contents iv List of Figures vii List of Tables viii Chapter 1 Introduction 1 1.1 Motivation 1 1.2 Related work 3 1.3 Objective 4 1.4 System Overview 5 1.5 Organization 7 Chapter 2 Preliminary 8 2.1 Camera Calibration 8 2.1.1 Camera Configuration 8 2.1.2 Transformation Formulation 10 2.1.3 Vanishing Point 11 2.2 Lane Detection Procedures 12 2.2.1 Feature Extraction 13 2.2.2 Search Methods 15 2.2.3 Lane Structure 16 2.3 Vehicle Detection Procedures 16 2.3.1 Hypothesis Generation 16 2.3.2 Hypothesis Verification 17 Chapter 3 Lane Hypothesis Generation 19 3.1 Overview 19 3.2 Formulation 20 3.2.1 Hybrid Lane Boundary Model 20 3.2.2 Configuration of a Lane Hypothesis 22 3.3 Line Marking Feature Extraction 23 3.3.1 Feature of Lane Marking 23 3.3.2 Line Segment Construction 24 3.4 Hypothesis Generation 26 3.4.1 Line Segment Pairing 26 3.4.2 Solution Graph Construction 27 3.4.3 Lane Boundary Model Fitting 30 3.4.4 Multiple Lane Extension 31 Chapter 4 Vehicle Hypothesis Generation 33 4.1 Overview 33 4.2 Vehicle Generation with Particle Filter 34 4.2.1 Initial Sampling 34 4.2.2 Propagation 36 4.2.3 Observation 37 4.3 Feature Cues of Vehicles 37 4.3.1 Bounding Box of a Sample 38 4.3.2 Underneath Cue 38 4.3.3 Vertical Edge Cue 39 4.3.4 Symmetry Cue 39 4.3.5 Taillight Cue 40 4.3.6 Cue Fusion 42 4.4 Hypothesis Generation 42 4.4.1 Sample Clustering for Mean-Shift 42 4.4.2 Hypothesis Tracking 43 Chapter 5 Integration framework 45 5.1 Overview 45 5.2 Integration Model 46 5.3 Confidence Initialization 47 5.3.1 Lane Confidence Initialization 48 5.3.2 Vehicle Confidence Initialization 49 5.4 Likelihood of Hypotheses 49 5.4.1 Vehicle Likelihood Estimation 49 5.4.2 Lane Likelihood Estimation 51 5.5 Hypotheses Integration 52 5.5.1 Reweighting of Confidence 53 5.5.2 Iterative Algorithm of Integration 53 5.5.3 Finding the Optimal Solution 54 Chapter 6 Experiments 56 6.1 Environment Description 56 6.2 Environment Results 56 6.3 Performance Analysis 58 Chapter 7 Conclusion 61 Reference 631139651 bytesapplication/pdfen-US多車輛偵測多車道線偵測資訊整合平台曲線車道線偵測multiple vehicle detectionmultiple lane detectionintegratio frameworkcurve lane detectionthesishttp://ntur.lib.ntu.edu.tw/bitstream/246246/53760/1/ntu-96-R94922043-1.pdf