林巍聳臺灣大學：電機工程學研究所劉培德Liu, Pei-DePei-DeLiu2007-11-262018-07-062007-11-262018-07-062007http://ntur.lib.ntu.edu.tw//handle/246246/53118用立體視訊攝影機作為感測器，智慧型碰撞預警系統不僅可以偵測和顯示碰撞熱點，還可以估測其三維空間資訊和運動參數做為避碰控制的依據，本文詳述製作智慧型碰撞預警系統所需的關鍵技術，其中包括用邊緣萃取技術分離出標的物、用邊緣視差圖求取空間資訊、用演算視注覺決定注目點、用注目點擴充法劃分標的物的範圍、用標的物追蹤法從連續視差圖估測運動參數、以及根據標的物的距離和動向決定碰撞熱點等。碰撞熱點表示標的物位於緊急區或位於危險區但是正向攝影機移動，除了可以在影像上面標示碰撞熱點和發送碰撞預警信號以外，還可以提供碰撞熱點的位置和動向給避碰控制器，相關技術被整合為一個完整系統並製作成試驗機，實驗結果顯示用立體攝影機製作智慧型碰撞預警系統確實可行，可以輔助駕駛人執行主動式安全防護動作。Sensing with stereo camera, intelligent collision alert system can not only detect and display collision hot spots but also estimate the three-dimension spatial information and motion parameters for collision avoidance control. To implement the intelligent collision alert system, the techniques of edge extraction to isolate target objects, generation of edge disparity map to find the spatial information, computational visual attention to find attentive spots, region of interest determination to encompass object areas, object tracking to estimate motion parameters from successive disparity maps, and the strategy to find collision hot spots by according to the distances and motion directions of the target objects are developed and integrated as a whole. Target objects falling in the emergency zone or in the danger zone but approaching the camera are determined as the collision hot spots. In addition to making collision alert and displaying the collision hot spots on the image, the three-dimension spatial information and motion directions are estimated for collision avoidance control. The key techniques are verified with a prototype implementation. Experimental results show that the intelligent collision alert system by stereo camera sensing is feasible and practical for driver assistance of active safety.中文摘要 iii Abstract iv Chapter 1 Introduction 1 1.1 Background, Motivation and Contribution 1 1.2 Organization of This Thesis 5 Chapter 2 Computational Stereovision System and Edge Disparity Map 7 2.1 Introduction 7 2.2 Computational Stereovision System 9 2.2.1 Theorem of Stereo Matching 9 2.2.2 Window-based Stereo Matching 10 2.2.3 The Sum of Squared Difference (SSD) 11 2.2.4 The Normalized Cross Covariance (NCC) 11 2.3 Disparity Map 14 2.4 3D Reconstruction 15 2.5 Edge Disparity Map 17 2.5.1 Noise Filter 18 2.5.2 Edge Detection 20 2.5.3 Sobel Edge Detector 21 Chapter 3 Saliency Detection 25 3.1 Introduction 25 3.2 Computational Visual Attention 25 3.2.1 Gaussian Pyramid 29 3.2.2 Center-Surround mechanism 32 3.2.3 Global Nonlinear Normalization Operator 34 3.2.4 Feature Extracting Method 36 3.2.5 Saliency Map 41 3.3 Saliency Detection 43 Chapter 4 Collision Alert 51 4.1 Introduction 51 4.2 Estimate object locations 52 4.3 Estimate directions of object move 53 4.3.1 Kernel Density Estimation 55 4.3.2 Image Representation 58 4.3.3 Similarity between Distributions 61 4.3.4 Result of estimating directions of object move 63 Chapter 5 Experimental Results 65 5.1 Experimental System 65 5.2 System Framework 65 5.3 Experiments and results 67 5.3.1 Edge Disparity Map Generation 70 5.3.2 Estimation of 3D locations 80 5.3.3 Collision Detection 84 5.5.4 Experimental Results 87 Chapter 6 Conclusion 92 Reference 941199007 bytesapplication/pdfen-US避碰碰撞預警立體視覺視差圖主動式安全駕駛人輔助系統collision avoidancecollision alertstereovisiondisparity mapactive safetydriver assistancethesishttp://ntur.lib.ntu.edu.tw/bitstream/246246/53118/1/ntu-96-R94921077-1.pdf